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/bitops.h>
19 #include <linux/slab.h>
21 #include <linux/proc_fs.h>
22 #include <linux/seq_file.h>
23 #include <linux/kasan.h>
24 #include <linux/cpu.h>
25 #include <linux/cpuset.h>
26 #include <linux/mempolicy.h>
27 #include <linux/ctype.h>
28 #include <linux/debugobjects.h>
29 #include <linux/kallsyms.h>
30 #include <linux/memory.h>
31 #include <linux/math64.h>
32 #include <linux/fault-inject.h>
33 #include <linux/stacktrace.h>
34 #include <linux/prefetch.h>
35 #include <linux/memcontrol.h>
36 #include <linux/random.h>
38 #include <trace/events/kmem.h>
44 * 1. slab_mutex (Global Mutex)
46 * 3. slab_lock(page) (Only on some arches and for debugging)
50 * The role of the slab_mutex is to protect the list of all the slabs
51 * and to synchronize major metadata changes to slab cache structures.
53 * The slab_lock is only used for debugging and on arches that do not
54 * have the ability to do a cmpxchg_double. It only protects:
55 * A. page->freelist -> List of object free in a page
56 * B. page->inuse -> Number of objects in use
57 * C. page->objects -> Number of objects in page
58 * D. page->frozen -> frozen state
60 * If a slab is frozen then it is exempt from list management. It is not
61 * on any list except per cpu partial list. The processor that froze the
62 * slab is the one who can perform list operations on the page. Other
63 * processors may put objects onto the freelist but the processor that
64 * froze the slab is the only one that can retrieve the objects from the
67 * The list_lock protects the partial and full list on each node and
68 * the partial slab counter. If taken then no new slabs may be added or
69 * removed from the lists nor make the number of partial slabs be modified.
70 * (Note that the total number of slabs is an atomic value that may be
71 * modified without taking the list lock).
73 * The list_lock is a centralized lock and thus we avoid taking it as
74 * much as possible. As long as SLUB does not have to handle partial
75 * slabs, operations can continue without any centralized lock. F.e.
76 * allocating a long series of objects that fill up slabs does not require
78 * Interrupts are disabled during allocation and deallocation in order to
79 * make the slab allocator safe to use in the context of an irq. In addition
80 * interrupts are disabled to ensure that the processor does not change
81 * while handling per_cpu slabs, due to kernel preemption.
83 * SLUB assigns one slab for allocation to each processor.
84 * Allocations only occur from these slabs called cpu slabs.
86 * Slabs with free elements are kept on a partial list and during regular
87 * operations no list for full slabs is used. If an object in a full slab is
88 * freed then the slab will show up again on the partial lists.
89 * We track full slabs for debugging purposes though because otherwise we
90 * cannot scan all objects.
92 * Slabs are freed when they become empty. Teardown and setup is
93 * minimal so we rely on the page allocators per cpu caches for
94 * fast frees and allocs.
96 * Overloading of page flags that are otherwise used for LRU management.
98 * PageActive The slab is frozen and exempt from list processing.
99 * This means that the slab is dedicated to a purpose
100 * such as satisfying allocations for a specific
101 * processor. Objects may be freed in the slab while
102 * it is frozen but slab_free will then skip the usual
103 * list operations. It is up to the processor holding
104 * the slab to integrate the slab into the slab lists
105 * when the slab is no longer needed.
107 * One use of this flag is to mark slabs that are
108 * used for allocations. Then such a slab becomes a cpu
109 * slab. The cpu slab may be equipped with an additional
110 * freelist that allows lockless access to
111 * free objects in addition to the regular freelist
112 * that requires the slab lock.
114 * PageError Slab requires special handling due to debug
115 * options set. This moves slab handling out of
116 * the fast path and disables lockless freelists.
119 static inline int kmem_cache_debug(struct kmem_cache *s)
121 #ifdef CONFIG_SLUB_DEBUG
122 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
128 void *fixup_red_left(struct kmem_cache *s, void *p)
130 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE)
131 p += s->red_left_pad;
136 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
138 #ifdef CONFIG_SLUB_CPU_PARTIAL
139 return !kmem_cache_debug(s);
146 * Issues still to be resolved:
148 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
150 * - Variable sizing of the per node arrays
153 /* Enable to test recovery from slab corruption on boot */
154 #undef SLUB_RESILIENCY_TEST
156 /* Enable to log cmpxchg failures */
157 #undef SLUB_DEBUG_CMPXCHG
160 * Mininum number of partial slabs. These will be left on the partial
161 * lists even if they are empty. kmem_cache_shrink may reclaim them.
163 #define MIN_PARTIAL 5
166 * Maximum number of desirable partial slabs.
167 * The existence of more partial slabs makes kmem_cache_shrink
168 * sort the partial list by the number of objects in use.
170 #define MAX_PARTIAL 10
172 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
173 SLAB_POISON | SLAB_STORE_USER)
176 * These debug flags cannot use CMPXCHG because there might be consistency
177 * issues when checking or reading debug information
179 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
184 * Debugging flags that require metadata to be stored in the slab. These get
185 * disabled when slub_debug=O is used and a cache's min order increases with
188 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
191 #define OO_MASK ((1 << OO_SHIFT) - 1)
192 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
194 /* Internal SLUB flags */
196 #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
197 /* Use cmpxchg_double */
198 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
201 * Tracking user of a slab.
203 #define TRACK_ADDRS_COUNT 16
205 unsigned long addr; /* Called from address */
206 #ifdef CONFIG_STACKTRACE
207 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
209 int cpu; /* Was running on cpu */
210 int pid; /* Pid context */
211 unsigned long when; /* When did the operation occur */
214 enum track_item { TRACK_ALLOC, TRACK_FREE };
217 static int sysfs_slab_add(struct kmem_cache *);
218 static int sysfs_slab_alias(struct kmem_cache *, const char *);
219 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
220 static void sysfs_slab_remove(struct kmem_cache *s);
222 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
223 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
225 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
226 static inline void sysfs_slab_remove(struct kmem_cache *s) { }
229 static inline void stat(const struct kmem_cache *s, enum stat_item si)
231 #ifdef CONFIG_SLUB_STATS
233 * The rmw is racy on a preemptible kernel but this is acceptable, so
234 * avoid this_cpu_add()'s irq-disable overhead.
236 raw_cpu_inc(s->cpu_slab->stat[si]);
240 /********************************************************************
241 * Core slab cache functions
242 *******************************************************************/
245 * Returns freelist pointer (ptr). With hardening, this is obfuscated
246 * with an XOR of the address where the pointer is held and a per-cache
249 static inline void *freelist_ptr(const struct kmem_cache *s, void *ptr,
250 unsigned long ptr_addr)
252 #ifdef CONFIG_SLAB_FREELIST_HARDENED
254 * When CONFIG_KASAN_SW_TAGS is enabled, ptr_addr might be tagged.
255 * Normally, this doesn't cause any issues, as both set_freepointer()
256 * and get_freepointer() are called with a pointer with the same tag.
257 * However, there are some issues with CONFIG_SLUB_DEBUG code. For
258 * example, when __free_slub() iterates over objects in a cache, it
259 * passes untagged pointers to check_object(). check_object() in turns
260 * calls get_freepointer() with an untagged pointer, which causes the
261 * freepointer to be restored incorrectly.
263 return (void *)((unsigned long)ptr ^ s->random ^
264 (unsigned long)kasan_reset_tag((void *)ptr_addr));
270 /* Returns the freelist pointer recorded at location ptr_addr. */
271 static inline void *freelist_dereference(const struct kmem_cache *s,
274 return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr),
275 (unsigned long)ptr_addr);
278 static inline void *get_freepointer(struct kmem_cache *s, void *object)
280 return freelist_dereference(s, object + s->offset);
283 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
285 prefetch(object + s->offset);
288 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
290 unsigned long freepointer_addr;
293 if (!debug_pagealloc_enabled())
294 return get_freepointer(s, object);
296 freepointer_addr = (unsigned long)object + s->offset;
297 probe_kernel_read(&p, (void **)freepointer_addr, sizeof(p));
298 return freelist_ptr(s, p, freepointer_addr);
301 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
303 unsigned long freeptr_addr = (unsigned long)object + s->offset;
305 #ifdef CONFIG_SLAB_FREELIST_HARDENED
306 BUG_ON(object == fp); /* naive detection of double free or corruption */
309 *(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr);
312 /* Loop over all objects in a slab */
313 #define for_each_object(__p, __s, __addr, __objects) \
314 for (__p = fixup_red_left(__s, __addr); \
315 __p < (__addr) + (__objects) * (__s)->size; \
318 /* Determine object index from a given position */
319 static inline unsigned int slab_index(void *p, struct kmem_cache *s, void *addr)
321 return (kasan_reset_tag(p) - addr) / s->size;
324 static inline unsigned int order_objects(unsigned int order, unsigned int size)
326 return ((unsigned int)PAGE_SIZE << order) / size;
329 static inline struct kmem_cache_order_objects oo_make(unsigned int order,
332 struct kmem_cache_order_objects x = {
333 (order << OO_SHIFT) + order_objects(order, size)
339 static inline unsigned int oo_order(struct kmem_cache_order_objects x)
341 return x.x >> OO_SHIFT;
344 static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
346 return x.x & OO_MASK;
350 * Per slab locking using the pagelock
352 static __always_inline void slab_lock(struct page *page)
354 VM_BUG_ON_PAGE(PageTail(page), page);
355 bit_spin_lock(PG_locked, &page->flags);
358 static __always_inline void slab_unlock(struct page *page)
360 VM_BUG_ON_PAGE(PageTail(page), page);
361 __bit_spin_unlock(PG_locked, &page->flags);
364 /* Interrupts must be disabled (for the fallback code to work right) */
365 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
366 void *freelist_old, unsigned long counters_old,
367 void *freelist_new, unsigned long counters_new,
370 VM_BUG_ON(!irqs_disabled());
371 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
372 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
373 if (s->flags & __CMPXCHG_DOUBLE) {
374 if (cmpxchg_double(&page->freelist, &page->counters,
375 freelist_old, counters_old,
376 freelist_new, counters_new))
382 if (page->freelist == freelist_old &&
383 page->counters == counters_old) {
384 page->freelist = freelist_new;
385 page->counters = counters_new;
393 stat(s, CMPXCHG_DOUBLE_FAIL);
395 #ifdef SLUB_DEBUG_CMPXCHG
396 pr_info("%s %s: cmpxchg double redo ", n, s->name);
402 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
403 void *freelist_old, unsigned long counters_old,
404 void *freelist_new, unsigned long counters_new,
407 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
408 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
409 if (s->flags & __CMPXCHG_DOUBLE) {
410 if (cmpxchg_double(&page->freelist, &page->counters,
411 freelist_old, counters_old,
412 freelist_new, counters_new))
419 local_irq_save(flags);
421 if (page->freelist == freelist_old &&
422 page->counters == counters_old) {
423 page->freelist = freelist_new;
424 page->counters = counters_new;
426 local_irq_restore(flags);
430 local_irq_restore(flags);
434 stat(s, CMPXCHG_DOUBLE_FAIL);
436 #ifdef SLUB_DEBUG_CMPXCHG
437 pr_info("%s %s: cmpxchg double redo ", n, s->name);
443 #ifdef CONFIG_SLUB_DEBUG
445 * Determine a map of object in use on a page.
447 * Node listlock must be held to guarantee that the page does
448 * not vanish from under us.
450 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
453 void *addr = page_address(page);
455 for (p = page->freelist; p; p = get_freepointer(s, p))
456 set_bit(slab_index(p, s, addr), map);
459 static inline unsigned int size_from_object(struct kmem_cache *s)
461 if (s->flags & SLAB_RED_ZONE)
462 return s->size - s->red_left_pad;
467 static inline void *restore_red_left(struct kmem_cache *s, void *p)
469 if (s->flags & SLAB_RED_ZONE)
470 p -= s->red_left_pad;
478 #if defined(CONFIG_SLUB_DEBUG_ON)
479 static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
481 static slab_flags_t slub_debug;
484 static char *slub_debug_slabs;
485 static int disable_higher_order_debug;
488 * slub is about to manipulate internal object metadata. This memory lies
489 * outside the range of the allocated object, so accessing it would normally
490 * be reported by kasan as a bounds error. metadata_access_enable() is used
491 * to tell kasan that these accesses are OK.
493 static inline void metadata_access_enable(void)
495 kasan_disable_current();
498 static inline void metadata_access_disable(void)
500 kasan_enable_current();
507 /* Verify that a pointer has an address that is valid within a slab page */
508 static inline int check_valid_pointer(struct kmem_cache *s,
509 struct page *page, void *object)
516 base = page_address(page);
517 object = kasan_reset_tag(object);
518 object = restore_red_left(s, object);
519 if (object < base || object >= base + page->objects * s->size ||
520 (object - base) % s->size) {
527 static void print_section(char *level, char *text, u8 *addr,
530 metadata_access_enable();
531 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
533 metadata_access_disable();
536 static struct track *get_track(struct kmem_cache *s, void *object,
537 enum track_item alloc)
542 p = object + s->offset + sizeof(void *);
544 p = object + s->inuse;
549 static void set_track(struct kmem_cache *s, void *object,
550 enum track_item alloc, unsigned long addr)
552 struct track *p = get_track(s, object, alloc);
555 #ifdef CONFIG_STACKTRACE
556 unsigned int nr_entries;
558 metadata_access_enable();
559 nr_entries = stack_trace_save(p->addrs, TRACK_ADDRS_COUNT, 3);
560 metadata_access_disable();
562 if (nr_entries < TRACK_ADDRS_COUNT)
563 p->addrs[nr_entries] = 0;
566 p->cpu = smp_processor_id();
567 p->pid = current->pid;
570 memset(p, 0, sizeof(struct track));
574 static void init_tracking(struct kmem_cache *s, void *object)
576 if (!(s->flags & SLAB_STORE_USER))
579 set_track(s, object, TRACK_FREE, 0UL);
580 set_track(s, object, TRACK_ALLOC, 0UL);
583 static void print_track(const char *s, struct track *t, unsigned long pr_time)
588 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
589 s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
590 #ifdef CONFIG_STACKTRACE
593 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
595 pr_err("\t%pS\n", (void *)t->addrs[i]);
602 static void print_tracking(struct kmem_cache *s, void *object)
604 unsigned long pr_time = jiffies;
605 if (!(s->flags & SLAB_STORE_USER))
608 print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
609 print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
612 static void print_page_info(struct page *page)
614 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
615 page, page->objects, page->inuse, page->freelist, page->flags);
619 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
621 struct va_format vaf;
627 pr_err("=============================================================================\n");
628 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
629 pr_err("-----------------------------------------------------------------------------\n\n");
631 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
635 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
637 struct va_format vaf;
643 pr_err("FIX %s: %pV\n", s->name, &vaf);
647 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
649 unsigned int off; /* Offset of last byte */
650 u8 *addr = page_address(page);
652 print_tracking(s, p);
654 print_page_info(page);
656 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
657 p, p - addr, get_freepointer(s, p));
659 if (s->flags & SLAB_RED_ZONE)
660 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
662 else if (p > addr + 16)
663 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
665 print_section(KERN_ERR, "Object ", p,
666 min_t(unsigned int, s->object_size, PAGE_SIZE));
667 if (s->flags & SLAB_RED_ZONE)
668 print_section(KERN_ERR, "Redzone ", p + s->object_size,
669 s->inuse - s->object_size);
672 off = s->offset + sizeof(void *);
676 if (s->flags & SLAB_STORE_USER)
677 off += 2 * sizeof(struct track);
679 off += kasan_metadata_size(s);
681 if (off != size_from_object(s))
682 /* Beginning of the filler is the free pointer */
683 print_section(KERN_ERR, "Padding ", p + off,
684 size_from_object(s) - off);
689 void object_err(struct kmem_cache *s, struct page *page,
690 u8 *object, char *reason)
692 slab_bug(s, "%s", reason);
693 print_trailer(s, page, object);
696 static __printf(3, 4) void slab_err(struct kmem_cache *s, struct page *page,
697 const char *fmt, ...)
703 vsnprintf(buf, sizeof(buf), fmt, args);
705 slab_bug(s, "%s", buf);
706 print_page_info(page);
710 static void init_object(struct kmem_cache *s, void *object, u8 val)
714 if (s->flags & SLAB_RED_ZONE)
715 memset(p - s->red_left_pad, val, s->red_left_pad);
717 if (s->flags & __OBJECT_POISON) {
718 memset(p, POISON_FREE, s->object_size - 1);
719 p[s->object_size - 1] = POISON_END;
722 if (s->flags & SLAB_RED_ZONE)
723 memset(p + s->object_size, val, s->inuse - s->object_size);
726 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
727 void *from, void *to)
729 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
730 memset(from, data, to - from);
733 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
734 u8 *object, char *what,
735 u8 *start, unsigned int value, unsigned int bytes)
740 metadata_access_enable();
741 fault = memchr_inv(start, value, bytes);
742 metadata_access_disable();
747 while (end > fault && end[-1] == value)
750 slab_bug(s, "%s overwritten", what);
751 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
752 fault, end - 1, fault[0], value);
753 print_trailer(s, page, object);
755 restore_bytes(s, what, value, fault, end);
763 * Bytes of the object to be managed.
764 * If the freepointer may overlay the object then the free
765 * pointer is the first word of the object.
767 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
770 * object + s->object_size
771 * Padding to reach word boundary. This is also used for Redzoning.
772 * Padding is extended by another word if Redzoning is enabled and
773 * object_size == inuse.
775 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
776 * 0xcc (RED_ACTIVE) for objects in use.
779 * Meta data starts here.
781 * A. Free pointer (if we cannot overwrite object on free)
782 * B. Tracking data for SLAB_STORE_USER
783 * C. Padding to reach required alignment boundary or at mininum
784 * one word if debugging is on to be able to detect writes
785 * before the word boundary.
787 * Padding is done using 0x5a (POISON_INUSE)
790 * Nothing is used beyond s->size.
792 * If slabcaches are merged then the object_size and inuse boundaries are mostly
793 * ignored. And therefore no slab options that rely on these boundaries
794 * may be used with merged slabcaches.
797 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
799 unsigned long off = s->inuse; /* The end of info */
802 /* Freepointer is placed after the object. */
803 off += sizeof(void *);
805 if (s->flags & SLAB_STORE_USER)
806 /* We also have user information there */
807 off += 2 * sizeof(struct track);
809 off += kasan_metadata_size(s);
811 if (size_from_object(s) == off)
814 return check_bytes_and_report(s, page, p, "Object padding",
815 p + off, POISON_INUSE, size_from_object(s) - off);
818 /* Check the pad bytes at the end of a slab page */
819 static int slab_pad_check(struct kmem_cache *s, struct page *page)
828 if (!(s->flags & SLAB_POISON))
831 start = page_address(page);
832 length = PAGE_SIZE << compound_order(page);
833 end = start + length;
834 remainder = length % s->size;
838 pad = end - remainder;
839 metadata_access_enable();
840 fault = memchr_inv(pad, POISON_INUSE, remainder);
841 metadata_access_disable();
844 while (end > fault && end[-1] == POISON_INUSE)
847 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
848 print_section(KERN_ERR, "Padding ", pad, remainder);
850 restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
854 static int check_object(struct kmem_cache *s, struct page *page,
855 void *object, u8 val)
858 u8 *endobject = object + s->object_size;
860 if (s->flags & SLAB_RED_ZONE) {
861 if (!check_bytes_and_report(s, page, object, "Redzone",
862 object - s->red_left_pad, val, s->red_left_pad))
865 if (!check_bytes_and_report(s, page, object, "Redzone",
866 endobject, val, s->inuse - s->object_size))
869 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
870 check_bytes_and_report(s, page, p, "Alignment padding",
871 endobject, POISON_INUSE,
872 s->inuse - s->object_size);
876 if (s->flags & SLAB_POISON) {
877 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
878 (!check_bytes_and_report(s, page, p, "Poison", p,
879 POISON_FREE, s->object_size - 1) ||
880 !check_bytes_and_report(s, page, p, "Poison",
881 p + s->object_size - 1, POISON_END, 1)))
884 * check_pad_bytes cleans up on its own.
886 check_pad_bytes(s, page, p);
889 if (!s->offset && val == SLUB_RED_ACTIVE)
891 * Object and freepointer overlap. Cannot check
892 * freepointer while object is allocated.
896 /* Check free pointer validity */
897 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
898 object_err(s, page, p, "Freepointer corrupt");
900 * No choice but to zap it and thus lose the remainder
901 * of the free objects in this slab. May cause
902 * another error because the object count is now wrong.
904 set_freepointer(s, p, NULL);
910 static int check_slab(struct kmem_cache *s, struct page *page)
914 VM_BUG_ON(!irqs_disabled());
916 if (!PageSlab(page)) {
917 slab_err(s, page, "Not a valid slab page");
921 maxobj = order_objects(compound_order(page), s->size);
922 if (page->objects > maxobj) {
923 slab_err(s, page, "objects %u > max %u",
924 page->objects, maxobj);
927 if (page->inuse > page->objects) {
928 slab_err(s, page, "inuse %u > max %u",
929 page->inuse, page->objects);
932 /* Slab_pad_check fixes things up after itself */
933 slab_pad_check(s, page);
938 * Determine if a certain object on a page is on the freelist. Must hold the
939 * slab lock to guarantee that the chains are in a consistent state.
941 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
949 while (fp && nr <= page->objects) {
952 if (!check_valid_pointer(s, page, fp)) {
954 object_err(s, page, object,
955 "Freechain corrupt");
956 set_freepointer(s, object, NULL);
958 slab_err(s, page, "Freepointer corrupt");
959 page->freelist = NULL;
960 page->inuse = page->objects;
961 slab_fix(s, "Freelist cleared");
967 fp = get_freepointer(s, object);
971 max_objects = order_objects(compound_order(page), s->size);
972 if (max_objects > MAX_OBJS_PER_PAGE)
973 max_objects = MAX_OBJS_PER_PAGE;
975 if (page->objects != max_objects) {
976 slab_err(s, page, "Wrong number of objects. Found %d but should be %d",
977 page->objects, max_objects);
978 page->objects = max_objects;
979 slab_fix(s, "Number of objects adjusted.");
981 if (page->inuse != page->objects - nr) {
982 slab_err(s, page, "Wrong object count. Counter is %d but counted were %d",
983 page->inuse, page->objects - nr);
984 page->inuse = page->objects - nr;
985 slab_fix(s, "Object count adjusted.");
987 return search == NULL;
990 static void trace(struct kmem_cache *s, struct page *page, void *object,
993 if (s->flags & SLAB_TRACE) {
994 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
996 alloc ? "alloc" : "free",
1001 print_section(KERN_INFO, "Object ", (void *)object,
1009 * Tracking of fully allocated slabs for debugging purposes.
1011 static void add_full(struct kmem_cache *s,
1012 struct kmem_cache_node *n, struct page *page)
1014 if (!(s->flags & SLAB_STORE_USER))
1017 lockdep_assert_held(&n->list_lock);
1018 list_add(&page->slab_list, &n->full);
1021 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
1023 if (!(s->flags & SLAB_STORE_USER))
1026 lockdep_assert_held(&n->list_lock);
1027 list_del(&page->slab_list);
1030 /* Tracking of the number of slabs for debugging purposes */
1031 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1033 struct kmem_cache_node *n = get_node(s, node);
1035 return atomic_long_read(&n->nr_slabs);
1038 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1040 return atomic_long_read(&n->nr_slabs);
1043 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1045 struct kmem_cache_node *n = get_node(s, node);
1048 * May be called early in order to allocate a slab for the
1049 * kmem_cache_node structure. Solve the chicken-egg
1050 * dilemma by deferring the increment of the count during
1051 * bootstrap (see early_kmem_cache_node_alloc).
1054 atomic_long_inc(&n->nr_slabs);
1055 atomic_long_add(objects, &n->total_objects);
1058 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1060 struct kmem_cache_node *n = get_node(s, node);
1062 atomic_long_dec(&n->nr_slabs);
1063 atomic_long_sub(objects, &n->total_objects);
1066 /* Object debug checks for alloc/free paths */
1067 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1070 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1073 init_object(s, object, SLUB_RED_INACTIVE);
1074 init_tracking(s, object);
1077 static void setup_page_debug(struct kmem_cache *s, void *addr, int order)
1079 if (!(s->flags & SLAB_POISON))
1082 metadata_access_enable();
1083 memset(addr, POISON_INUSE, PAGE_SIZE << order);
1084 metadata_access_disable();
1087 static inline int alloc_consistency_checks(struct kmem_cache *s,
1088 struct page *page, void *object)
1090 if (!check_slab(s, page))
1093 if (!check_valid_pointer(s, page, object)) {
1094 object_err(s, page, object, "Freelist Pointer check fails");
1098 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1104 static noinline int alloc_debug_processing(struct kmem_cache *s,
1106 void *object, unsigned long addr)
1108 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1109 if (!alloc_consistency_checks(s, page, object))
1113 /* Success perform special debug activities for allocs */
1114 if (s->flags & SLAB_STORE_USER)
1115 set_track(s, object, TRACK_ALLOC, addr);
1116 trace(s, page, object, 1);
1117 init_object(s, object, SLUB_RED_ACTIVE);
1121 if (PageSlab(page)) {
1123 * If this is a slab page then lets do the best we can
1124 * to avoid issues in the future. Marking all objects
1125 * as used avoids touching the remaining objects.
1127 slab_fix(s, "Marking all objects used");
1128 page->inuse = page->objects;
1129 page->freelist = NULL;
1134 static inline int free_consistency_checks(struct kmem_cache *s,
1135 struct page *page, void *object, unsigned long addr)
1137 if (!check_valid_pointer(s, page, object)) {
1138 slab_err(s, page, "Invalid object pointer 0x%p", object);
1142 if (on_freelist(s, page, object)) {
1143 object_err(s, page, object, "Object already free");
1147 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1150 if (unlikely(s != page->slab_cache)) {
1151 if (!PageSlab(page)) {
1152 slab_err(s, page, "Attempt to free object(0x%p) outside of slab",
1154 } else if (!page->slab_cache) {
1155 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1159 object_err(s, page, object,
1160 "page slab pointer corrupt.");
1166 /* Supports checking bulk free of a constructed freelist */
1167 static noinline int free_debug_processing(
1168 struct kmem_cache *s, struct page *page,
1169 void *head, void *tail, int bulk_cnt,
1172 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1173 void *object = head;
1175 unsigned long uninitialized_var(flags);
1178 spin_lock_irqsave(&n->list_lock, flags);
1181 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1182 if (!check_slab(s, page))
1189 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1190 if (!free_consistency_checks(s, page, object, addr))
1194 if (s->flags & SLAB_STORE_USER)
1195 set_track(s, object, TRACK_FREE, addr);
1196 trace(s, page, object, 0);
1197 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1198 init_object(s, object, SLUB_RED_INACTIVE);
1200 /* Reached end of constructed freelist yet? */
1201 if (object != tail) {
1202 object = get_freepointer(s, object);
1208 if (cnt != bulk_cnt)
1209 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1213 spin_unlock_irqrestore(&n->list_lock, flags);
1215 slab_fix(s, "Object at 0x%p not freed", object);
1219 static int __init setup_slub_debug(char *str)
1221 slub_debug = DEBUG_DEFAULT_FLAGS;
1222 if (*str++ != '=' || !*str)
1224 * No options specified. Switch on full debugging.
1230 * No options but restriction on slabs. This means full
1231 * debugging for slabs matching a pattern.
1238 * Switch off all debugging measures.
1243 * Determine which debug features should be switched on
1245 for (; *str && *str != ','; str++) {
1246 switch (tolower(*str)) {
1248 slub_debug |= SLAB_CONSISTENCY_CHECKS;
1251 slub_debug |= SLAB_RED_ZONE;
1254 slub_debug |= SLAB_POISON;
1257 slub_debug |= SLAB_STORE_USER;
1260 slub_debug |= SLAB_TRACE;
1263 slub_debug |= SLAB_FAILSLAB;
1267 * Avoid enabling debugging on caches if its minimum
1268 * order would increase as a result.
1270 disable_higher_order_debug = 1;
1273 pr_err("slub_debug option '%c' unknown. skipped\n",
1280 slub_debug_slabs = str + 1;
1282 if ((static_branch_unlikely(&init_on_alloc) ||
1283 static_branch_unlikely(&init_on_free)) &&
1284 (slub_debug & SLAB_POISON))
1285 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1289 __setup("slub_debug", setup_slub_debug);
1292 * kmem_cache_flags - apply debugging options to the cache
1293 * @object_size: the size of an object without meta data
1294 * @flags: flags to set
1295 * @name: name of the cache
1296 * @ctor: constructor function
1298 * Debug option(s) are applied to @flags. In addition to the debug
1299 * option(s), if a slab name (or multiple) is specified i.e.
1300 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1301 * then only the select slabs will receive the debug option(s).
1303 slab_flags_t kmem_cache_flags(unsigned int object_size,
1304 slab_flags_t flags, const char *name,
1305 void (*ctor)(void *))
1310 /* If slub_debug = 0, it folds into the if conditional. */
1311 if (!slub_debug_slabs)
1312 return flags | slub_debug;
1315 iter = slub_debug_slabs;
1320 end = strchrnul(iter, ',');
1322 glob = strnchr(iter, end - iter, '*');
1324 cmplen = glob - iter;
1326 cmplen = max_t(size_t, len, (end - iter));
1328 if (!strncmp(name, iter, cmplen)) {
1329 flags |= slub_debug;
1340 #else /* !CONFIG_SLUB_DEBUG */
1341 static inline void setup_object_debug(struct kmem_cache *s,
1342 struct page *page, void *object) {}
1343 static inline void setup_page_debug(struct kmem_cache *s,
1344 void *addr, int order) {}
1346 static inline int alloc_debug_processing(struct kmem_cache *s,
1347 struct page *page, void *object, unsigned long addr) { return 0; }
1349 static inline int free_debug_processing(
1350 struct kmem_cache *s, struct page *page,
1351 void *head, void *tail, int bulk_cnt,
1352 unsigned long addr) { return 0; }
1354 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1356 static inline int check_object(struct kmem_cache *s, struct page *page,
1357 void *object, u8 val) { return 1; }
1358 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1359 struct page *page) {}
1360 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1361 struct page *page) {}
1362 slab_flags_t kmem_cache_flags(unsigned int object_size,
1363 slab_flags_t flags, const char *name,
1364 void (*ctor)(void *))
1368 #define slub_debug 0
1370 #define disable_higher_order_debug 0
1372 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1374 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1376 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1378 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1381 #endif /* CONFIG_SLUB_DEBUG */
1384 * Hooks for other subsystems that check memory allocations. In a typical
1385 * production configuration these hooks all should produce no code at all.
1387 static inline void *kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1389 ptr = kasan_kmalloc_large(ptr, size, flags);
1390 /* As ptr might get tagged, call kmemleak hook after KASAN. */
1391 kmemleak_alloc(ptr, size, 1, flags);
1395 static __always_inline void kfree_hook(void *x)
1398 kasan_kfree_large(x, _RET_IP_);
1401 static __always_inline bool slab_free_hook(struct kmem_cache *s, void *x)
1403 kmemleak_free_recursive(x, s->flags);
1406 * Trouble is that we may no longer disable interrupts in the fast path
1407 * So in order to make the debug calls that expect irqs to be
1408 * disabled we need to disable interrupts temporarily.
1410 #ifdef CONFIG_LOCKDEP
1412 unsigned long flags;
1414 local_irq_save(flags);
1415 debug_check_no_locks_freed(x, s->object_size);
1416 local_irq_restore(flags);
1419 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1420 debug_check_no_obj_freed(x, s->object_size);
1422 /* KASAN might put x into memory quarantine, delaying its reuse */
1423 return kasan_slab_free(s, x, _RET_IP_);
1426 static inline bool slab_free_freelist_hook(struct kmem_cache *s,
1427 void **head, void **tail)
1432 void *old_tail = *tail ? *tail : *head;
1435 if (slab_want_init_on_free(s))
1438 next = get_freepointer(s, object);
1440 * Clear the object and the metadata, but don't touch
1443 memset(object, 0, s->object_size);
1444 rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad
1446 memset((char *)object + s->inuse, 0,
1447 s->size - s->inuse - rsize);
1448 set_freepointer(s, object, next);
1449 } while (object != old_tail);
1452 * Compiler cannot detect this function can be removed if slab_free_hook()
1453 * evaluates to nothing. Thus, catch all relevant config debug options here.
1455 #if defined(CONFIG_LOCKDEP) || \
1456 defined(CONFIG_DEBUG_KMEMLEAK) || \
1457 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1458 defined(CONFIG_KASAN)
1462 /* Head and tail of the reconstructed freelist */
1468 next = get_freepointer(s, object);
1469 /* If object's reuse doesn't have to be delayed */
1470 if (!slab_free_hook(s, object)) {
1471 /* Move object to the new freelist */
1472 set_freepointer(s, object, *head);
1477 } while (object != old_tail);
1482 return *head != NULL;
1488 static void *setup_object(struct kmem_cache *s, struct page *page,
1491 setup_object_debug(s, page, object);
1492 object = kasan_init_slab_obj(s, object);
1493 if (unlikely(s->ctor)) {
1494 kasan_unpoison_object_data(s, object);
1496 kasan_poison_object_data(s, object);
1502 * Slab allocation and freeing
1504 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1505 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1508 unsigned int order = oo_order(oo);
1510 if (node == NUMA_NO_NODE)
1511 page = alloc_pages(flags, order);
1513 page = __alloc_pages_node(node, flags, order);
1515 if (page && charge_slab_page(page, flags, order, s)) {
1516 __free_pages(page, order);
1523 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1524 /* Pre-initialize the random sequence cache */
1525 static int init_cache_random_seq(struct kmem_cache *s)
1527 unsigned int count = oo_objects(s->oo);
1530 /* Bailout if already initialised */
1534 err = cache_random_seq_create(s, count, GFP_KERNEL);
1536 pr_err("SLUB: Unable to initialize free list for %s\n",
1541 /* Transform to an offset on the set of pages */
1542 if (s->random_seq) {
1545 for (i = 0; i < count; i++)
1546 s->random_seq[i] *= s->size;
1551 /* Initialize each random sequence freelist per cache */
1552 static void __init init_freelist_randomization(void)
1554 struct kmem_cache *s;
1556 mutex_lock(&slab_mutex);
1558 list_for_each_entry(s, &slab_caches, list)
1559 init_cache_random_seq(s);
1561 mutex_unlock(&slab_mutex);
1564 /* Get the next entry on the pre-computed freelist randomized */
1565 static void *next_freelist_entry(struct kmem_cache *s, struct page *page,
1566 unsigned long *pos, void *start,
1567 unsigned long page_limit,
1568 unsigned long freelist_count)
1573 * If the target page allocation failed, the number of objects on the
1574 * page might be smaller than the usual size defined by the cache.
1577 idx = s->random_seq[*pos];
1579 if (*pos >= freelist_count)
1581 } while (unlikely(idx >= page_limit));
1583 return (char *)start + idx;
1586 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1587 static bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1592 unsigned long idx, pos, page_limit, freelist_count;
1594 if (page->objects < 2 || !s->random_seq)
1597 freelist_count = oo_objects(s->oo);
1598 pos = get_random_int() % freelist_count;
1600 page_limit = page->objects * s->size;
1601 start = fixup_red_left(s, page_address(page));
1603 /* First entry is used as the base of the freelist */
1604 cur = next_freelist_entry(s, page, &pos, start, page_limit,
1606 cur = setup_object(s, page, cur);
1607 page->freelist = cur;
1609 for (idx = 1; idx < page->objects; idx++) {
1610 next = next_freelist_entry(s, page, &pos, start, page_limit,
1612 next = setup_object(s, page, next);
1613 set_freepointer(s, cur, next);
1616 set_freepointer(s, cur, NULL);
1621 static inline int init_cache_random_seq(struct kmem_cache *s)
1625 static inline void init_freelist_randomization(void) { }
1626 static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1630 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1632 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1635 struct kmem_cache_order_objects oo = s->oo;
1637 void *start, *p, *next;
1641 flags &= gfp_allowed_mask;
1643 if (gfpflags_allow_blocking(flags))
1646 flags |= s->allocflags;
1649 * Let the initial higher-order allocation fail under memory pressure
1650 * so we fall-back to the minimum order allocation.
1652 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1653 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1654 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
1656 page = alloc_slab_page(s, alloc_gfp, node, oo);
1657 if (unlikely(!page)) {
1661 * Allocation may have failed due to fragmentation.
1662 * Try a lower order alloc if possible
1664 page = alloc_slab_page(s, alloc_gfp, node, oo);
1665 if (unlikely(!page))
1667 stat(s, ORDER_FALLBACK);
1670 page->objects = oo_objects(oo);
1672 order = compound_order(page);
1673 page->slab_cache = s;
1674 __SetPageSlab(page);
1675 if (page_is_pfmemalloc(page))
1676 SetPageSlabPfmemalloc(page);
1678 kasan_poison_slab(page);
1680 start = page_address(page);
1682 setup_page_debug(s, start, order);
1684 shuffle = shuffle_freelist(s, page);
1687 start = fixup_red_left(s, start);
1688 start = setup_object(s, page, start);
1689 page->freelist = start;
1690 for (idx = 0, p = start; idx < page->objects - 1; idx++) {
1692 next = setup_object(s, page, next);
1693 set_freepointer(s, p, next);
1696 set_freepointer(s, p, NULL);
1699 page->inuse = page->objects;
1703 if (gfpflags_allow_blocking(flags))
1704 local_irq_disable();
1708 inc_slabs_node(s, page_to_nid(page), page->objects);
1713 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1715 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1716 gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK;
1717 flags &= ~GFP_SLAB_BUG_MASK;
1718 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
1719 invalid_mask, &invalid_mask, flags, &flags);
1723 return allocate_slab(s,
1724 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1727 static void __free_slab(struct kmem_cache *s, struct page *page)
1729 int order = compound_order(page);
1730 int pages = 1 << order;
1732 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1735 slab_pad_check(s, page);
1736 for_each_object(p, s, page_address(page),
1738 check_object(s, page, p, SLUB_RED_INACTIVE);
1741 __ClearPageSlabPfmemalloc(page);
1742 __ClearPageSlab(page);
1744 page->mapping = NULL;
1745 if (current->reclaim_state)
1746 current->reclaim_state->reclaimed_slab += pages;
1747 uncharge_slab_page(page, order, s);
1748 __free_pages(page, order);
1751 static void rcu_free_slab(struct rcu_head *h)
1753 struct page *page = container_of(h, struct page, rcu_head);
1755 __free_slab(page->slab_cache, page);
1758 static void free_slab(struct kmem_cache *s, struct page *page)
1760 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) {
1761 call_rcu(&page->rcu_head, rcu_free_slab);
1763 __free_slab(s, page);
1766 static void discard_slab(struct kmem_cache *s, struct page *page)
1768 dec_slabs_node(s, page_to_nid(page), page->objects);
1773 * Management of partially allocated slabs.
1776 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1779 if (tail == DEACTIVATE_TO_TAIL)
1780 list_add_tail(&page->slab_list, &n->partial);
1782 list_add(&page->slab_list, &n->partial);
1785 static inline void add_partial(struct kmem_cache_node *n,
1786 struct page *page, int tail)
1788 lockdep_assert_held(&n->list_lock);
1789 __add_partial(n, page, tail);
1792 static inline void remove_partial(struct kmem_cache_node *n,
1795 lockdep_assert_held(&n->list_lock);
1796 list_del(&page->slab_list);
1801 * Remove slab from the partial list, freeze it and
1802 * return the pointer to the freelist.
1804 * Returns a list of objects or NULL if it fails.
1806 static inline void *acquire_slab(struct kmem_cache *s,
1807 struct kmem_cache_node *n, struct page *page,
1808 int mode, int *objects)
1811 unsigned long counters;
1814 lockdep_assert_held(&n->list_lock);
1817 * Zap the freelist and set the frozen bit.
1818 * The old freelist is the list of objects for the
1819 * per cpu allocation list.
1821 freelist = page->freelist;
1822 counters = page->counters;
1823 new.counters = counters;
1824 *objects = new.objects - new.inuse;
1826 new.inuse = page->objects;
1827 new.freelist = NULL;
1829 new.freelist = freelist;
1832 VM_BUG_ON(new.frozen);
1835 if (!__cmpxchg_double_slab(s, page,
1837 new.freelist, new.counters,
1841 remove_partial(n, page);
1846 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1847 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1850 * Try to allocate a partial slab from a specific node.
1852 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1853 struct kmem_cache_cpu *c, gfp_t flags)
1855 struct page *page, *page2;
1856 void *object = NULL;
1857 unsigned int available = 0;
1861 * Racy check. If we mistakenly see no partial slabs then we
1862 * just allocate an empty slab. If we mistakenly try to get a
1863 * partial slab and there is none available then get_partials()
1866 if (!n || !n->nr_partial)
1869 spin_lock(&n->list_lock);
1870 list_for_each_entry_safe(page, page2, &n->partial, slab_list) {
1873 if (!pfmemalloc_match(page, flags))
1876 t = acquire_slab(s, n, page, object == NULL, &objects);
1880 available += objects;
1883 stat(s, ALLOC_FROM_PARTIAL);
1886 put_cpu_partial(s, page, 0);
1887 stat(s, CPU_PARTIAL_NODE);
1889 if (!kmem_cache_has_cpu_partial(s)
1890 || available > slub_cpu_partial(s) / 2)
1894 spin_unlock(&n->list_lock);
1899 * Get a page from somewhere. Search in increasing NUMA distances.
1901 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1902 struct kmem_cache_cpu *c)
1905 struct zonelist *zonelist;
1908 enum zone_type high_zoneidx = gfp_zone(flags);
1910 unsigned int cpuset_mems_cookie;
1913 * The defrag ratio allows a configuration of the tradeoffs between
1914 * inter node defragmentation and node local allocations. A lower
1915 * defrag_ratio increases the tendency to do local allocations
1916 * instead of attempting to obtain partial slabs from other nodes.
1918 * If the defrag_ratio is set to 0 then kmalloc() always
1919 * returns node local objects. If the ratio is higher then kmalloc()
1920 * may return off node objects because partial slabs are obtained
1921 * from other nodes and filled up.
1923 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
1924 * (which makes defrag_ratio = 1000) then every (well almost)
1925 * allocation will first attempt to defrag slab caches on other nodes.
1926 * This means scanning over all nodes to look for partial slabs which
1927 * may be expensive if we do it every time we are trying to find a slab
1928 * with available objects.
1930 if (!s->remote_node_defrag_ratio ||
1931 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1935 cpuset_mems_cookie = read_mems_allowed_begin();
1936 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1937 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1938 struct kmem_cache_node *n;
1940 n = get_node(s, zone_to_nid(zone));
1942 if (n && cpuset_zone_allowed(zone, flags) &&
1943 n->nr_partial > s->min_partial) {
1944 object = get_partial_node(s, n, c, flags);
1947 * Don't check read_mems_allowed_retry()
1948 * here - if mems_allowed was updated in
1949 * parallel, that was a harmless race
1950 * between allocation and the cpuset
1957 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1958 #endif /* CONFIG_NUMA */
1963 * Get a partial page, lock it and return it.
1965 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1966 struct kmem_cache_cpu *c)
1969 int searchnode = node;
1971 if (node == NUMA_NO_NODE)
1972 searchnode = numa_mem_id();
1973 else if (!node_present_pages(node))
1974 searchnode = node_to_mem_node(node);
1976 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1977 if (object || node != NUMA_NO_NODE)
1980 return get_any_partial(s, flags, c);
1983 #ifdef CONFIG_PREEMPT
1985 * Calculate the next globally unique transaction for disambiguiation
1986 * during cmpxchg. The transactions start with the cpu number and are then
1987 * incremented by CONFIG_NR_CPUS.
1989 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1992 * No preemption supported therefore also no need to check for
1998 static inline unsigned long next_tid(unsigned long tid)
2000 return tid + TID_STEP;
2003 static inline unsigned int tid_to_cpu(unsigned long tid)
2005 return tid % TID_STEP;
2008 static inline unsigned long tid_to_event(unsigned long tid)
2010 return tid / TID_STEP;
2013 static inline unsigned int init_tid(int cpu)
2018 static inline void note_cmpxchg_failure(const char *n,
2019 const struct kmem_cache *s, unsigned long tid)
2021 #ifdef SLUB_DEBUG_CMPXCHG
2022 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
2024 pr_info("%s %s: cmpxchg redo ", n, s->name);
2026 #ifdef CONFIG_PREEMPT
2027 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
2028 pr_warn("due to cpu change %d -> %d\n",
2029 tid_to_cpu(tid), tid_to_cpu(actual_tid));
2032 if (tid_to_event(tid) != tid_to_event(actual_tid))
2033 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2034 tid_to_event(tid), tid_to_event(actual_tid));
2036 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2037 actual_tid, tid, next_tid(tid));
2039 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2042 static void init_kmem_cache_cpus(struct kmem_cache *s)
2046 for_each_possible_cpu(cpu)
2047 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
2051 * Remove the cpu slab
2053 static void deactivate_slab(struct kmem_cache *s, struct page *page,
2054 void *freelist, struct kmem_cache_cpu *c)
2056 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
2057 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
2059 enum slab_modes l = M_NONE, m = M_NONE;
2061 int tail = DEACTIVATE_TO_HEAD;
2065 if (page->freelist) {
2066 stat(s, DEACTIVATE_REMOTE_FREES);
2067 tail = DEACTIVATE_TO_TAIL;
2071 * Stage one: Free all available per cpu objects back
2072 * to the page freelist while it is still frozen. Leave the
2075 * There is no need to take the list->lock because the page
2078 while (freelist && (nextfree = get_freepointer(s, freelist))) {
2080 unsigned long counters;
2083 prior = page->freelist;
2084 counters = page->counters;
2085 set_freepointer(s, freelist, prior);
2086 new.counters = counters;
2088 VM_BUG_ON(!new.frozen);
2090 } while (!__cmpxchg_double_slab(s, page,
2092 freelist, new.counters,
2093 "drain percpu freelist"));
2095 freelist = nextfree;
2099 * Stage two: Ensure that the page is unfrozen while the
2100 * list presence reflects the actual number of objects
2103 * We setup the list membership and then perform a cmpxchg
2104 * with the count. If there is a mismatch then the page
2105 * is not unfrozen but the page is on the wrong list.
2107 * Then we restart the process which may have to remove
2108 * the page from the list that we just put it on again
2109 * because the number of objects in the slab may have
2114 old.freelist = page->freelist;
2115 old.counters = page->counters;
2116 VM_BUG_ON(!old.frozen);
2118 /* Determine target state of the slab */
2119 new.counters = old.counters;
2122 set_freepointer(s, freelist, old.freelist);
2123 new.freelist = freelist;
2125 new.freelist = old.freelist;
2129 if (!new.inuse && n->nr_partial >= s->min_partial)
2131 else if (new.freelist) {
2136 * Taking the spinlock removes the possibility
2137 * that acquire_slab() will see a slab page that
2140 spin_lock(&n->list_lock);
2144 if (kmem_cache_debug(s) && !lock) {
2147 * This also ensures that the scanning of full
2148 * slabs from diagnostic functions will not see
2151 spin_lock(&n->list_lock);
2157 remove_partial(n, page);
2158 else if (l == M_FULL)
2159 remove_full(s, n, page);
2162 add_partial(n, page, tail);
2163 else if (m == M_FULL)
2164 add_full(s, n, page);
2168 if (!__cmpxchg_double_slab(s, page,
2169 old.freelist, old.counters,
2170 new.freelist, new.counters,
2175 spin_unlock(&n->list_lock);
2179 else if (m == M_FULL)
2180 stat(s, DEACTIVATE_FULL);
2181 else if (m == M_FREE) {
2182 stat(s, DEACTIVATE_EMPTY);
2183 discard_slab(s, page);
2192 * Unfreeze all the cpu partial slabs.
2194 * This function must be called with interrupts disabled
2195 * for the cpu using c (or some other guarantee must be there
2196 * to guarantee no concurrent accesses).
2198 static void unfreeze_partials(struct kmem_cache *s,
2199 struct kmem_cache_cpu *c)
2201 #ifdef CONFIG_SLUB_CPU_PARTIAL
2202 struct kmem_cache_node *n = NULL, *n2 = NULL;
2203 struct page *page, *discard_page = NULL;
2205 while ((page = c->partial)) {
2209 c->partial = page->next;
2211 n2 = get_node(s, page_to_nid(page));
2214 spin_unlock(&n->list_lock);
2217 spin_lock(&n->list_lock);
2222 old.freelist = page->freelist;
2223 old.counters = page->counters;
2224 VM_BUG_ON(!old.frozen);
2226 new.counters = old.counters;
2227 new.freelist = old.freelist;
2231 } while (!__cmpxchg_double_slab(s, page,
2232 old.freelist, old.counters,
2233 new.freelist, new.counters,
2234 "unfreezing slab"));
2236 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2237 page->next = discard_page;
2238 discard_page = page;
2240 add_partial(n, page, DEACTIVATE_TO_TAIL);
2241 stat(s, FREE_ADD_PARTIAL);
2246 spin_unlock(&n->list_lock);
2248 while (discard_page) {
2249 page = discard_page;
2250 discard_page = discard_page->next;
2252 stat(s, DEACTIVATE_EMPTY);
2253 discard_slab(s, page);
2256 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2260 * Put a page that was just frozen (in __slab_free|get_partial_node) into a
2261 * partial page slot if available.
2263 * If we did not find a slot then simply move all the partials to the
2264 * per node partial list.
2266 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2268 #ifdef CONFIG_SLUB_CPU_PARTIAL
2269 struct page *oldpage;
2277 oldpage = this_cpu_read(s->cpu_slab->partial);
2280 pobjects = oldpage->pobjects;
2281 pages = oldpage->pages;
2282 if (drain && pobjects > s->cpu_partial) {
2283 unsigned long flags;
2285 * partial array is full. Move the existing
2286 * set to the per node partial list.
2288 local_irq_save(flags);
2289 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2290 local_irq_restore(flags);
2294 stat(s, CPU_PARTIAL_DRAIN);
2299 pobjects += page->objects - page->inuse;
2301 page->pages = pages;
2302 page->pobjects = pobjects;
2303 page->next = oldpage;
2305 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2307 if (unlikely(!s->cpu_partial)) {
2308 unsigned long flags;
2310 local_irq_save(flags);
2311 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2312 local_irq_restore(flags);
2315 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2318 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2320 stat(s, CPUSLAB_FLUSH);
2321 deactivate_slab(s, c->page, c->freelist, c);
2323 c->tid = next_tid(c->tid);
2329 * Called from IPI handler with interrupts disabled.
2331 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2333 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2338 unfreeze_partials(s, c);
2341 static void flush_cpu_slab(void *d)
2343 struct kmem_cache *s = d;
2345 __flush_cpu_slab(s, smp_processor_id());
2348 static bool has_cpu_slab(int cpu, void *info)
2350 struct kmem_cache *s = info;
2351 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2353 return c->page || slub_percpu_partial(c);
2356 static void flush_all(struct kmem_cache *s)
2358 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2362 * Use the cpu notifier to insure that the cpu slabs are flushed when
2365 static int slub_cpu_dead(unsigned int cpu)
2367 struct kmem_cache *s;
2368 unsigned long flags;
2370 mutex_lock(&slab_mutex);
2371 list_for_each_entry(s, &slab_caches, list) {
2372 local_irq_save(flags);
2373 __flush_cpu_slab(s, cpu);
2374 local_irq_restore(flags);
2376 mutex_unlock(&slab_mutex);
2381 * Check if the objects in a per cpu structure fit numa
2382 * locality expectations.
2384 static inline int node_match(struct page *page, int node)
2387 if (node != NUMA_NO_NODE && page_to_nid(page) != node)
2393 #ifdef CONFIG_SLUB_DEBUG
2394 static int count_free(struct page *page)
2396 return page->objects - page->inuse;
2399 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2401 return atomic_long_read(&n->total_objects);
2403 #endif /* CONFIG_SLUB_DEBUG */
2405 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2406 static unsigned long count_partial(struct kmem_cache_node *n,
2407 int (*get_count)(struct page *))
2409 unsigned long flags;
2410 unsigned long x = 0;
2413 spin_lock_irqsave(&n->list_lock, flags);
2414 list_for_each_entry(page, &n->partial, slab_list)
2415 x += get_count(page);
2416 spin_unlock_irqrestore(&n->list_lock, flags);
2419 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2421 static noinline void
2422 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2424 #ifdef CONFIG_SLUB_DEBUG
2425 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2426 DEFAULT_RATELIMIT_BURST);
2428 struct kmem_cache_node *n;
2430 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2433 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2434 nid, gfpflags, &gfpflags);
2435 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2436 s->name, s->object_size, s->size, oo_order(s->oo),
2439 if (oo_order(s->min) > get_order(s->object_size))
2440 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2443 for_each_kmem_cache_node(s, node, n) {
2444 unsigned long nr_slabs;
2445 unsigned long nr_objs;
2446 unsigned long nr_free;
2448 nr_free = count_partial(n, count_free);
2449 nr_slabs = node_nr_slabs(n);
2450 nr_objs = node_nr_objs(n);
2452 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2453 node, nr_slabs, nr_objs, nr_free);
2458 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2459 int node, struct kmem_cache_cpu **pc)
2462 struct kmem_cache_cpu *c = *pc;
2465 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
2467 freelist = get_partial(s, flags, node, c);
2472 page = new_slab(s, flags, node);
2474 c = raw_cpu_ptr(s->cpu_slab);
2479 * No other reference to the page yet so we can
2480 * muck around with it freely without cmpxchg
2482 freelist = page->freelist;
2483 page->freelist = NULL;
2485 stat(s, ALLOC_SLAB);
2493 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2495 if (unlikely(PageSlabPfmemalloc(page)))
2496 return gfp_pfmemalloc_allowed(gfpflags);
2502 * Check the page->freelist of a page and either transfer the freelist to the
2503 * per cpu freelist or deactivate the page.
2505 * The page is still frozen if the return value is not NULL.
2507 * If this function returns NULL then the page has been unfrozen.
2509 * This function must be called with interrupt disabled.
2511 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2514 unsigned long counters;
2518 freelist = page->freelist;
2519 counters = page->counters;
2521 new.counters = counters;
2522 VM_BUG_ON(!new.frozen);
2524 new.inuse = page->objects;
2525 new.frozen = freelist != NULL;
2527 } while (!__cmpxchg_double_slab(s, page,
2536 * Slow path. The lockless freelist is empty or we need to perform
2539 * Processing is still very fast if new objects have been freed to the
2540 * regular freelist. In that case we simply take over the regular freelist
2541 * as the lockless freelist and zap the regular freelist.
2543 * If that is not working then we fall back to the partial lists. We take the
2544 * first element of the freelist as the object to allocate now and move the
2545 * rest of the freelist to the lockless freelist.
2547 * And if we were unable to get a new slab from the partial slab lists then
2548 * we need to allocate a new slab. This is the slowest path since it involves
2549 * a call to the page allocator and the setup of a new slab.
2551 * Version of __slab_alloc to use when we know that interrupts are
2552 * already disabled (which is the case for bulk allocation).
2554 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2555 unsigned long addr, struct kmem_cache_cpu *c)
2565 if (unlikely(!node_match(page, node))) {
2566 int searchnode = node;
2568 if (node != NUMA_NO_NODE && !node_present_pages(node))
2569 searchnode = node_to_mem_node(node);
2571 if (unlikely(!node_match(page, searchnode))) {
2572 stat(s, ALLOC_NODE_MISMATCH);
2573 deactivate_slab(s, page, c->freelist, c);
2579 * By rights, we should be searching for a slab page that was
2580 * PFMEMALLOC but right now, we are losing the pfmemalloc
2581 * information when the page leaves the per-cpu allocator
2583 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2584 deactivate_slab(s, page, c->freelist, c);
2588 /* must check again c->freelist in case of cpu migration or IRQ */
2589 freelist = c->freelist;
2593 freelist = get_freelist(s, page);
2597 stat(s, DEACTIVATE_BYPASS);
2601 stat(s, ALLOC_REFILL);
2605 * freelist is pointing to the list of objects to be used.
2606 * page is pointing to the page from which the objects are obtained.
2607 * That page must be frozen for per cpu allocations to work.
2609 VM_BUG_ON(!c->page->frozen);
2610 c->freelist = get_freepointer(s, freelist);
2611 c->tid = next_tid(c->tid);
2616 if (slub_percpu_partial(c)) {
2617 page = c->page = slub_percpu_partial(c);
2618 slub_set_percpu_partial(c, page);
2619 stat(s, CPU_PARTIAL_ALLOC);
2623 freelist = new_slab_objects(s, gfpflags, node, &c);
2625 if (unlikely(!freelist)) {
2626 slab_out_of_memory(s, gfpflags, node);
2631 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2634 /* Only entered in the debug case */
2635 if (kmem_cache_debug(s) &&
2636 !alloc_debug_processing(s, page, freelist, addr))
2637 goto new_slab; /* Slab failed checks. Next slab needed */
2639 deactivate_slab(s, page, get_freepointer(s, freelist), c);
2644 * Another one that disabled interrupt and compensates for possible
2645 * cpu changes by refetching the per cpu area pointer.
2647 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2648 unsigned long addr, struct kmem_cache_cpu *c)
2651 unsigned long flags;
2653 local_irq_save(flags);
2654 #ifdef CONFIG_PREEMPT
2656 * We may have been preempted and rescheduled on a different
2657 * cpu before disabling interrupts. Need to reload cpu area
2660 c = this_cpu_ptr(s->cpu_slab);
2663 p = ___slab_alloc(s, gfpflags, node, addr, c);
2664 local_irq_restore(flags);
2669 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2670 * have the fastpath folded into their functions. So no function call
2671 * overhead for requests that can be satisfied on the fastpath.
2673 * The fastpath works by first checking if the lockless freelist can be used.
2674 * If not then __slab_alloc is called for slow processing.
2676 * Otherwise we can simply pick the next object from the lockless free list.
2678 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2679 gfp_t gfpflags, int node, unsigned long addr)
2682 struct kmem_cache_cpu *c;
2686 s = slab_pre_alloc_hook(s, gfpflags);
2691 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2692 * enabled. We may switch back and forth between cpus while
2693 * reading from one cpu area. That does not matter as long
2694 * as we end up on the original cpu again when doing the cmpxchg.
2696 * We should guarantee that tid and kmem_cache are retrieved on
2697 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2698 * to check if it is matched or not.
2701 tid = this_cpu_read(s->cpu_slab->tid);
2702 c = raw_cpu_ptr(s->cpu_slab);
2703 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2704 unlikely(tid != READ_ONCE(c->tid)));
2707 * Irqless object alloc/free algorithm used here depends on sequence
2708 * of fetching cpu_slab's data. tid should be fetched before anything
2709 * on c to guarantee that object and page associated with previous tid
2710 * won't be used with current tid. If we fetch tid first, object and
2711 * page could be one associated with next tid and our alloc/free
2712 * request will be failed. In this case, we will retry. So, no problem.
2717 * The transaction ids are globally unique per cpu and per operation on
2718 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2719 * occurs on the right processor and that there was no operation on the
2720 * linked list in between.
2723 object = c->freelist;
2725 if (unlikely(!object || !node_match(page, node))) {
2726 object = __slab_alloc(s, gfpflags, node, addr, c);
2727 stat(s, ALLOC_SLOWPATH);
2729 void *next_object = get_freepointer_safe(s, object);
2732 * The cmpxchg will only match if there was no additional
2733 * operation and if we are on the right processor.
2735 * The cmpxchg does the following atomically (without lock
2737 * 1. Relocate first pointer to the current per cpu area.
2738 * 2. Verify that tid and freelist have not been changed
2739 * 3. If they were not changed replace tid and freelist
2741 * Since this is without lock semantics the protection is only
2742 * against code executing on this cpu *not* from access by
2745 if (unlikely(!this_cpu_cmpxchg_double(
2746 s->cpu_slab->freelist, s->cpu_slab->tid,
2748 next_object, next_tid(tid)))) {
2750 note_cmpxchg_failure("slab_alloc", s, tid);
2753 prefetch_freepointer(s, next_object);
2754 stat(s, ALLOC_FASTPATH);
2757 * If the object has been wiped upon free, make sure it's fully
2758 * initialized by zeroing out freelist pointer.
2760 if (unlikely(slab_want_init_on_free(s)) && object)
2761 memset(object + s->offset, 0, sizeof(void *));
2763 if (unlikely(slab_want_init_on_alloc(gfpflags, s)) && object)
2764 memset(object, 0, s->object_size);
2766 slab_post_alloc_hook(s, gfpflags, 1, &object);
2771 static __always_inline void *slab_alloc(struct kmem_cache *s,
2772 gfp_t gfpflags, unsigned long addr)
2774 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2777 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2779 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2781 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2786 EXPORT_SYMBOL(kmem_cache_alloc);
2788 #ifdef CONFIG_TRACING
2789 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2791 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2792 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2793 ret = kasan_kmalloc(s, ret, size, gfpflags);
2796 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2800 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2802 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2804 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2805 s->object_size, s->size, gfpflags, node);
2809 EXPORT_SYMBOL(kmem_cache_alloc_node);
2811 #ifdef CONFIG_TRACING
2812 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2814 int node, size_t size)
2816 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2818 trace_kmalloc_node(_RET_IP_, ret,
2819 size, s->size, gfpflags, node);
2821 ret = kasan_kmalloc(s, ret, size, gfpflags);
2824 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2826 #endif /* CONFIG_NUMA */
2829 * Slow path handling. This may still be called frequently since objects
2830 * have a longer lifetime than the cpu slabs in most processing loads.
2832 * So we still attempt to reduce cache line usage. Just take the slab
2833 * lock and free the item. If there is no additional partial page
2834 * handling required then we can return immediately.
2836 static void __slab_free(struct kmem_cache *s, struct page *page,
2837 void *head, void *tail, int cnt,
2844 unsigned long counters;
2845 struct kmem_cache_node *n = NULL;
2846 unsigned long uninitialized_var(flags);
2848 stat(s, FREE_SLOWPATH);
2850 if (kmem_cache_debug(s) &&
2851 !free_debug_processing(s, page, head, tail, cnt, addr))
2856 spin_unlock_irqrestore(&n->list_lock, flags);
2859 prior = page->freelist;
2860 counters = page->counters;
2861 set_freepointer(s, tail, prior);
2862 new.counters = counters;
2863 was_frozen = new.frozen;
2865 if ((!new.inuse || !prior) && !was_frozen) {
2867 if (kmem_cache_has_cpu_partial(s) && !prior) {
2870 * Slab was on no list before and will be
2872 * We can defer the list move and instead
2877 } else { /* Needs to be taken off a list */
2879 n = get_node(s, page_to_nid(page));
2881 * Speculatively acquire the list_lock.
2882 * If the cmpxchg does not succeed then we may
2883 * drop the list_lock without any processing.
2885 * Otherwise the list_lock will synchronize with
2886 * other processors updating the list of slabs.
2888 spin_lock_irqsave(&n->list_lock, flags);
2893 } while (!cmpxchg_double_slab(s, page,
2901 * If we just froze the page then put it onto the
2902 * per cpu partial list.
2904 if (new.frozen && !was_frozen) {
2905 put_cpu_partial(s, page, 1);
2906 stat(s, CPU_PARTIAL_FREE);
2909 * The list lock was not taken therefore no list
2910 * activity can be necessary.
2913 stat(s, FREE_FROZEN);
2917 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2921 * Objects left in the slab. If it was not on the partial list before
2924 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2925 remove_full(s, n, page);
2926 add_partial(n, page, DEACTIVATE_TO_TAIL);
2927 stat(s, FREE_ADD_PARTIAL);
2929 spin_unlock_irqrestore(&n->list_lock, flags);
2935 * Slab on the partial list.
2937 remove_partial(n, page);
2938 stat(s, FREE_REMOVE_PARTIAL);
2940 /* Slab must be on the full list */
2941 remove_full(s, n, page);
2944 spin_unlock_irqrestore(&n->list_lock, flags);
2946 discard_slab(s, page);
2950 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2951 * can perform fastpath freeing without additional function calls.
2953 * The fastpath is only possible if we are freeing to the current cpu slab
2954 * of this processor. This typically the case if we have just allocated
2957 * If fastpath is not possible then fall back to __slab_free where we deal
2958 * with all sorts of special processing.
2960 * Bulk free of a freelist with several objects (all pointing to the
2961 * same page) possible by specifying head and tail ptr, plus objects
2962 * count (cnt). Bulk free indicated by tail pointer being set.
2964 static __always_inline void do_slab_free(struct kmem_cache *s,
2965 struct page *page, void *head, void *tail,
2966 int cnt, unsigned long addr)
2968 void *tail_obj = tail ? : head;
2969 struct kmem_cache_cpu *c;
2973 * Determine the currently cpus per cpu slab.
2974 * The cpu may change afterward. However that does not matter since
2975 * data is retrieved via this pointer. If we are on the same cpu
2976 * during the cmpxchg then the free will succeed.
2979 tid = this_cpu_read(s->cpu_slab->tid);
2980 c = raw_cpu_ptr(s->cpu_slab);
2981 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2982 unlikely(tid != READ_ONCE(c->tid)));
2984 /* Same with comment on barrier() in slab_alloc_node() */
2987 if (likely(page == c->page)) {
2988 set_freepointer(s, tail_obj, c->freelist);
2990 if (unlikely(!this_cpu_cmpxchg_double(
2991 s->cpu_slab->freelist, s->cpu_slab->tid,
2993 head, next_tid(tid)))) {
2995 note_cmpxchg_failure("slab_free", s, tid);
2998 stat(s, FREE_FASTPATH);
3000 __slab_free(s, page, head, tail_obj, cnt, addr);
3004 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
3005 void *head, void *tail, int cnt,
3009 * With KASAN enabled slab_free_freelist_hook modifies the freelist
3010 * to remove objects, whose reuse must be delayed.
3012 if (slab_free_freelist_hook(s, &head, &tail))
3013 do_slab_free(s, page, head, tail, cnt, addr);
3016 #ifdef CONFIG_KASAN_GENERIC
3017 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
3019 do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr);
3023 void kmem_cache_free(struct kmem_cache *s, void *x)
3025 s = cache_from_obj(s, x);
3028 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
3029 trace_kmem_cache_free(_RET_IP_, x);
3031 EXPORT_SYMBOL(kmem_cache_free);
3033 struct detached_freelist {
3038 struct kmem_cache *s;
3042 * This function progressively scans the array with free objects (with
3043 * a limited look ahead) and extract objects belonging to the same
3044 * page. It builds a detached freelist directly within the given
3045 * page/objects. This can happen without any need for
3046 * synchronization, because the objects are owned by running process.
3047 * The freelist is build up as a single linked list in the objects.
3048 * The idea is, that this detached freelist can then be bulk
3049 * transferred to the real freelist(s), but only requiring a single
3050 * synchronization primitive. Look ahead in the array is limited due
3051 * to performance reasons.
3054 int build_detached_freelist(struct kmem_cache *s, size_t size,
3055 void **p, struct detached_freelist *df)
3057 size_t first_skipped_index = 0;
3062 /* Always re-init detached_freelist */
3067 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3068 } while (!object && size);
3073 page = virt_to_head_page(object);
3075 /* Handle kalloc'ed objects */
3076 if (unlikely(!PageSlab(page))) {
3077 BUG_ON(!PageCompound(page));
3079 __free_pages(page, compound_order(page));
3080 p[size] = NULL; /* mark object processed */
3083 /* Derive kmem_cache from object */
3084 df->s = page->slab_cache;
3086 df->s = cache_from_obj(s, object); /* Support for memcg */
3089 /* Start new detached freelist */
3091 set_freepointer(df->s, object, NULL);
3093 df->freelist = object;
3094 p[size] = NULL; /* mark object processed */
3100 continue; /* Skip processed objects */
3102 /* df->page is always set at this point */
3103 if (df->page == virt_to_head_page(object)) {
3104 /* Opportunity build freelist */
3105 set_freepointer(df->s, object, df->freelist);
3106 df->freelist = object;
3108 p[size] = NULL; /* mark object processed */
3113 /* Limit look ahead search */
3117 if (!first_skipped_index)
3118 first_skipped_index = size + 1;
3121 return first_skipped_index;
3124 /* Note that interrupts must be enabled when calling this function. */
3125 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3131 struct detached_freelist df;
3133 size = build_detached_freelist(s, size, p, &df);
3137 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
3138 } while (likely(size));
3140 EXPORT_SYMBOL(kmem_cache_free_bulk);
3142 /* Note that interrupts must be enabled when calling this function. */
3143 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3146 struct kmem_cache_cpu *c;
3149 /* memcg and kmem_cache debug support */
3150 s = slab_pre_alloc_hook(s, flags);
3154 * Drain objects in the per cpu slab, while disabling local
3155 * IRQs, which protects against PREEMPT and interrupts
3156 * handlers invoking normal fastpath.
3158 local_irq_disable();
3159 c = this_cpu_ptr(s->cpu_slab);
3161 for (i = 0; i < size; i++) {
3162 void *object = c->freelist;
3164 if (unlikely(!object)) {
3166 * Invoking slow path likely have side-effect
3167 * of re-populating per CPU c->freelist
3169 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3171 if (unlikely(!p[i]))
3174 c = this_cpu_ptr(s->cpu_slab);
3175 continue; /* goto for-loop */
3177 c->freelist = get_freepointer(s, object);
3180 c->tid = next_tid(c->tid);
3183 /* Clear memory outside IRQ disabled fastpath loop */
3184 if (unlikely(slab_want_init_on_alloc(flags, s))) {
3187 for (j = 0; j < i; j++)
3188 memset(p[j], 0, s->object_size);
3191 /* memcg and kmem_cache debug support */
3192 slab_post_alloc_hook(s, flags, size, p);
3196 slab_post_alloc_hook(s, flags, i, p);
3197 __kmem_cache_free_bulk(s, i, p);
3200 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3204 * Object placement in a slab is made very easy because we always start at
3205 * offset 0. If we tune the size of the object to the alignment then we can
3206 * get the required alignment by putting one properly sized object after
3209 * Notice that the allocation order determines the sizes of the per cpu
3210 * caches. Each processor has always one slab available for allocations.
3211 * Increasing the allocation order reduces the number of times that slabs
3212 * must be moved on and off the partial lists and is therefore a factor in
3217 * Mininum / Maximum order of slab pages. This influences locking overhead
3218 * and slab fragmentation. A higher order reduces the number of partial slabs
3219 * and increases the number of allocations possible without having to
3220 * take the list_lock.
3222 static unsigned int slub_min_order;
3223 static unsigned int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3224 static unsigned int slub_min_objects;
3227 * Calculate the order of allocation given an slab object size.
3229 * The order of allocation has significant impact on performance and other
3230 * system components. Generally order 0 allocations should be preferred since
3231 * order 0 does not cause fragmentation in the page allocator. Larger objects
3232 * be problematic to put into order 0 slabs because there may be too much
3233 * unused space left. We go to a higher order if more than 1/16th of the slab
3236 * In order to reach satisfactory performance we must ensure that a minimum
3237 * number of objects is in one slab. Otherwise we may generate too much
3238 * activity on the partial lists which requires taking the list_lock. This is
3239 * less a concern for large slabs though which are rarely used.
3241 * slub_max_order specifies the order where we begin to stop considering the
3242 * number of objects in a slab as critical. If we reach slub_max_order then
3243 * we try to keep the page order as low as possible. So we accept more waste
3244 * of space in favor of a small page order.
3246 * Higher order allocations also allow the placement of more objects in a
3247 * slab and thereby reduce object handling overhead. If the user has
3248 * requested a higher mininum order then we start with that one instead of
3249 * the smallest order which will fit the object.
3251 static inline unsigned int slab_order(unsigned int size,
3252 unsigned int min_objects, unsigned int max_order,
3253 unsigned int fract_leftover)
3255 unsigned int min_order = slub_min_order;
3258 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
3259 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3261 for (order = max(min_order, (unsigned int)get_order(min_objects * size));
3262 order <= max_order; order++) {
3264 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
3267 rem = slab_size % size;
3269 if (rem <= slab_size / fract_leftover)
3276 static inline int calculate_order(unsigned int size)
3279 unsigned int min_objects;
3280 unsigned int max_objects;
3283 * Attempt to find best configuration for a slab. This
3284 * works by first attempting to generate a layout with
3285 * the best configuration and backing off gradually.
3287 * First we increase the acceptable waste in a slab. Then
3288 * we reduce the minimum objects required in a slab.
3290 min_objects = slub_min_objects;
3292 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3293 max_objects = order_objects(slub_max_order, size);
3294 min_objects = min(min_objects, max_objects);
3296 while (min_objects > 1) {
3297 unsigned int fraction;
3300 while (fraction >= 4) {
3301 order = slab_order(size, min_objects,
3302 slub_max_order, fraction);
3303 if (order <= slub_max_order)
3311 * We were unable to place multiple objects in a slab. Now
3312 * lets see if we can place a single object there.
3314 order = slab_order(size, 1, slub_max_order, 1);
3315 if (order <= slub_max_order)
3319 * Doh this slab cannot be placed using slub_max_order.
3321 order = slab_order(size, 1, MAX_ORDER, 1);
3322 if (order < MAX_ORDER)
3328 init_kmem_cache_node(struct kmem_cache_node *n)
3331 spin_lock_init(&n->list_lock);
3332 INIT_LIST_HEAD(&n->partial);
3333 #ifdef CONFIG_SLUB_DEBUG
3334 atomic_long_set(&n->nr_slabs, 0);
3335 atomic_long_set(&n->total_objects, 0);
3336 INIT_LIST_HEAD(&n->full);
3340 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3342 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3343 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3346 * Must align to double word boundary for the double cmpxchg
3347 * instructions to work; see __pcpu_double_call_return_bool().
3349 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3350 2 * sizeof(void *));
3355 init_kmem_cache_cpus(s);
3360 static struct kmem_cache *kmem_cache_node;
3363 * No kmalloc_node yet so do it by hand. We know that this is the first
3364 * slab on the node for this slabcache. There are no concurrent accesses
3367 * Note that this function only works on the kmem_cache_node
3368 * when allocating for the kmem_cache_node. This is used for bootstrapping
3369 * memory on a fresh node that has no slab structures yet.
3371 static void early_kmem_cache_node_alloc(int node)
3374 struct kmem_cache_node *n;
3376 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3378 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3381 if (page_to_nid(page) != node) {
3382 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3383 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3388 #ifdef CONFIG_SLUB_DEBUG
3389 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3390 init_tracking(kmem_cache_node, n);
3392 n = kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node),
3394 page->freelist = get_freepointer(kmem_cache_node, n);
3397 kmem_cache_node->node[node] = n;
3398 init_kmem_cache_node(n);
3399 inc_slabs_node(kmem_cache_node, node, page->objects);
3402 * No locks need to be taken here as it has just been
3403 * initialized and there is no concurrent access.
3405 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3408 static void free_kmem_cache_nodes(struct kmem_cache *s)
3411 struct kmem_cache_node *n;
3413 for_each_kmem_cache_node(s, node, n) {
3414 s->node[node] = NULL;
3415 kmem_cache_free(kmem_cache_node, n);
3419 void __kmem_cache_release(struct kmem_cache *s)
3421 cache_random_seq_destroy(s);
3422 free_percpu(s->cpu_slab);
3423 free_kmem_cache_nodes(s);
3426 static int init_kmem_cache_nodes(struct kmem_cache *s)
3430 for_each_node_state(node, N_NORMAL_MEMORY) {
3431 struct kmem_cache_node *n;
3433 if (slab_state == DOWN) {
3434 early_kmem_cache_node_alloc(node);
3437 n = kmem_cache_alloc_node(kmem_cache_node,
3441 free_kmem_cache_nodes(s);
3445 init_kmem_cache_node(n);
3451 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3453 if (min < MIN_PARTIAL)
3455 else if (min > MAX_PARTIAL)
3457 s->min_partial = min;
3460 static void set_cpu_partial(struct kmem_cache *s)
3462 #ifdef CONFIG_SLUB_CPU_PARTIAL
3464 * cpu_partial determined the maximum number of objects kept in the
3465 * per cpu partial lists of a processor.
3467 * Per cpu partial lists mainly contain slabs that just have one
3468 * object freed. If they are used for allocation then they can be
3469 * filled up again with minimal effort. The slab will never hit the
3470 * per node partial lists and therefore no locking will be required.
3472 * This setting also determines
3474 * A) The number of objects from per cpu partial slabs dumped to the
3475 * per node list when we reach the limit.
3476 * B) The number of objects in cpu partial slabs to extract from the
3477 * per node list when we run out of per cpu objects. We only fetch
3478 * 50% to keep some capacity around for frees.
3480 if (!kmem_cache_has_cpu_partial(s))
3482 else if (s->size >= PAGE_SIZE)
3484 else if (s->size >= 1024)
3486 else if (s->size >= 256)
3487 s->cpu_partial = 13;
3489 s->cpu_partial = 30;
3494 * calculate_sizes() determines the order and the distribution of data within
3497 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3499 slab_flags_t flags = s->flags;
3500 unsigned int size = s->object_size;
3504 * Round up object size to the next word boundary. We can only
3505 * place the free pointer at word boundaries and this determines
3506 * the possible location of the free pointer.
3508 size = ALIGN(size, sizeof(void *));
3510 #ifdef CONFIG_SLUB_DEBUG
3512 * Determine if we can poison the object itself. If the user of
3513 * the slab may touch the object after free or before allocation
3514 * then we should never poison the object itself.
3516 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
3518 s->flags |= __OBJECT_POISON;
3520 s->flags &= ~__OBJECT_POISON;
3524 * If we are Redzoning then check if there is some space between the
3525 * end of the object and the free pointer. If not then add an
3526 * additional word to have some bytes to store Redzone information.
3528 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3529 size += sizeof(void *);
3533 * With that we have determined the number of bytes in actual use
3534 * by the object. This is the potential offset to the free pointer.
3538 if (((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
3541 * Relocate free pointer after the object if it is not
3542 * permitted to overwrite the first word of the object on
3545 * This is the case if we do RCU, have a constructor or
3546 * destructor or are poisoning the objects.
3549 size += sizeof(void *);
3552 #ifdef CONFIG_SLUB_DEBUG
3553 if (flags & SLAB_STORE_USER)
3555 * Need to store information about allocs and frees after
3558 size += 2 * sizeof(struct track);
3561 kasan_cache_create(s, &size, &s->flags);
3562 #ifdef CONFIG_SLUB_DEBUG
3563 if (flags & SLAB_RED_ZONE) {
3565 * Add some empty padding so that we can catch
3566 * overwrites from earlier objects rather than let
3567 * tracking information or the free pointer be
3568 * corrupted if a user writes before the start
3571 size += sizeof(void *);
3573 s->red_left_pad = sizeof(void *);
3574 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3575 size += s->red_left_pad;
3580 * SLUB stores one object immediately after another beginning from
3581 * offset 0. In order to align the objects we have to simply size
3582 * each object to conform to the alignment.
3584 size = ALIGN(size, s->align);
3586 if (forced_order >= 0)
3587 order = forced_order;
3589 order = calculate_order(size);
3596 s->allocflags |= __GFP_COMP;
3598 if (s->flags & SLAB_CACHE_DMA)
3599 s->allocflags |= GFP_DMA;
3601 if (s->flags & SLAB_CACHE_DMA32)
3602 s->allocflags |= GFP_DMA32;
3604 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3605 s->allocflags |= __GFP_RECLAIMABLE;
3608 * Determine the number of objects per slab
3610 s->oo = oo_make(order, size);
3611 s->min = oo_make(get_order(size), size);
3612 if (oo_objects(s->oo) > oo_objects(s->max))
3615 return !!oo_objects(s->oo);
3618 static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
3620 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3621 #ifdef CONFIG_SLAB_FREELIST_HARDENED
3622 s->random = get_random_long();
3625 if (!calculate_sizes(s, -1))
3627 if (disable_higher_order_debug) {
3629 * Disable debugging flags that store metadata if the min slab
3632 if (get_order(s->size) > get_order(s->object_size)) {
3633 s->flags &= ~DEBUG_METADATA_FLAGS;
3635 if (!calculate_sizes(s, -1))
3640 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3641 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3642 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
3643 /* Enable fast mode */
3644 s->flags |= __CMPXCHG_DOUBLE;
3648 * The larger the object size is, the more pages we want on the partial
3649 * list to avoid pounding the page allocator excessively.
3651 set_min_partial(s, ilog2(s->size) / 2);
3656 s->remote_node_defrag_ratio = 1000;
3659 /* Initialize the pre-computed randomized freelist if slab is up */
3660 if (slab_state >= UP) {
3661 if (init_cache_random_seq(s))
3665 if (!init_kmem_cache_nodes(s))
3668 if (alloc_kmem_cache_cpus(s))
3671 free_kmem_cache_nodes(s);
3676 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3679 #ifdef CONFIG_SLUB_DEBUG
3680 void *addr = page_address(page);
3682 unsigned long *map = bitmap_zalloc(page->objects, GFP_ATOMIC);
3685 slab_err(s, page, text, s->name);
3688 get_map(s, page, map);
3689 for_each_object(p, s, addr, page->objects) {
3691 if (!test_bit(slab_index(p, s, addr), map)) {
3692 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3693 print_tracking(s, p);
3702 * Attempt to free all partial slabs on a node.
3703 * This is called from __kmem_cache_shutdown(). We must take list_lock
3704 * because sysfs file might still access partial list after the shutdowning.
3706 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3709 struct page *page, *h;
3711 BUG_ON(irqs_disabled());
3712 spin_lock_irq(&n->list_lock);
3713 list_for_each_entry_safe(page, h, &n->partial, slab_list) {
3715 remove_partial(n, page);
3716 list_add(&page->slab_list, &discard);
3718 list_slab_objects(s, page,
3719 "Objects remaining in %s on __kmem_cache_shutdown()");
3722 spin_unlock_irq(&n->list_lock);
3724 list_for_each_entry_safe(page, h, &discard, slab_list)
3725 discard_slab(s, page);
3728 bool __kmem_cache_empty(struct kmem_cache *s)
3731 struct kmem_cache_node *n;
3733 for_each_kmem_cache_node(s, node, n)
3734 if (n->nr_partial || slabs_node(s, node))
3740 * Release all resources used by a slab cache.
3742 int __kmem_cache_shutdown(struct kmem_cache *s)
3745 struct kmem_cache_node *n;
3748 /* Attempt to free all objects */
3749 for_each_kmem_cache_node(s, node, n) {
3751 if (n->nr_partial || slabs_node(s, node))
3754 sysfs_slab_remove(s);
3758 /********************************************************************
3760 *******************************************************************/
3762 static int __init setup_slub_min_order(char *str)
3764 get_option(&str, (int *)&slub_min_order);
3769 __setup("slub_min_order=", setup_slub_min_order);
3771 static int __init setup_slub_max_order(char *str)
3773 get_option(&str, (int *)&slub_max_order);
3774 slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1);
3779 __setup("slub_max_order=", setup_slub_max_order);
3781 static int __init setup_slub_min_objects(char *str)
3783 get_option(&str, (int *)&slub_min_objects);
3788 __setup("slub_min_objects=", setup_slub_min_objects);
3790 void *__kmalloc(size_t size, gfp_t flags)
3792 struct kmem_cache *s;
3795 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3796 return kmalloc_large(size, flags);
3798 s = kmalloc_slab(size, flags);
3800 if (unlikely(ZERO_OR_NULL_PTR(s)))
3803 ret = slab_alloc(s, flags, _RET_IP_);
3805 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3807 ret = kasan_kmalloc(s, ret, size, flags);
3811 EXPORT_SYMBOL(__kmalloc);
3814 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3819 flags |= __GFP_COMP;
3820 page = alloc_pages_node(node, flags, get_order(size));
3822 ptr = page_address(page);
3824 return 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 ret = kasan_kmalloc(s, ret, size, flags);
3855 EXPORT_SYMBOL(__kmalloc_node);
3856 #endif /* CONFIG_NUMA */
3858 #ifdef CONFIG_HARDENED_USERCOPY
3860 * Rejects incorrectly sized objects and objects that are to be copied
3861 * to/from userspace but do not fall entirely within the containing slab
3862 * cache's usercopy region.
3864 * Returns NULL if check passes, otherwise const char * to name of cache
3865 * to indicate an error.
3867 void __check_heap_object(const void *ptr, unsigned long n, struct page *page,
3870 struct kmem_cache *s;
3871 unsigned int offset;
3874 ptr = kasan_reset_tag(ptr);
3876 /* Find object and usable object size. */
3877 s = page->slab_cache;
3879 /* Reject impossible pointers. */
3880 if (ptr < page_address(page))
3881 usercopy_abort("SLUB object not in SLUB page?!", NULL,
3884 /* Find offset within object. */
3885 offset = (ptr - page_address(page)) % s->size;
3887 /* Adjust for redzone and reject if within the redzone. */
3888 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE) {
3889 if (offset < s->red_left_pad)
3890 usercopy_abort("SLUB object in left red zone",
3891 s->name, to_user, offset, n);
3892 offset -= s->red_left_pad;
3895 /* Allow address range falling entirely within usercopy region. */
3896 if (offset >= s->useroffset &&
3897 offset - s->useroffset <= s->usersize &&
3898 n <= s->useroffset - offset + s->usersize)
3902 * If the copy is still within the allocated object, produce
3903 * a warning instead of rejecting the copy. This is intended
3904 * to be a temporary method to find any missing usercopy
3907 object_size = slab_ksize(s);
3908 if (usercopy_fallback &&
3909 offset <= object_size && n <= object_size - offset) {
3910 usercopy_warn("SLUB object", s->name, to_user, offset, n);
3914 usercopy_abort("SLUB object", s->name, to_user, offset, n);
3916 #endif /* CONFIG_HARDENED_USERCOPY */
3918 size_t __ksize(const void *object)
3922 if (unlikely(object == ZERO_SIZE_PTR))
3925 page = virt_to_head_page(object);
3927 if (unlikely(!PageSlab(page))) {
3928 WARN_ON(!PageCompound(page));
3929 return PAGE_SIZE << compound_order(page);
3932 return slab_ksize(page->slab_cache);
3934 EXPORT_SYMBOL(__ksize);
3936 void kfree(const void *x)
3939 void *object = (void *)x;
3941 trace_kfree(_RET_IP_, x);
3943 if (unlikely(ZERO_OR_NULL_PTR(x)))
3946 page = virt_to_head_page(x);
3947 if (unlikely(!PageSlab(page))) {
3948 BUG_ON(!PageCompound(page));
3950 __free_pages(page, compound_order(page));
3953 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
3955 EXPORT_SYMBOL(kfree);
3957 #define SHRINK_PROMOTE_MAX 32
3960 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3961 * up most to the head of the partial lists. New allocations will then
3962 * fill those up and thus they can be removed from the partial lists.
3964 * The slabs with the least items are placed last. This results in them
3965 * being allocated from last increasing the chance that the last objects
3966 * are freed in them.
3968 int __kmem_cache_shrink(struct kmem_cache *s)
3972 struct kmem_cache_node *n;
3975 struct list_head discard;
3976 struct list_head promote[SHRINK_PROMOTE_MAX];
3977 unsigned long flags;
3981 for_each_kmem_cache_node(s, node, n) {
3982 INIT_LIST_HEAD(&discard);
3983 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3984 INIT_LIST_HEAD(promote + i);
3986 spin_lock_irqsave(&n->list_lock, flags);
3989 * Build lists of slabs to discard or promote.
3991 * Note that concurrent frees may occur while we hold the
3992 * list_lock. page->inuse here is the upper limit.
3994 list_for_each_entry_safe(page, t, &n->partial, slab_list) {
3995 int free = page->objects - page->inuse;
3997 /* Do not reread page->inuse */
4000 /* We do not keep full slabs on the list */
4003 if (free == page->objects) {
4004 list_move(&page->slab_list, &discard);
4006 } else if (free <= SHRINK_PROMOTE_MAX)
4007 list_move(&page->slab_list, promote + free - 1);
4011 * Promote the slabs filled up most to the head of the
4014 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
4015 list_splice(promote + i, &n->partial);
4017 spin_unlock_irqrestore(&n->list_lock, flags);
4019 /* Release empty slabs */
4020 list_for_each_entry_safe(page, t, &discard, slab_list)
4021 discard_slab(s, page);
4023 if (slabs_node(s, node))
4031 void __kmemcg_cache_deactivate_after_rcu(struct kmem_cache *s)
4034 * Called with all the locks held after a sched RCU grace period.
4035 * Even if @s becomes empty after shrinking, we can't know that @s
4036 * doesn't have allocations already in-flight and thus can't
4037 * destroy @s until the associated memcg is released.
4039 * However, let's remove the sysfs files for empty caches here.
4040 * Each cache has a lot of interface files which aren't
4041 * particularly useful for empty draining caches; otherwise, we can
4042 * easily end up with millions of unnecessary sysfs files on
4043 * systems which have a lot of memory and transient cgroups.
4045 if (!__kmem_cache_shrink(s))
4046 sysfs_slab_remove(s);
4049 void __kmemcg_cache_deactivate(struct kmem_cache *s)
4052 * Disable empty slabs caching. Used to avoid pinning offline
4053 * memory cgroups by kmem pages that can be freed.
4055 slub_set_cpu_partial(s, 0);
4058 #endif /* CONFIG_MEMCG */
4060 static int slab_mem_going_offline_callback(void *arg)
4062 struct kmem_cache *s;
4064 mutex_lock(&slab_mutex);
4065 list_for_each_entry(s, &slab_caches, list)
4066 __kmem_cache_shrink(s);
4067 mutex_unlock(&slab_mutex);
4072 static void slab_mem_offline_callback(void *arg)
4074 struct kmem_cache_node *n;
4075 struct kmem_cache *s;
4076 struct memory_notify *marg = arg;
4079 offline_node = marg->status_change_nid_normal;
4082 * If the node still has available memory. we need kmem_cache_node
4085 if (offline_node < 0)
4088 mutex_lock(&slab_mutex);
4089 list_for_each_entry(s, &slab_caches, list) {
4090 n = get_node(s, offline_node);
4093 * if n->nr_slabs > 0, slabs still exist on the node
4094 * that is going down. We were unable to free them,
4095 * and offline_pages() function shouldn't call this
4096 * callback. So, we must fail.
4098 BUG_ON(slabs_node(s, offline_node));
4100 s->node[offline_node] = NULL;
4101 kmem_cache_free(kmem_cache_node, n);
4104 mutex_unlock(&slab_mutex);
4107 static int slab_mem_going_online_callback(void *arg)
4109 struct kmem_cache_node *n;
4110 struct kmem_cache *s;
4111 struct memory_notify *marg = arg;
4112 int nid = marg->status_change_nid_normal;
4116 * If the node's memory is already available, then kmem_cache_node is
4117 * already created. Nothing to do.
4123 * We are bringing a node online. No memory is available yet. We must
4124 * allocate a kmem_cache_node structure in order to bring the node
4127 mutex_lock(&slab_mutex);
4128 list_for_each_entry(s, &slab_caches, list) {
4130 * XXX: kmem_cache_alloc_node will fallback to other nodes
4131 * since memory is not yet available from the node that
4134 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4139 init_kmem_cache_node(n);
4143 mutex_unlock(&slab_mutex);
4147 static int slab_memory_callback(struct notifier_block *self,
4148 unsigned long action, void *arg)
4153 case MEM_GOING_ONLINE:
4154 ret = slab_mem_going_online_callback(arg);
4156 case MEM_GOING_OFFLINE:
4157 ret = slab_mem_going_offline_callback(arg);
4160 case MEM_CANCEL_ONLINE:
4161 slab_mem_offline_callback(arg);
4164 case MEM_CANCEL_OFFLINE:
4168 ret = notifier_from_errno(ret);
4174 static struct notifier_block slab_memory_callback_nb = {
4175 .notifier_call = slab_memory_callback,
4176 .priority = SLAB_CALLBACK_PRI,
4179 /********************************************************************
4180 * Basic setup of slabs
4181 *******************************************************************/
4184 * Used for early kmem_cache structures that were allocated using
4185 * the page allocator. Allocate them properly then fix up the pointers
4186 * that may be pointing to the wrong kmem_cache structure.
4189 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4192 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4193 struct kmem_cache_node *n;
4195 memcpy(s, static_cache, kmem_cache->object_size);
4198 * This runs very early, and only the boot processor is supposed to be
4199 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4202 __flush_cpu_slab(s, smp_processor_id());
4203 for_each_kmem_cache_node(s, node, n) {
4206 list_for_each_entry(p, &n->partial, slab_list)
4209 #ifdef CONFIG_SLUB_DEBUG
4210 list_for_each_entry(p, &n->full, slab_list)
4214 slab_init_memcg_params(s);
4215 list_add(&s->list, &slab_caches);
4216 memcg_link_cache(s, NULL);
4220 void __init kmem_cache_init(void)
4222 static __initdata struct kmem_cache boot_kmem_cache,
4223 boot_kmem_cache_node;
4225 if (debug_guardpage_minorder())
4228 kmem_cache_node = &boot_kmem_cache_node;
4229 kmem_cache = &boot_kmem_cache;
4231 create_boot_cache(kmem_cache_node, "kmem_cache_node",
4232 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
4234 register_hotmemory_notifier(&slab_memory_callback_nb);
4236 /* Able to allocate the per node structures */
4237 slab_state = PARTIAL;
4239 create_boot_cache(kmem_cache, "kmem_cache",
4240 offsetof(struct kmem_cache, node) +
4241 nr_node_ids * sizeof(struct kmem_cache_node *),
4242 SLAB_HWCACHE_ALIGN, 0, 0);
4244 kmem_cache = bootstrap(&boot_kmem_cache);
4245 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
4247 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4248 setup_kmalloc_cache_index_table();
4249 create_kmalloc_caches(0);
4251 /* Setup random freelists for each cache */
4252 init_freelist_randomization();
4254 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
4257 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
4259 slub_min_order, slub_max_order, slub_min_objects,
4260 nr_cpu_ids, nr_node_ids);
4263 void __init kmem_cache_init_late(void)
4268 __kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
4269 slab_flags_t flags, void (*ctor)(void *))
4271 struct kmem_cache *s, *c;
4273 s = find_mergeable(size, align, flags, name, ctor);
4278 * Adjust the object sizes so that we clear
4279 * the complete object on kzalloc.
4281 s->object_size = max(s->object_size, size);
4282 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
4284 for_each_memcg_cache(c, s) {
4285 c->object_size = s->object_size;
4286 c->inuse = max(c->inuse, ALIGN(size, sizeof(void *)));
4289 if (sysfs_slab_alias(s, name)) {
4298 int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
4302 err = kmem_cache_open(s, flags);
4306 /* Mutex is not taken during early boot */
4307 if (slab_state <= UP)
4310 memcg_propagate_slab_attrs(s);
4311 err = sysfs_slab_add(s);
4313 __kmem_cache_release(s);
4318 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4320 struct kmem_cache *s;
4323 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4324 return kmalloc_large(size, gfpflags);
4326 s = kmalloc_slab(size, gfpflags);
4328 if (unlikely(ZERO_OR_NULL_PTR(s)))
4331 ret = slab_alloc(s, gfpflags, caller);
4333 /* Honor the call site pointer we received. */
4334 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4340 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4341 int node, unsigned long caller)
4343 struct kmem_cache *s;
4346 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4347 ret = kmalloc_large_node(size, gfpflags, node);
4349 trace_kmalloc_node(caller, ret,
4350 size, PAGE_SIZE << get_order(size),
4356 s = kmalloc_slab(size, gfpflags);
4358 if (unlikely(ZERO_OR_NULL_PTR(s)))
4361 ret = slab_alloc_node(s, gfpflags, node, caller);
4363 /* Honor the call site pointer we received. */
4364 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4371 static int count_inuse(struct page *page)
4376 static int count_total(struct page *page)
4378 return page->objects;
4382 #ifdef CONFIG_SLUB_DEBUG
4383 static int validate_slab(struct kmem_cache *s, struct page *page,
4387 void *addr = page_address(page);
4389 if (!check_slab(s, page) ||
4390 !on_freelist(s, page, NULL))
4393 /* Now we know that a valid freelist exists */
4394 bitmap_zero(map, page->objects);
4396 get_map(s, page, map);
4397 for_each_object(p, s, addr, page->objects) {
4398 if (test_bit(slab_index(p, s, addr), map))
4399 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4403 for_each_object(p, s, addr, page->objects)
4404 if (!test_bit(slab_index(p, s, addr), map))
4405 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4410 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4414 validate_slab(s, page, map);
4418 static int validate_slab_node(struct kmem_cache *s,
4419 struct kmem_cache_node *n, unsigned long *map)
4421 unsigned long count = 0;
4423 unsigned long flags;
4425 spin_lock_irqsave(&n->list_lock, flags);
4427 list_for_each_entry(page, &n->partial, slab_list) {
4428 validate_slab_slab(s, page, map);
4431 if (count != n->nr_partial)
4432 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4433 s->name, count, n->nr_partial);
4435 if (!(s->flags & SLAB_STORE_USER))
4438 list_for_each_entry(page, &n->full, slab_list) {
4439 validate_slab_slab(s, page, map);
4442 if (count != atomic_long_read(&n->nr_slabs))
4443 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4444 s->name, count, atomic_long_read(&n->nr_slabs));
4447 spin_unlock_irqrestore(&n->list_lock, flags);
4451 static long validate_slab_cache(struct kmem_cache *s)
4454 unsigned long count = 0;
4455 struct kmem_cache_node *n;
4456 unsigned long *map = bitmap_alloc(oo_objects(s->max), GFP_KERNEL);
4462 for_each_kmem_cache_node(s, node, n)
4463 count += validate_slab_node(s, n, map);
4468 * Generate lists of code addresses where slabcache objects are allocated
4473 unsigned long count;
4480 DECLARE_BITMAP(cpus, NR_CPUS);
4486 unsigned long count;
4487 struct location *loc;
4490 static void free_loc_track(struct loc_track *t)
4493 free_pages((unsigned long)t->loc,
4494 get_order(sizeof(struct location) * t->max));
4497 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4502 order = get_order(sizeof(struct location) * max);
4504 l = (void *)__get_free_pages(flags, order);
4509 memcpy(l, t->loc, sizeof(struct location) * t->count);
4517 static int add_location(struct loc_track *t, struct kmem_cache *s,
4518 const struct track *track)
4520 long start, end, pos;
4522 unsigned long caddr;
4523 unsigned long age = jiffies - track->when;
4529 pos = start + (end - start + 1) / 2;
4532 * There is nothing at "end". If we end up there
4533 * we need to add something to before end.
4538 caddr = t->loc[pos].addr;
4539 if (track->addr == caddr) {
4545 if (age < l->min_time)
4547 if (age > l->max_time)
4550 if (track->pid < l->min_pid)
4551 l->min_pid = track->pid;
4552 if (track->pid > l->max_pid)
4553 l->max_pid = track->pid;
4555 cpumask_set_cpu(track->cpu,
4556 to_cpumask(l->cpus));
4558 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4562 if (track->addr < caddr)
4569 * Not found. Insert new tracking element.
4571 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4577 (t->count - pos) * sizeof(struct location));
4580 l->addr = track->addr;
4584 l->min_pid = track->pid;
4585 l->max_pid = track->pid;
4586 cpumask_clear(to_cpumask(l->cpus));
4587 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4588 nodes_clear(l->nodes);
4589 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4593 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4594 struct page *page, enum track_item alloc,
4597 void *addr = page_address(page);
4600 bitmap_zero(map, page->objects);
4601 get_map(s, page, map);
4603 for_each_object(p, s, addr, page->objects)
4604 if (!test_bit(slab_index(p, s, addr), map))
4605 add_location(t, s, get_track(s, p, alloc));
4608 static int list_locations(struct kmem_cache *s, char *buf,
4609 enum track_item alloc)
4613 struct loc_track t = { 0, 0, NULL };
4615 struct kmem_cache_node *n;
4616 unsigned long *map = bitmap_alloc(oo_objects(s->max), GFP_KERNEL);
4618 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4621 return sprintf(buf, "Out of memory\n");
4623 /* Push back cpu slabs */
4626 for_each_kmem_cache_node(s, node, n) {
4627 unsigned long flags;
4630 if (!atomic_long_read(&n->nr_slabs))
4633 spin_lock_irqsave(&n->list_lock, flags);
4634 list_for_each_entry(page, &n->partial, slab_list)
4635 process_slab(&t, s, page, alloc, map);
4636 list_for_each_entry(page, &n->full, slab_list)
4637 process_slab(&t, s, page, alloc, map);
4638 spin_unlock_irqrestore(&n->list_lock, flags);
4641 for (i = 0; i < t.count; i++) {
4642 struct location *l = &t.loc[i];
4644 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4646 len += sprintf(buf + len, "%7ld ", l->count);
4649 len += sprintf(buf + len, "%pS", (void *)l->addr);
4651 len += sprintf(buf + len, "<not-available>");
4653 if (l->sum_time != l->min_time) {
4654 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4656 (long)div_u64(l->sum_time, l->count),
4659 len += sprintf(buf + len, " age=%ld",
4662 if (l->min_pid != l->max_pid)
4663 len += sprintf(buf + len, " pid=%ld-%ld",
4664 l->min_pid, l->max_pid);
4666 len += sprintf(buf + len, " pid=%ld",
4669 if (num_online_cpus() > 1 &&
4670 !cpumask_empty(to_cpumask(l->cpus)) &&
4671 len < PAGE_SIZE - 60)
4672 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4674 cpumask_pr_args(to_cpumask(l->cpus)));
4676 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4677 len < PAGE_SIZE - 60)
4678 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4680 nodemask_pr_args(&l->nodes));
4682 len += sprintf(buf + len, "\n");
4688 len += sprintf(buf, "No data\n");
4691 #endif /* CONFIG_SLUB_DEBUG */
4693 #ifdef SLUB_RESILIENCY_TEST
4694 static void __init resiliency_test(void)
4697 int type = KMALLOC_NORMAL;
4699 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4701 pr_err("SLUB resiliency testing\n");
4702 pr_err("-----------------------\n");
4703 pr_err("A. Corruption after allocation\n");
4705 p = kzalloc(16, GFP_KERNEL);
4707 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4710 validate_slab_cache(kmalloc_caches[type][4]);
4712 /* Hmmm... The next two are dangerous */
4713 p = kzalloc(32, GFP_KERNEL);
4714 p[32 + sizeof(void *)] = 0x34;
4715 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4717 pr_err("If allocated object is overwritten then not detectable\n\n");
4719 validate_slab_cache(kmalloc_caches[type][5]);
4720 p = kzalloc(64, GFP_KERNEL);
4721 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4723 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4725 pr_err("If allocated object is overwritten then not detectable\n\n");
4726 validate_slab_cache(kmalloc_caches[type][6]);
4728 pr_err("\nB. Corruption after free\n");
4729 p = kzalloc(128, GFP_KERNEL);
4732 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4733 validate_slab_cache(kmalloc_caches[type][7]);
4735 p = kzalloc(256, GFP_KERNEL);
4738 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4739 validate_slab_cache(kmalloc_caches[type][8]);
4741 p = kzalloc(512, GFP_KERNEL);
4744 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4745 validate_slab_cache(kmalloc_caches[type][9]);
4749 static void resiliency_test(void) {};
4751 #endif /* SLUB_RESILIENCY_TEST */
4754 enum slab_stat_type {
4755 SL_ALL, /* All slabs */
4756 SL_PARTIAL, /* Only partially allocated slabs */
4757 SL_CPU, /* Only slabs used for cpu caches */
4758 SL_OBJECTS, /* Determine allocated objects not slabs */
4759 SL_TOTAL /* Determine object capacity not slabs */
4762 #define SO_ALL (1 << SL_ALL)
4763 #define SO_PARTIAL (1 << SL_PARTIAL)
4764 #define SO_CPU (1 << SL_CPU)
4765 #define SO_OBJECTS (1 << SL_OBJECTS)
4766 #define SO_TOTAL (1 << SL_TOTAL)
4769 static bool memcg_sysfs_enabled = IS_ENABLED(CONFIG_SLUB_MEMCG_SYSFS_ON);
4771 static int __init setup_slub_memcg_sysfs(char *str)
4775 if (get_option(&str, &v) > 0)
4776 memcg_sysfs_enabled = v;
4781 __setup("slub_memcg_sysfs=", setup_slub_memcg_sysfs);
4784 static ssize_t show_slab_objects(struct kmem_cache *s,
4785 char *buf, unsigned long flags)
4787 unsigned long total = 0;
4790 unsigned long *nodes;
4792 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
4796 if (flags & SO_CPU) {
4799 for_each_possible_cpu(cpu) {
4800 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4805 page = READ_ONCE(c->page);
4809 node = page_to_nid(page);
4810 if (flags & SO_TOTAL)
4812 else if (flags & SO_OBJECTS)
4820 page = slub_percpu_partial_read_once(c);
4822 node = page_to_nid(page);
4823 if (flags & SO_TOTAL)
4825 else if (flags & SO_OBJECTS)
4836 #ifdef CONFIG_SLUB_DEBUG
4837 if (flags & SO_ALL) {
4838 struct kmem_cache_node *n;
4840 for_each_kmem_cache_node(s, node, n) {
4842 if (flags & SO_TOTAL)
4843 x = atomic_long_read(&n->total_objects);
4844 else if (flags & SO_OBJECTS)
4845 x = atomic_long_read(&n->total_objects) -
4846 count_partial(n, count_free);
4848 x = atomic_long_read(&n->nr_slabs);
4855 if (flags & SO_PARTIAL) {
4856 struct kmem_cache_node *n;
4858 for_each_kmem_cache_node(s, node, n) {
4859 if (flags & SO_TOTAL)
4860 x = count_partial(n, count_total);
4861 else if (flags & SO_OBJECTS)
4862 x = count_partial(n, count_inuse);
4869 x = sprintf(buf, "%lu", total);
4871 for (node = 0; node < nr_node_ids; node++)
4873 x += sprintf(buf + x, " N%d=%lu",
4878 return x + sprintf(buf + x, "\n");
4881 #ifdef CONFIG_SLUB_DEBUG
4882 static int any_slab_objects(struct kmem_cache *s)
4885 struct kmem_cache_node *n;
4887 for_each_kmem_cache_node(s, node, n)
4888 if (atomic_long_read(&n->total_objects))
4895 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4896 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4898 struct slab_attribute {
4899 struct attribute attr;
4900 ssize_t (*show)(struct kmem_cache *s, char *buf);
4901 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4904 #define SLAB_ATTR_RO(_name) \
4905 static struct slab_attribute _name##_attr = \
4906 __ATTR(_name, 0400, _name##_show, NULL)
4908 #define SLAB_ATTR(_name) \
4909 static struct slab_attribute _name##_attr = \
4910 __ATTR(_name, 0600, _name##_show, _name##_store)
4912 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4914 return sprintf(buf, "%u\n", s->size);
4916 SLAB_ATTR_RO(slab_size);
4918 static ssize_t align_show(struct kmem_cache *s, char *buf)
4920 return sprintf(buf, "%u\n", s->align);
4922 SLAB_ATTR_RO(align);
4924 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4926 return sprintf(buf, "%u\n", s->object_size);
4928 SLAB_ATTR_RO(object_size);
4930 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4932 return sprintf(buf, "%u\n", oo_objects(s->oo));
4934 SLAB_ATTR_RO(objs_per_slab);
4936 static ssize_t order_store(struct kmem_cache *s,
4937 const char *buf, size_t length)
4942 err = kstrtouint(buf, 10, &order);
4946 if (order > slub_max_order || order < slub_min_order)
4949 calculate_sizes(s, order);
4953 static ssize_t order_show(struct kmem_cache *s, char *buf)
4955 return sprintf(buf, "%u\n", oo_order(s->oo));
4959 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4961 return sprintf(buf, "%lu\n", s->min_partial);
4964 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4970 err = kstrtoul(buf, 10, &min);
4974 set_min_partial(s, min);
4977 SLAB_ATTR(min_partial);
4979 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4981 return sprintf(buf, "%u\n", slub_cpu_partial(s));
4984 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4987 unsigned int objects;
4990 err = kstrtouint(buf, 10, &objects);
4993 if (objects && !kmem_cache_has_cpu_partial(s))
4996 slub_set_cpu_partial(s, objects);
5000 SLAB_ATTR(cpu_partial);
5002 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
5006 return sprintf(buf, "%pS\n", s->ctor);
5010 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
5012 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
5014 SLAB_ATTR_RO(aliases);
5016 static ssize_t partial_show(struct kmem_cache *s, char *buf)
5018 return show_slab_objects(s, buf, SO_PARTIAL);
5020 SLAB_ATTR_RO(partial);
5022 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
5024 return show_slab_objects(s, buf, SO_CPU);
5026 SLAB_ATTR_RO(cpu_slabs);
5028 static ssize_t objects_show(struct kmem_cache *s, char *buf)
5030 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5032 SLAB_ATTR_RO(objects);
5034 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5036 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5038 SLAB_ATTR_RO(objects_partial);
5040 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5047 for_each_online_cpu(cpu) {
5050 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5053 pages += page->pages;
5054 objects += page->pobjects;
5058 len = sprintf(buf, "%d(%d)", objects, pages);
5061 for_each_online_cpu(cpu) {
5064 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5066 if (page && len < PAGE_SIZE - 20)
5067 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
5068 page->pobjects, page->pages);
5071 return len + sprintf(buf + len, "\n");
5073 SLAB_ATTR_RO(slabs_cpu_partial);
5075 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5077 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5080 static ssize_t reclaim_account_store(struct kmem_cache *s,
5081 const char *buf, size_t length)
5083 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
5085 s->flags |= SLAB_RECLAIM_ACCOUNT;
5088 SLAB_ATTR(reclaim_account);
5090 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5092 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5094 SLAB_ATTR_RO(hwcache_align);
5096 #ifdef CONFIG_ZONE_DMA
5097 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5099 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5101 SLAB_ATTR_RO(cache_dma);
5104 static ssize_t usersize_show(struct kmem_cache *s, char *buf)
5106 return sprintf(buf, "%u\n", s->usersize);
5108 SLAB_ATTR_RO(usersize);
5110 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5112 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
5114 SLAB_ATTR_RO(destroy_by_rcu);
5116 #ifdef CONFIG_SLUB_DEBUG
5117 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5119 return show_slab_objects(s, buf, SO_ALL);
5121 SLAB_ATTR_RO(slabs);
5123 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5125 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5127 SLAB_ATTR_RO(total_objects);
5129 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5131 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5134 static ssize_t sanity_checks_store(struct kmem_cache *s,
5135 const char *buf, size_t length)
5137 s->flags &= ~SLAB_CONSISTENCY_CHECKS;
5138 if (buf[0] == '1') {
5139 s->flags &= ~__CMPXCHG_DOUBLE;
5140 s->flags |= SLAB_CONSISTENCY_CHECKS;
5144 SLAB_ATTR(sanity_checks);
5146 static ssize_t trace_show(struct kmem_cache *s, char *buf)
5148 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5151 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
5155 * Tracing a merged cache is going to give confusing results
5156 * as well as cause other issues like converting a mergeable
5157 * cache into an umergeable one.
5159 if (s->refcount > 1)
5162 s->flags &= ~SLAB_TRACE;
5163 if (buf[0] == '1') {
5164 s->flags &= ~__CMPXCHG_DOUBLE;
5165 s->flags |= SLAB_TRACE;
5171 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5173 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5176 static ssize_t red_zone_store(struct kmem_cache *s,
5177 const char *buf, size_t length)
5179 if (any_slab_objects(s))
5182 s->flags &= ~SLAB_RED_ZONE;
5183 if (buf[0] == '1') {
5184 s->flags |= SLAB_RED_ZONE;
5186 calculate_sizes(s, -1);
5189 SLAB_ATTR(red_zone);
5191 static ssize_t poison_show(struct kmem_cache *s, char *buf)
5193 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
5196 static ssize_t poison_store(struct kmem_cache *s,
5197 const char *buf, size_t length)
5199 if (any_slab_objects(s))
5202 s->flags &= ~SLAB_POISON;
5203 if (buf[0] == '1') {
5204 s->flags |= SLAB_POISON;
5206 calculate_sizes(s, -1);
5211 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5213 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5216 static ssize_t store_user_store(struct kmem_cache *s,
5217 const char *buf, size_t length)
5219 if (any_slab_objects(s))
5222 s->flags &= ~SLAB_STORE_USER;
5223 if (buf[0] == '1') {
5224 s->flags &= ~__CMPXCHG_DOUBLE;
5225 s->flags |= SLAB_STORE_USER;
5227 calculate_sizes(s, -1);
5230 SLAB_ATTR(store_user);
5232 static ssize_t validate_show(struct kmem_cache *s, char *buf)
5237 static ssize_t validate_store(struct kmem_cache *s,
5238 const char *buf, size_t length)
5242 if (buf[0] == '1') {
5243 ret = validate_slab_cache(s);
5249 SLAB_ATTR(validate);
5251 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
5253 if (!(s->flags & SLAB_STORE_USER))
5255 return list_locations(s, buf, TRACK_ALLOC);
5257 SLAB_ATTR_RO(alloc_calls);
5259 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
5261 if (!(s->flags & SLAB_STORE_USER))
5263 return list_locations(s, buf, TRACK_FREE);
5265 SLAB_ATTR_RO(free_calls);
5266 #endif /* CONFIG_SLUB_DEBUG */
5268 #ifdef CONFIG_FAILSLAB
5269 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5271 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5274 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
5277 if (s->refcount > 1)
5280 s->flags &= ~SLAB_FAILSLAB;
5282 s->flags |= SLAB_FAILSLAB;
5285 SLAB_ATTR(failslab);
5288 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5293 static ssize_t shrink_store(struct kmem_cache *s,
5294 const char *buf, size_t length)
5297 kmem_cache_shrink(s);
5305 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5307 return sprintf(buf, "%u\n", s->remote_node_defrag_ratio / 10);
5310 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5311 const char *buf, size_t length)
5316 err = kstrtouint(buf, 10, &ratio);
5322 s->remote_node_defrag_ratio = ratio * 10;
5326 SLAB_ATTR(remote_node_defrag_ratio);
5329 #ifdef CONFIG_SLUB_STATS
5330 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5332 unsigned long sum = 0;
5335 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
5340 for_each_online_cpu(cpu) {
5341 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5347 len = sprintf(buf, "%lu", sum);
5350 for_each_online_cpu(cpu) {
5351 if (data[cpu] && len < PAGE_SIZE - 20)
5352 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5356 return len + sprintf(buf + len, "\n");
5359 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5363 for_each_online_cpu(cpu)
5364 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5367 #define STAT_ATTR(si, text) \
5368 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5370 return show_stat(s, buf, si); \
5372 static ssize_t text##_store(struct kmem_cache *s, \
5373 const char *buf, size_t length) \
5375 if (buf[0] != '0') \
5377 clear_stat(s, si); \
5382 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5383 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5384 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5385 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5386 STAT_ATTR(FREE_FROZEN, free_frozen);
5387 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5388 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5389 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5390 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5391 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5392 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5393 STAT_ATTR(FREE_SLAB, free_slab);
5394 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5395 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5396 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5397 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5398 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5399 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5400 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5401 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5402 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5403 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5404 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5405 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5406 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5407 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5408 #endif /* CONFIG_SLUB_STATS */
5410 static struct attribute *slab_attrs[] = {
5411 &slab_size_attr.attr,
5412 &object_size_attr.attr,
5413 &objs_per_slab_attr.attr,
5415 &min_partial_attr.attr,
5416 &cpu_partial_attr.attr,
5418 &objects_partial_attr.attr,
5420 &cpu_slabs_attr.attr,
5424 &hwcache_align_attr.attr,
5425 &reclaim_account_attr.attr,
5426 &destroy_by_rcu_attr.attr,
5428 &slabs_cpu_partial_attr.attr,
5429 #ifdef CONFIG_SLUB_DEBUG
5430 &total_objects_attr.attr,
5432 &sanity_checks_attr.attr,
5434 &red_zone_attr.attr,
5436 &store_user_attr.attr,
5437 &validate_attr.attr,
5438 &alloc_calls_attr.attr,
5439 &free_calls_attr.attr,
5441 #ifdef CONFIG_ZONE_DMA
5442 &cache_dma_attr.attr,
5445 &remote_node_defrag_ratio_attr.attr,
5447 #ifdef CONFIG_SLUB_STATS
5448 &alloc_fastpath_attr.attr,
5449 &alloc_slowpath_attr.attr,
5450 &free_fastpath_attr.attr,
5451 &free_slowpath_attr.attr,
5452 &free_frozen_attr.attr,
5453 &free_add_partial_attr.attr,
5454 &free_remove_partial_attr.attr,
5455 &alloc_from_partial_attr.attr,
5456 &alloc_slab_attr.attr,
5457 &alloc_refill_attr.attr,
5458 &alloc_node_mismatch_attr.attr,
5459 &free_slab_attr.attr,
5460 &cpuslab_flush_attr.attr,
5461 &deactivate_full_attr.attr,
5462 &deactivate_empty_attr.attr,
5463 &deactivate_to_head_attr.attr,
5464 &deactivate_to_tail_attr.attr,
5465 &deactivate_remote_frees_attr.attr,
5466 &deactivate_bypass_attr.attr,
5467 &order_fallback_attr.attr,
5468 &cmpxchg_double_fail_attr.attr,
5469 &cmpxchg_double_cpu_fail_attr.attr,
5470 &cpu_partial_alloc_attr.attr,
5471 &cpu_partial_free_attr.attr,
5472 &cpu_partial_node_attr.attr,
5473 &cpu_partial_drain_attr.attr,
5475 #ifdef CONFIG_FAILSLAB
5476 &failslab_attr.attr,
5478 &usersize_attr.attr,
5483 static const struct attribute_group slab_attr_group = {
5484 .attrs = slab_attrs,
5487 static ssize_t slab_attr_show(struct kobject *kobj,
5488 struct attribute *attr,
5491 struct slab_attribute *attribute;
5492 struct kmem_cache *s;
5495 attribute = to_slab_attr(attr);
5498 if (!attribute->show)
5501 err = attribute->show(s, buf);
5506 static ssize_t slab_attr_store(struct kobject *kobj,
5507 struct attribute *attr,
5508 const char *buf, size_t len)
5510 struct slab_attribute *attribute;
5511 struct kmem_cache *s;
5514 attribute = to_slab_attr(attr);
5517 if (!attribute->store)
5520 err = attribute->store(s, buf, len);
5522 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5523 struct kmem_cache *c;
5525 mutex_lock(&slab_mutex);
5526 if (s->max_attr_size < len)
5527 s->max_attr_size = len;
5530 * This is a best effort propagation, so this function's return
5531 * value will be determined by the parent cache only. This is
5532 * basically because not all attributes will have a well
5533 * defined semantics for rollbacks - most of the actions will
5534 * have permanent effects.
5536 * Returning the error value of any of the children that fail
5537 * is not 100 % defined, in the sense that users seeing the
5538 * error code won't be able to know anything about the state of
5541 * Only returning the error code for the parent cache at least
5542 * has well defined semantics. The cache being written to
5543 * directly either failed or succeeded, in which case we loop
5544 * through the descendants with best-effort propagation.
5546 for_each_memcg_cache(c, s)
5547 attribute->store(c, buf, len);
5548 mutex_unlock(&slab_mutex);
5554 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5558 char *buffer = NULL;
5559 struct kmem_cache *root_cache;
5561 if (is_root_cache(s))
5564 root_cache = s->memcg_params.root_cache;
5567 * This mean this cache had no attribute written. Therefore, no point
5568 * in copying default values around
5570 if (!root_cache->max_attr_size)
5573 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5576 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5579 if (!attr || !attr->store || !attr->show)
5583 * It is really bad that we have to allocate here, so we will
5584 * do it only as a fallback. If we actually allocate, though,
5585 * we can just use the allocated buffer until the end.
5587 * Most of the slub attributes will tend to be very small in
5588 * size, but sysfs allows buffers up to a page, so they can
5589 * theoretically happen.
5593 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5596 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5597 if (WARN_ON(!buffer))
5602 len = attr->show(root_cache, buf);
5604 attr->store(s, buf, len);
5608 free_page((unsigned long)buffer);
5609 #endif /* CONFIG_MEMCG */
5612 static void kmem_cache_release(struct kobject *k)
5614 slab_kmem_cache_release(to_slab(k));
5617 static const struct sysfs_ops slab_sysfs_ops = {
5618 .show = slab_attr_show,
5619 .store = slab_attr_store,
5622 static struct kobj_type slab_ktype = {
5623 .sysfs_ops = &slab_sysfs_ops,
5624 .release = kmem_cache_release,
5627 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5629 struct kobj_type *ktype = get_ktype(kobj);
5631 if (ktype == &slab_ktype)
5636 static const struct kset_uevent_ops slab_uevent_ops = {
5637 .filter = uevent_filter,
5640 static struct kset *slab_kset;
5642 static inline struct kset *cache_kset(struct kmem_cache *s)
5645 if (!is_root_cache(s))
5646 return s->memcg_params.root_cache->memcg_kset;
5651 #define ID_STR_LENGTH 64
5653 /* Create a unique string id for a slab cache:
5655 * Format :[flags-]size
5657 static char *create_unique_id(struct kmem_cache *s)
5659 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5666 * First flags affecting slabcache operations. We will only
5667 * get here for aliasable slabs so we do not need to support
5668 * too many flags. The flags here must cover all flags that
5669 * are matched during merging to guarantee that the id is
5672 if (s->flags & SLAB_CACHE_DMA)
5674 if (s->flags & SLAB_CACHE_DMA32)
5676 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5678 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5680 if (s->flags & SLAB_ACCOUNT)
5684 p += sprintf(p, "%07u", s->size);
5686 BUG_ON(p > name + ID_STR_LENGTH - 1);
5690 static void sysfs_slab_remove_workfn(struct work_struct *work)
5692 struct kmem_cache *s =
5693 container_of(work, struct kmem_cache, kobj_remove_work);
5695 if (!s->kobj.state_in_sysfs)
5697 * For a memcg cache, this may be called during
5698 * deactivation and again on shutdown. Remove only once.
5699 * A cache is never shut down before deactivation is
5700 * complete, so no need to worry about synchronization.
5705 kset_unregister(s->memcg_kset);
5707 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5709 kobject_put(&s->kobj);
5712 static int sysfs_slab_add(struct kmem_cache *s)
5716 struct kset *kset = cache_kset(s);
5717 int unmergeable = slab_unmergeable(s);
5719 INIT_WORK(&s->kobj_remove_work, sysfs_slab_remove_workfn);
5722 kobject_init(&s->kobj, &slab_ktype);
5726 if (!unmergeable && disable_higher_order_debug &&
5727 (slub_debug & DEBUG_METADATA_FLAGS))
5732 * Slabcache can never be merged so we can use the name proper.
5733 * This is typically the case for debug situations. In that
5734 * case we can catch duplicate names easily.
5736 sysfs_remove_link(&slab_kset->kobj, s->name);
5740 * Create a unique name for the slab as a target
5743 name = create_unique_id(s);
5746 s->kobj.kset = kset;
5747 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5751 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5756 if (is_root_cache(s) && memcg_sysfs_enabled) {
5757 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5758 if (!s->memcg_kset) {
5765 kobject_uevent(&s->kobj, KOBJ_ADD);
5767 /* Setup first alias */
5768 sysfs_slab_alias(s, s->name);
5775 kobject_del(&s->kobj);
5779 static void sysfs_slab_remove(struct kmem_cache *s)
5781 if (slab_state < FULL)
5783 * Sysfs has not been setup yet so no need to remove the
5788 kobject_get(&s->kobj);
5789 schedule_work(&s->kobj_remove_work);
5792 void sysfs_slab_unlink(struct kmem_cache *s)
5794 if (slab_state >= FULL)
5795 kobject_del(&s->kobj);
5798 void sysfs_slab_release(struct kmem_cache *s)
5800 if (slab_state >= FULL)
5801 kobject_put(&s->kobj);
5805 * Need to buffer aliases during bootup until sysfs becomes
5806 * available lest we lose that information.
5808 struct saved_alias {
5809 struct kmem_cache *s;
5811 struct saved_alias *next;
5814 static struct saved_alias *alias_list;
5816 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5818 struct saved_alias *al;
5820 if (slab_state == FULL) {
5822 * If we have a leftover link then remove it.
5824 sysfs_remove_link(&slab_kset->kobj, name);
5825 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5828 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5834 al->next = alias_list;
5839 static int __init slab_sysfs_init(void)
5841 struct kmem_cache *s;
5844 mutex_lock(&slab_mutex);
5846 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5848 mutex_unlock(&slab_mutex);
5849 pr_err("Cannot register slab subsystem.\n");
5855 list_for_each_entry(s, &slab_caches, list) {
5856 err = sysfs_slab_add(s);
5858 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5862 while (alias_list) {
5863 struct saved_alias *al = alias_list;
5865 alias_list = alias_list->next;
5866 err = sysfs_slab_alias(al->s, al->name);
5868 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5873 mutex_unlock(&slab_mutex);
5878 __initcall(slab_sysfs_init);
5879 #endif /* CONFIG_SYSFS */
5882 * The /proc/slabinfo ABI
5884 #ifdef CONFIG_SLUB_DEBUG
5885 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5887 unsigned long nr_slabs = 0;
5888 unsigned long nr_objs = 0;
5889 unsigned long nr_free = 0;
5891 struct kmem_cache_node *n;
5893 for_each_kmem_cache_node(s, node, n) {
5894 nr_slabs += node_nr_slabs(n);
5895 nr_objs += node_nr_objs(n);
5896 nr_free += count_partial(n, count_free);
5899 sinfo->active_objs = nr_objs - nr_free;
5900 sinfo->num_objs = nr_objs;
5901 sinfo->active_slabs = nr_slabs;
5902 sinfo->num_slabs = nr_slabs;
5903 sinfo->objects_per_slab = oo_objects(s->oo);
5904 sinfo->cache_order = oo_order(s->oo);
5907 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5911 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5912 size_t count, loff_t *ppos)
5916 #endif /* CONFIG_SLUB_DEBUG */