1 // SPDX-License-Identifier: GPL-2.0
3 * SLUB: A slab allocator that limits cache line use instead of queuing
4 * objects in per cpu and per node lists.
6 * The allocator synchronizes using per slab locks or atomic operatios
7 * and only uses a centralized lock to manage a pool of partial slabs.
9 * (C) 2007 SGI, Christoph Lameter
10 * (C) 2011 Linux Foundation, Christoph Lameter
14 #include <linux/swap.h> /* struct reclaim_state */
15 #include <linux/module.h>
16 #include <linux/bit_spinlock.h>
17 #include <linux/interrupt.h>
18 #include <linux/swab.h>
19 #include <linux/bitops.h>
20 #include <linux/slab.h>
22 #include <linux/proc_fs.h>
23 #include <linux/seq_file.h>
24 #include <linux/kasan.h>
25 #include <linux/cpu.h>
26 #include <linux/cpuset.h>
27 #include <linux/mempolicy.h>
28 #include <linux/ctype.h>
29 #include <linux/debugobjects.h>
30 #include <linux/kallsyms.h>
31 #include <linux/memory.h>
32 #include <linux/math64.h>
33 #include <linux/fault-inject.h>
34 #include <linux/stacktrace.h>
35 #include <linux/prefetch.h>
36 #include <linux/memcontrol.h>
37 #include <linux/random.h>
39 #include <trace/events/kmem.h>
45 * 1. slab_mutex (Global Mutex)
47 * 3. slab_lock(page) (Only on some arches and for debugging)
51 * The role of the slab_mutex is to protect the list of all the slabs
52 * and to synchronize major metadata changes to slab cache structures.
54 * The slab_lock is only used for debugging and on arches that do not
55 * have the ability to do a cmpxchg_double. It only protects:
56 * A. page->freelist -> List of object free in a page
57 * B. page->inuse -> Number of objects in use
58 * C. page->objects -> Number of objects in page
59 * D. page->frozen -> frozen state
61 * If a slab is frozen then it is exempt from list management. It is not
62 * on any list. The processor that froze the slab is the one who can
63 * perform list operations on the page. Other processors may put objects
64 * onto the freelist but the processor that froze the slab is the only
65 * one that can retrieve the objects from the page's freelist.
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
253 return (void *)((unsigned long)ptr ^ s->random ^ swab(ptr_addr));
259 /* Returns the freelist pointer recorded at location ptr_addr. */
260 static inline void *freelist_dereference(const struct kmem_cache *s,
263 return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr),
264 (unsigned long)ptr_addr);
267 static inline void *get_freepointer(struct kmem_cache *s, void *object)
269 return freelist_dereference(s, object + s->offset);
272 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
274 prefetch(object + s->offset);
277 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
279 unsigned long freepointer_addr;
282 if (!debug_pagealloc_enabled())
283 return get_freepointer(s, object);
285 freepointer_addr = (unsigned long)object + s->offset;
286 probe_kernel_read(&p, (void **)freepointer_addr, sizeof(p));
287 return freelist_ptr(s, p, freepointer_addr);
290 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
292 unsigned long freeptr_addr = (unsigned long)object + s->offset;
294 #ifdef CONFIG_SLAB_FREELIST_HARDENED
295 BUG_ON(object == fp); /* naive detection of double free or corruption */
298 *(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr);
301 /* Loop over all objects in a slab */
302 #define for_each_object(__p, __s, __addr, __objects) \
303 for (__p = fixup_red_left(__s, __addr); \
304 __p < (__addr) + (__objects) * (__s)->size; \
307 #define for_each_object_idx(__p, __idx, __s, __addr, __objects) \
308 for (__p = fixup_red_left(__s, __addr), __idx = 1; \
309 __idx <= __objects; \
310 __p += (__s)->size, __idx++)
312 /* Determine object index from a given position */
313 static inline unsigned int slab_index(void *p, struct kmem_cache *s, void *addr)
315 return (p - addr) / s->size;
318 static inline unsigned int order_objects(unsigned int order, unsigned int size)
320 return ((unsigned int)PAGE_SIZE << order) / size;
323 static inline struct kmem_cache_order_objects oo_make(unsigned int order,
326 struct kmem_cache_order_objects x = {
327 (order << OO_SHIFT) + order_objects(order, size)
333 static inline unsigned int oo_order(struct kmem_cache_order_objects x)
335 return x.x >> OO_SHIFT;
338 static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
340 return x.x & OO_MASK;
344 * Per slab locking using the pagelock
346 static __always_inline void slab_lock(struct page *page)
348 VM_BUG_ON_PAGE(PageTail(page), page);
349 bit_spin_lock(PG_locked, &page->flags);
352 static __always_inline void slab_unlock(struct page *page)
354 VM_BUG_ON_PAGE(PageTail(page), page);
355 __bit_spin_unlock(PG_locked, &page->flags);
358 /* Interrupts must be disabled (for the fallback code to work right) */
359 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
360 void *freelist_old, unsigned long counters_old,
361 void *freelist_new, unsigned long counters_new,
364 VM_BUG_ON(!irqs_disabled());
365 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
366 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
367 if (s->flags & __CMPXCHG_DOUBLE) {
368 if (cmpxchg_double(&page->freelist, &page->counters,
369 freelist_old, counters_old,
370 freelist_new, counters_new))
376 if (page->freelist == freelist_old &&
377 page->counters == counters_old) {
378 page->freelist = freelist_new;
379 page->counters = counters_new;
387 stat(s, CMPXCHG_DOUBLE_FAIL);
389 #ifdef SLUB_DEBUG_CMPXCHG
390 pr_info("%s %s: cmpxchg double redo ", n, s->name);
396 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
397 void *freelist_old, unsigned long counters_old,
398 void *freelist_new, unsigned long counters_new,
401 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
402 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
403 if (s->flags & __CMPXCHG_DOUBLE) {
404 if (cmpxchg_double(&page->freelist, &page->counters,
405 freelist_old, counters_old,
406 freelist_new, counters_new))
413 local_irq_save(flags);
415 if (page->freelist == freelist_old &&
416 page->counters == counters_old) {
417 page->freelist = freelist_new;
418 page->counters = counters_new;
420 local_irq_restore(flags);
424 local_irq_restore(flags);
428 stat(s, CMPXCHG_DOUBLE_FAIL);
430 #ifdef SLUB_DEBUG_CMPXCHG
431 pr_info("%s %s: cmpxchg double redo ", n, s->name);
437 #ifdef CONFIG_SLUB_DEBUG
439 * Determine a map of object in use on a page.
441 * Node listlock must be held to guarantee that the page does
442 * not vanish from under us.
444 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
447 void *addr = page_address(page);
449 for (p = page->freelist; p; p = get_freepointer(s, p))
450 set_bit(slab_index(p, s, addr), map);
453 static inline unsigned int size_from_object(struct kmem_cache *s)
455 if (s->flags & SLAB_RED_ZONE)
456 return s->size - s->red_left_pad;
461 static inline void *restore_red_left(struct kmem_cache *s, void *p)
463 if (s->flags & SLAB_RED_ZONE)
464 p -= s->red_left_pad;
472 #if defined(CONFIG_SLUB_DEBUG_ON)
473 static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
475 static slab_flags_t slub_debug;
478 static char *slub_debug_slabs;
479 static int disable_higher_order_debug;
482 * slub is about to manipulate internal object metadata. This memory lies
483 * outside the range of the allocated object, so accessing it would normally
484 * be reported by kasan as a bounds error. metadata_access_enable() is used
485 * to tell kasan that these accesses are OK.
487 static inline void metadata_access_enable(void)
489 kasan_disable_current();
492 static inline void metadata_access_disable(void)
494 kasan_enable_current();
501 /* Verify that a pointer has an address that is valid within a slab page */
502 static inline int check_valid_pointer(struct kmem_cache *s,
503 struct page *page, void *object)
510 base = page_address(page);
511 object = restore_red_left(s, object);
512 if (object < base || object >= base + page->objects * s->size ||
513 (object - base) % s->size) {
520 static void print_section(char *level, char *text, u8 *addr,
523 metadata_access_enable();
524 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
526 metadata_access_disable();
529 static struct track *get_track(struct kmem_cache *s, void *object,
530 enum track_item alloc)
535 p = object + s->offset + sizeof(void *);
537 p = object + s->inuse;
542 static void set_track(struct kmem_cache *s, void *object,
543 enum track_item alloc, unsigned long addr)
545 struct track *p = get_track(s, object, alloc);
548 #ifdef CONFIG_STACKTRACE
549 struct stack_trace trace;
552 trace.nr_entries = 0;
553 trace.max_entries = TRACK_ADDRS_COUNT;
554 trace.entries = p->addrs;
556 metadata_access_enable();
557 save_stack_trace(&trace);
558 metadata_access_disable();
560 /* See rant in lockdep.c */
561 if (trace.nr_entries != 0 &&
562 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
565 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
569 p->cpu = smp_processor_id();
570 p->pid = current->pid;
573 memset(p, 0, sizeof(struct track));
576 static void init_tracking(struct kmem_cache *s, void *object)
578 if (!(s->flags & SLAB_STORE_USER))
581 set_track(s, object, TRACK_FREE, 0UL);
582 set_track(s, object, TRACK_ALLOC, 0UL);
585 static void print_track(const char *s, struct track *t, unsigned long pr_time)
590 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
591 s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
592 #ifdef CONFIG_STACKTRACE
595 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
597 pr_err("\t%pS\n", (void *)t->addrs[i]);
604 static void print_tracking(struct kmem_cache *s, void *object)
606 unsigned long pr_time = jiffies;
607 if (!(s->flags & SLAB_STORE_USER))
610 print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
611 print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
614 static void print_page_info(struct page *page)
616 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
617 page, page->objects, page->inuse, page->freelist, page->flags);
621 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
623 struct va_format vaf;
629 pr_err("=============================================================================\n");
630 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
631 pr_err("-----------------------------------------------------------------------------\n\n");
633 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
637 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
639 struct va_format vaf;
645 pr_err("FIX %s: %pV\n", s->name, &vaf);
649 static bool freelist_corrupted(struct kmem_cache *s, struct page *page,
650 void **freelist, void *nextfree)
652 if ((s->flags & SLAB_CONSISTENCY_CHECKS) &&
653 !check_valid_pointer(s, page, nextfree) && freelist) {
654 object_err(s, page, *freelist, "Freechain corrupt");
656 slab_fix(s, "Isolate corrupted freechain");
663 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
665 unsigned int off; /* Offset of last byte */
666 u8 *addr = page_address(page);
668 print_tracking(s, p);
670 print_page_info(page);
672 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
673 p, p - addr, get_freepointer(s, p));
675 if (s->flags & SLAB_RED_ZONE)
676 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
678 else if (p > addr + 16)
679 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
681 print_section(KERN_ERR, "Object ", p,
682 min_t(unsigned int, s->object_size, PAGE_SIZE));
683 if (s->flags & SLAB_RED_ZONE)
684 print_section(KERN_ERR, "Redzone ", p + s->object_size,
685 s->inuse - s->object_size);
688 off = s->offset + sizeof(void *);
692 if (s->flags & SLAB_STORE_USER)
693 off += 2 * sizeof(struct track);
695 off += kasan_metadata_size(s);
697 if (off != size_from_object(s))
698 /* Beginning of the filler is the free pointer */
699 print_section(KERN_ERR, "Padding ", p + off,
700 size_from_object(s) - off);
705 void object_err(struct kmem_cache *s, struct page *page,
706 u8 *object, char *reason)
708 slab_bug(s, "%s", reason);
709 print_trailer(s, page, object);
712 static __printf(3, 4) void slab_err(struct kmem_cache *s, struct page *page,
713 const char *fmt, ...)
719 vsnprintf(buf, sizeof(buf), fmt, args);
721 slab_bug(s, "%s", buf);
722 print_page_info(page);
726 static void init_object(struct kmem_cache *s, void *object, u8 val)
730 if (s->flags & SLAB_RED_ZONE)
731 memset(p - s->red_left_pad, val, s->red_left_pad);
733 if (s->flags & __OBJECT_POISON) {
734 memset(p, POISON_FREE, s->object_size - 1);
735 p[s->object_size - 1] = POISON_END;
738 if (s->flags & SLAB_RED_ZONE)
739 memset(p + s->object_size, val, s->inuse - s->object_size);
742 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
743 void *from, void *to)
745 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
746 memset(from, data, to - from);
749 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
750 u8 *object, char *what,
751 u8 *start, unsigned int value, unsigned int bytes)
756 metadata_access_enable();
757 fault = memchr_inv(start, value, bytes);
758 metadata_access_disable();
763 while (end > fault && end[-1] == value)
766 slab_bug(s, "%s overwritten", what);
767 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
768 fault, end - 1, fault[0], value);
769 print_trailer(s, page, object);
771 restore_bytes(s, what, value, fault, end);
779 * Bytes of the object to be managed.
780 * If the freepointer may overlay the object then the free
781 * pointer is the first word of the object.
783 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
786 * object + s->object_size
787 * Padding to reach word boundary. This is also used for Redzoning.
788 * Padding is extended by another word if Redzoning is enabled and
789 * object_size == inuse.
791 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
792 * 0xcc (RED_ACTIVE) for objects in use.
795 * Meta data starts here.
797 * A. Free pointer (if we cannot overwrite object on free)
798 * B. Tracking data for SLAB_STORE_USER
799 * C. Padding to reach required alignment boundary or at mininum
800 * one word if debugging is on to be able to detect writes
801 * before the word boundary.
803 * Padding is done using 0x5a (POISON_INUSE)
806 * Nothing is used beyond s->size.
808 * If slabcaches are merged then the object_size and inuse boundaries are mostly
809 * ignored. And therefore no slab options that rely on these boundaries
810 * may be used with merged slabcaches.
813 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
815 unsigned long off = s->inuse; /* The end of info */
818 /* Freepointer is placed after the object. */
819 off += sizeof(void *);
821 if (s->flags & SLAB_STORE_USER)
822 /* We also have user information there */
823 off += 2 * sizeof(struct track);
825 off += kasan_metadata_size(s);
827 if (size_from_object(s) == off)
830 return check_bytes_and_report(s, page, p, "Object padding",
831 p + off, POISON_INUSE, size_from_object(s) - off);
834 /* Check the pad bytes at the end of a slab page */
835 static int slab_pad_check(struct kmem_cache *s, struct page *page)
844 if (!(s->flags & SLAB_POISON))
847 start = page_address(page);
848 length = PAGE_SIZE << compound_order(page);
849 end = start + length;
850 remainder = length % s->size;
854 pad = end - remainder;
855 metadata_access_enable();
856 fault = memchr_inv(pad, POISON_INUSE, remainder);
857 metadata_access_disable();
860 while (end > fault && end[-1] == POISON_INUSE)
863 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
864 print_section(KERN_ERR, "Padding ", pad, remainder);
866 restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
870 static int check_object(struct kmem_cache *s, struct page *page,
871 void *object, u8 val)
874 u8 *endobject = object + s->object_size;
876 if (s->flags & SLAB_RED_ZONE) {
877 if (!check_bytes_and_report(s, page, object, "Left Redzone",
878 object - s->red_left_pad, val, s->red_left_pad))
881 if (!check_bytes_and_report(s, page, object, "Right Redzone",
882 endobject, val, s->inuse - s->object_size))
885 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
886 check_bytes_and_report(s, page, p, "Alignment padding",
887 endobject, POISON_INUSE,
888 s->inuse - s->object_size);
892 if (s->flags & SLAB_POISON) {
893 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
894 (!check_bytes_and_report(s, page, p, "Poison", p,
895 POISON_FREE, s->object_size - 1) ||
896 !check_bytes_and_report(s, page, p, "End Poison",
897 p + s->object_size - 1, POISON_END, 1)))
900 * check_pad_bytes cleans up on its own.
902 check_pad_bytes(s, page, p);
905 if (!s->offset && val == SLUB_RED_ACTIVE)
907 * Object and freepointer overlap. Cannot check
908 * freepointer while object is allocated.
912 /* Check free pointer validity */
913 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
914 object_err(s, page, p, "Freepointer corrupt");
916 * No choice but to zap it and thus lose the remainder
917 * of the free objects in this slab. May cause
918 * another error because the object count is now wrong.
920 set_freepointer(s, p, NULL);
926 static int check_slab(struct kmem_cache *s, struct page *page)
930 VM_BUG_ON(!irqs_disabled());
932 if (!PageSlab(page)) {
933 slab_err(s, page, "Not a valid slab page");
937 maxobj = order_objects(compound_order(page), s->size);
938 if (page->objects > maxobj) {
939 slab_err(s, page, "objects %u > max %u",
940 page->objects, maxobj);
943 if (page->inuse > page->objects) {
944 slab_err(s, page, "inuse %u > max %u",
945 page->inuse, page->objects);
948 /* Slab_pad_check fixes things up after itself */
949 slab_pad_check(s, page);
954 * Determine if a certain object on a page is on the freelist. Must hold the
955 * slab lock to guarantee that the chains are in a consistent state.
957 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
965 while (fp && nr <= page->objects) {
968 if (!check_valid_pointer(s, page, fp)) {
970 object_err(s, page, object,
971 "Freechain corrupt");
972 set_freepointer(s, object, NULL);
974 slab_err(s, page, "Freepointer corrupt");
975 page->freelist = NULL;
976 page->inuse = page->objects;
977 slab_fix(s, "Freelist cleared");
983 fp = get_freepointer(s, object);
987 max_objects = order_objects(compound_order(page), s->size);
988 if (max_objects > MAX_OBJS_PER_PAGE)
989 max_objects = MAX_OBJS_PER_PAGE;
991 if (page->objects != max_objects) {
992 slab_err(s, page, "Wrong number of objects. Found %d but should be %d",
993 page->objects, max_objects);
994 page->objects = max_objects;
995 slab_fix(s, "Number of objects adjusted.");
997 if (page->inuse != page->objects - nr) {
998 slab_err(s, page, "Wrong object count. Counter is %d but counted were %d",
999 page->inuse, page->objects - nr);
1000 page->inuse = page->objects - nr;
1001 slab_fix(s, "Object count adjusted.");
1003 return search == NULL;
1006 static void trace(struct kmem_cache *s, struct page *page, void *object,
1009 if (s->flags & SLAB_TRACE) {
1010 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1012 alloc ? "alloc" : "free",
1013 object, page->inuse,
1017 print_section(KERN_INFO, "Object ", (void *)object,
1025 * Tracking of fully allocated slabs for debugging purposes.
1027 static void add_full(struct kmem_cache *s,
1028 struct kmem_cache_node *n, struct page *page)
1030 if (!(s->flags & SLAB_STORE_USER))
1033 lockdep_assert_held(&n->list_lock);
1034 list_add(&page->lru, &n->full);
1037 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
1039 if (!(s->flags & SLAB_STORE_USER))
1042 lockdep_assert_held(&n->list_lock);
1043 list_del(&page->lru);
1046 /* Tracking of the number of slabs for debugging purposes */
1047 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1049 struct kmem_cache_node *n = get_node(s, node);
1051 return atomic_long_read(&n->nr_slabs);
1054 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1056 return atomic_long_read(&n->nr_slabs);
1059 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1061 struct kmem_cache_node *n = get_node(s, node);
1064 * May be called early in order to allocate a slab for the
1065 * kmem_cache_node structure. Solve the chicken-egg
1066 * dilemma by deferring the increment of the count during
1067 * bootstrap (see early_kmem_cache_node_alloc).
1070 atomic_long_inc(&n->nr_slabs);
1071 atomic_long_add(objects, &n->total_objects);
1074 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1076 struct kmem_cache_node *n = get_node(s, node);
1078 atomic_long_dec(&n->nr_slabs);
1079 atomic_long_sub(objects, &n->total_objects);
1082 /* Object debug checks for alloc/free paths */
1083 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1086 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1089 init_object(s, object, SLUB_RED_INACTIVE);
1090 init_tracking(s, object);
1093 static inline int alloc_consistency_checks(struct kmem_cache *s,
1095 void *object, unsigned long addr)
1097 if (!check_slab(s, page))
1100 if (!check_valid_pointer(s, page, object)) {
1101 object_err(s, page, object, "Freelist Pointer check fails");
1105 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1111 static noinline int alloc_debug_processing(struct kmem_cache *s,
1113 void *object, unsigned long addr)
1115 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1116 if (!alloc_consistency_checks(s, page, object, addr))
1120 /* Success perform special debug activities for allocs */
1121 if (s->flags & SLAB_STORE_USER)
1122 set_track(s, object, TRACK_ALLOC, addr);
1123 trace(s, page, object, 1);
1124 init_object(s, object, SLUB_RED_ACTIVE);
1128 if (PageSlab(page)) {
1130 * If this is a slab page then lets do the best we can
1131 * to avoid issues in the future. Marking all objects
1132 * as used avoids touching the remaining objects.
1134 slab_fix(s, "Marking all objects used");
1135 page->inuse = page->objects;
1136 page->freelist = NULL;
1141 static inline int free_consistency_checks(struct kmem_cache *s,
1142 struct page *page, void *object, unsigned long addr)
1144 if (!check_valid_pointer(s, page, object)) {
1145 slab_err(s, page, "Invalid object pointer 0x%p", object);
1149 if (on_freelist(s, page, object)) {
1150 object_err(s, page, object, "Object already free");
1154 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1157 if (unlikely(s != page->slab_cache)) {
1158 if (!PageSlab(page)) {
1159 slab_err(s, page, "Attempt to free object(0x%p) outside of slab",
1161 } else if (!page->slab_cache) {
1162 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1166 object_err(s, page, object,
1167 "page slab pointer corrupt.");
1173 /* Supports checking bulk free of a constructed freelist */
1174 static noinline int free_debug_processing(
1175 struct kmem_cache *s, struct page *page,
1176 void *head, void *tail, int bulk_cnt,
1179 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1180 void *object = head;
1182 unsigned long uninitialized_var(flags);
1185 spin_lock_irqsave(&n->list_lock, flags);
1188 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1189 if (!check_slab(s, page))
1196 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1197 if (!free_consistency_checks(s, page, object, addr))
1201 if (s->flags & SLAB_STORE_USER)
1202 set_track(s, object, TRACK_FREE, addr);
1203 trace(s, page, object, 0);
1204 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1205 init_object(s, object, SLUB_RED_INACTIVE);
1207 /* Reached end of constructed freelist yet? */
1208 if (object != tail) {
1209 object = get_freepointer(s, object);
1215 if (cnt != bulk_cnt)
1216 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1220 spin_unlock_irqrestore(&n->list_lock, flags);
1222 slab_fix(s, "Object at 0x%p not freed", object);
1226 static int __init setup_slub_debug(char *str)
1228 slub_debug = DEBUG_DEFAULT_FLAGS;
1229 if (*str++ != '=' || !*str)
1231 * No options specified. Switch on full debugging.
1237 * No options but restriction on slabs. This means full
1238 * debugging for slabs matching a pattern.
1245 * Switch off all debugging measures.
1250 * Determine which debug features should be switched on
1252 for (; *str && *str != ','; str++) {
1253 switch (tolower(*str)) {
1255 slub_debug |= SLAB_CONSISTENCY_CHECKS;
1258 slub_debug |= SLAB_RED_ZONE;
1261 slub_debug |= SLAB_POISON;
1264 slub_debug |= SLAB_STORE_USER;
1267 slub_debug |= SLAB_TRACE;
1270 slub_debug |= SLAB_FAILSLAB;
1274 * Avoid enabling debugging on caches if its minimum
1275 * order would increase as a result.
1277 disable_higher_order_debug = 1;
1280 pr_err("slub_debug option '%c' unknown. skipped\n",
1287 slub_debug_slabs = str + 1;
1292 __setup("slub_debug", setup_slub_debug);
1294 slab_flags_t kmem_cache_flags(unsigned int object_size,
1295 slab_flags_t flags, const char *name,
1296 void (*ctor)(void *))
1299 * Enable debugging if selected on the kernel commandline.
1301 if (slub_debug && (!slub_debug_slabs || (name &&
1302 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1303 flags |= slub_debug;
1307 #else /* !CONFIG_SLUB_DEBUG */
1308 static inline void setup_object_debug(struct kmem_cache *s,
1309 struct page *page, void *object) {}
1311 static inline int alloc_debug_processing(struct kmem_cache *s,
1312 struct page *page, void *object, unsigned long addr) { return 0; }
1314 static inline int free_debug_processing(
1315 struct kmem_cache *s, struct page *page,
1316 void *head, void *tail, int bulk_cnt,
1317 unsigned long addr) { return 0; }
1319 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1321 static inline int check_object(struct kmem_cache *s, struct page *page,
1322 void *object, u8 val) { return 1; }
1323 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1324 struct page *page) {}
1325 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1326 struct page *page) {}
1327 slab_flags_t kmem_cache_flags(unsigned int object_size,
1328 slab_flags_t flags, const char *name,
1329 void (*ctor)(void *))
1333 #define slub_debug 0
1335 #define disable_higher_order_debug 0
1337 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1339 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1341 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1343 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1346 static bool freelist_corrupted(struct kmem_cache *s, struct page *page,
1347 void **freelist, void *nextfree)
1351 #endif /* CONFIG_SLUB_DEBUG */
1354 * Hooks for other subsystems that check memory allocations. In a typical
1355 * production configuration these hooks all should produce no code at all.
1357 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1359 kmemleak_alloc(ptr, size, 1, flags);
1360 kasan_kmalloc_large(ptr, size, flags);
1363 static __always_inline void kfree_hook(void *x)
1366 kasan_kfree_large(x, _RET_IP_);
1369 static __always_inline bool slab_free_hook(struct kmem_cache *s, void *x)
1371 kmemleak_free_recursive(x, s->flags);
1374 * Trouble is that we may no longer disable interrupts in the fast path
1375 * So in order to make the debug calls that expect irqs to be
1376 * disabled we need to disable interrupts temporarily.
1378 #ifdef CONFIG_LOCKDEP
1380 unsigned long flags;
1382 local_irq_save(flags);
1383 debug_check_no_locks_freed(x, s->object_size);
1384 local_irq_restore(flags);
1387 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1388 debug_check_no_obj_freed(x, s->object_size);
1390 /* KASAN might put x into memory quarantine, delaying its reuse */
1391 return kasan_slab_free(s, x, _RET_IP_);
1394 static inline bool slab_free_freelist_hook(struct kmem_cache *s,
1395 void **head, void **tail)
1398 * Compiler cannot detect this function can be removed if slab_free_hook()
1399 * evaluates to nothing. Thus, catch all relevant config debug options here.
1401 #if defined(CONFIG_LOCKDEP) || \
1402 defined(CONFIG_DEBUG_KMEMLEAK) || \
1403 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1404 defined(CONFIG_KASAN)
1408 void *old_tail = *tail ? *tail : *head;
1410 /* Head and tail of the reconstructed freelist */
1416 next = get_freepointer(s, object);
1417 /* If object's reuse doesn't have to be delayed */
1418 if (!slab_free_hook(s, object)) {
1419 /* Move object to the new freelist */
1420 set_freepointer(s, object, *head);
1425 } while (object != old_tail);
1430 return *head != NULL;
1436 static void setup_object(struct kmem_cache *s, struct page *page,
1439 setup_object_debug(s, page, object);
1440 kasan_init_slab_obj(s, object);
1441 if (unlikely(s->ctor)) {
1442 kasan_unpoison_object_data(s, object);
1444 kasan_poison_object_data(s, object);
1449 * Slab allocation and freeing
1451 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1452 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1455 unsigned int order = oo_order(oo);
1457 if (node == NUMA_NO_NODE)
1458 page = alloc_pages(flags, order);
1460 page = __alloc_pages_node(node, flags, order);
1462 if (page && memcg_charge_slab(page, flags, order, s)) {
1463 __free_pages(page, order);
1470 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1471 /* Pre-initialize the random sequence cache */
1472 static int init_cache_random_seq(struct kmem_cache *s)
1474 unsigned int count = oo_objects(s->oo);
1477 /* Bailout if already initialised */
1481 err = cache_random_seq_create(s, count, GFP_KERNEL);
1483 pr_err("SLUB: Unable to initialize free list for %s\n",
1488 /* Transform to an offset on the set of pages */
1489 if (s->random_seq) {
1492 for (i = 0; i < count; i++)
1493 s->random_seq[i] *= s->size;
1498 /* Initialize each random sequence freelist per cache */
1499 static void __init init_freelist_randomization(void)
1501 struct kmem_cache *s;
1503 mutex_lock(&slab_mutex);
1505 list_for_each_entry(s, &slab_caches, list)
1506 init_cache_random_seq(s);
1508 mutex_unlock(&slab_mutex);
1511 /* Get the next entry on the pre-computed freelist randomized */
1512 static void *next_freelist_entry(struct kmem_cache *s, struct page *page,
1513 unsigned long *pos, void *start,
1514 unsigned long page_limit,
1515 unsigned long freelist_count)
1520 * If the target page allocation failed, the number of objects on the
1521 * page might be smaller than the usual size defined by the cache.
1524 idx = s->random_seq[*pos];
1526 if (*pos >= freelist_count)
1528 } while (unlikely(idx >= page_limit));
1530 return (char *)start + idx;
1533 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1534 static bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1539 unsigned long idx, pos, page_limit, freelist_count;
1541 if (page->objects < 2 || !s->random_seq)
1544 freelist_count = oo_objects(s->oo);
1545 pos = get_random_int() % freelist_count;
1547 page_limit = page->objects * s->size;
1548 start = fixup_red_left(s, page_address(page));
1550 /* First entry is used as the base of the freelist */
1551 cur = next_freelist_entry(s, page, &pos, start, page_limit,
1553 page->freelist = cur;
1555 for (idx = 1; idx < page->objects; idx++) {
1556 setup_object(s, page, cur);
1557 next = next_freelist_entry(s, page, &pos, start, page_limit,
1559 set_freepointer(s, cur, next);
1562 setup_object(s, page, cur);
1563 set_freepointer(s, cur, NULL);
1568 static inline int init_cache_random_seq(struct kmem_cache *s)
1572 static inline void init_freelist_randomization(void) { }
1573 static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1577 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1579 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1582 struct kmem_cache_order_objects oo = s->oo;
1588 flags &= gfp_allowed_mask;
1590 if (gfpflags_allow_blocking(flags))
1593 flags |= s->allocflags;
1596 * Let the initial higher-order allocation fail under memory pressure
1597 * so we fall-back to the minimum order allocation.
1599 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1600 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1601 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
1603 page = alloc_slab_page(s, alloc_gfp, node, oo);
1604 if (unlikely(!page)) {
1608 * Allocation may have failed due to fragmentation.
1609 * Try a lower order alloc if possible
1611 page = alloc_slab_page(s, alloc_gfp, node, oo);
1612 if (unlikely(!page))
1614 stat(s, ORDER_FALLBACK);
1617 page->objects = oo_objects(oo);
1619 order = compound_order(page);
1620 page->slab_cache = s;
1621 __SetPageSlab(page);
1622 if (page_is_pfmemalloc(page))
1623 SetPageSlabPfmemalloc(page);
1625 start = page_address(page);
1627 if (unlikely(s->flags & SLAB_POISON))
1628 memset(start, POISON_INUSE, PAGE_SIZE << order);
1630 kasan_poison_slab(page);
1632 shuffle = shuffle_freelist(s, page);
1635 for_each_object_idx(p, idx, s, start, page->objects) {
1636 setup_object(s, page, p);
1637 if (likely(idx < page->objects))
1638 set_freepointer(s, p, p + s->size);
1640 set_freepointer(s, p, NULL);
1642 page->freelist = fixup_red_left(s, start);
1645 page->inuse = page->objects;
1649 if (gfpflags_allow_blocking(flags))
1650 local_irq_disable();
1654 mod_lruvec_page_state(page,
1655 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1656 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1659 inc_slabs_node(s, page_to_nid(page), page->objects);
1664 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1666 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1667 gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK;
1668 flags &= ~GFP_SLAB_BUG_MASK;
1669 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
1670 invalid_mask, &invalid_mask, flags, &flags);
1674 return allocate_slab(s,
1675 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1678 static void __free_slab(struct kmem_cache *s, struct page *page)
1680 int order = compound_order(page);
1681 int pages = 1 << order;
1683 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1686 slab_pad_check(s, page);
1687 for_each_object(p, s, page_address(page),
1689 check_object(s, page, p, SLUB_RED_INACTIVE);
1692 mod_lruvec_page_state(page,
1693 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1694 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1697 __ClearPageSlabPfmemalloc(page);
1698 __ClearPageSlab(page);
1700 page->mapping = NULL;
1701 if (current->reclaim_state)
1702 current->reclaim_state->reclaimed_slab += pages;
1703 memcg_uncharge_slab(page, order, s);
1704 __free_pages(page, order);
1707 static void rcu_free_slab(struct rcu_head *h)
1709 struct page *page = container_of(h, struct page, rcu_head);
1711 __free_slab(page->slab_cache, page);
1714 static void free_slab(struct kmem_cache *s, struct page *page)
1716 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) {
1717 call_rcu(&page->rcu_head, rcu_free_slab);
1719 __free_slab(s, page);
1722 static void discard_slab(struct kmem_cache *s, struct page *page)
1724 dec_slabs_node(s, page_to_nid(page), page->objects);
1729 * Management of partially allocated slabs.
1732 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1735 if (tail == DEACTIVATE_TO_TAIL)
1736 list_add_tail(&page->lru, &n->partial);
1738 list_add(&page->lru, &n->partial);
1741 static inline void add_partial(struct kmem_cache_node *n,
1742 struct page *page, int tail)
1744 lockdep_assert_held(&n->list_lock);
1745 __add_partial(n, page, tail);
1748 static inline void remove_partial(struct kmem_cache_node *n,
1751 lockdep_assert_held(&n->list_lock);
1752 list_del(&page->lru);
1757 * Remove slab from the partial list, freeze it and
1758 * return the pointer to the freelist.
1760 * Returns a list of objects or NULL if it fails.
1762 static inline void *acquire_slab(struct kmem_cache *s,
1763 struct kmem_cache_node *n, struct page *page,
1764 int mode, int *objects)
1767 unsigned long counters;
1770 lockdep_assert_held(&n->list_lock);
1773 * Zap the freelist and set the frozen bit.
1774 * The old freelist is the list of objects for the
1775 * per cpu allocation list.
1777 freelist = page->freelist;
1778 counters = page->counters;
1779 new.counters = counters;
1780 *objects = new.objects - new.inuse;
1782 new.inuse = page->objects;
1783 new.freelist = NULL;
1785 new.freelist = freelist;
1788 VM_BUG_ON(new.frozen);
1791 if (!__cmpxchg_double_slab(s, page,
1793 new.freelist, new.counters,
1797 remove_partial(n, page);
1802 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1803 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1806 * Try to allocate a partial slab from a specific node.
1808 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1809 struct kmem_cache_cpu *c, gfp_t flags)
1811 struct page *page, *page2;
1812 void *object = NULL;
1813 unsigned int available = 0;
1817 * Racy check. If we mistakenly see no partial slabs then we
1818 * just allocate an empty slab. If we mistakenly try to get a
1819 * partial slab and there is none available then get_partials()
1822 if (!n || !n->nr_partial)
1825 spin_lock(&n->list_lock);
1826 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1829 if (!pfmemalloc_match(page, flags))
1832 t = acquire_slab(s, n, page, object == NULL, &objects);
1836 available += objects;
1839 stat(s, ALLOC_FROM_PARTIAL);
1842 put_cpu_partial(s, page, 0);
1843 stat(s, CPU_PARTIAL_NODE);
1845 if (!kmem_cache_has_cpu_partial(s)
1846 || available > slub_cpu_partial(s) / 2)
1850 spin_unlock(&n->list_lock);
1855 * Get a page from somewhere. Search in increasing NUMA distances.
1857 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1858 struct kmem_cache_cpu *c)
1861 struct zonelist *zonelist;
1864 enum zone_type high_zoneidx = gfp_zone(flags);
1866 unsigned int cpuset_mems_cookie;
1869 * The defrag ratio allows a configuration of the tradeoffs between
1870 * inter node defragmentation and node local allocations. A lower
1871 * defrag_ratio increases the tendency to do local allocations
1872 * instead of attempting to obtain partial slabs from other nodes.
1874 * If the defrag_ratio is set to 0 then kmalloc() always
1875 * returns node local objects. If the ratio is higher then kmalloc()
1876 * may return off node objects because partial slabs are obtained
1877 * from other nodes and filled up.
1879 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
1880 * (which makes defrag_ratio = 1000) then every (well almost)
1881 * allocation will first attempt to defrag slab caches on other nodes.
1882 * This means scanning over all nodes to look for partial slabs which
1883 * may be expensive if we do it every time we are trying to find a slab
1884 * with available objects.
1886 if (!s->remote_node_defrag_ratio ||
1887 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1891 cpuset_mems_cookie = read_mems_allowed_begin();
1892 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1893 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1894 struct kmem_cache_node *n;
1896 n = get_node(s, zone_to_nid(zone));
1898 if (n && cpuset_zone_allowed(zone, flags) &&
1899 n->nr_partial > s->min_partial) {
1900 object = get_partial_node(s, n, c, flags);
1903 * Don't check read_mems_allowed_retry()
1904 * here - if mems_allowed was updated in
1905 * parallel, that was a harmless race
1906 * between allocation and the cpuset
1913 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1919 * Get a partial page, lock it and return it.
1921 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1922 struct kmem_cache_cpu *c)
1925 int searchnode = node;
1927 if (node == NUMA_NO_NODE)
1928 searchnode = numa_mem_id();
1930 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1931 if (object || node != NUMA_NO_NODE)
1934 return get_any_partial(s, flags, c);
1937 #ifdef CONFIG_PREEMPT
1939 * Calculate the next globally unique transaction for disambiguiation
1940 * during cmpxchg. The transactions start with the cpu number and are then
1941 * incremented by CONFIG_NR_CPUS.
1943 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1946 * No preemption supported therefore also no need to check for
1952 static inline unsigned long next_tid(unsigned long tid)
1954 return tid + TID_STEP;
1957 static inline unsigned int tid_to_cpu(unsigned long tid)
1959 return tid % TID_STEP;
1962 static inline unsigned long tid_to_event(unsigned long tid)
1964 return tid / TID_STEP;
1967 static inline unsigned int init_tid(int cpu)
1972 static inline void note_cmpxchg_failure(const char *n,
1973 const struct kmem_cache *s, unsigned long tid)
1975 #ifdef SLUB_DEBUG_CMPXCHG
1976 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1978 pr_info("%s %s: cmpxchg redo ", n, s->name);
1980 #ifdef CONFIG_PREEMPT
1981 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1982 pr_warn("due to cpu change %d -> %d\n",
1983 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1986 if (tid_to_event(tid) != tid_to_event(actual_tid))
1987 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1988 tid_to_event(tid), tid_to_event(actual_tid));
1990 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1991 actual_tid, tid, next_tid(tid));
1993 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1996 static void init_kmem_cache_cpus(struct kmem_cache *s)
2000 for_each_possible_cpu(cpu)
2001 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
2005 * Remove the cpu slab
2007 static void deactivate_slab(struct kmem_cache *s, struct page *page,
2008 void *freelist, struct kmem_cache_cpu *c)
2010 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
2011 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
2013 enum slab_modes l = M_NONE, m = M_NONE;
2015 int tail = DEACTIVATE_TO_HEAD;
2019 if (page->freelist) {
2020 stat(s, DEACTIVATE_REMOTE_FREES);
2021 tail = DEACTIVATE_TO_TAIL;
2025 * Stage one: Free all available per cpu objects back
2026 * to the page freelist while it is still frozen. Leave the
2029 * There is no need to take the list->lock because the page
2032 while (freelist && (nextfree = get_freepointer(s, freelist))) {
2034 unsigned long counters;
2037 * If 'nextfree' is invalid, it is possible that the object at
2038 * 'freelist' is already corrupted. So isolate all objects
2039 * starting at 'freelist'.
2041 if (freelist_corrupted(s, page, &freelist, nextfree))
2045 prior = page->freelist;
2046 counters = page->counters;
2047 set_freepointer(s, freelist, prior);
2048 new.counters = counters;
2050 VM_BUG_ON(!new.frozen);
2052 } while (!__cmpxchg_double_slab(s, page,
2054 freelist, new.counters,
2055 "drain percpu freelist"));
2057 freelist = nextfree;
2061 * Stage two: Ensure that the page is unfrozen while the
2062 * list presence reflects the actual number of objects
2065 * We setup the list membership and then perform a cmpxchg
2066 * with the count. If there is a mismatch then the page
2067 * is not unfrozen but the page is on the wrong list.
2069 * Then we restart the process which may have to remove
2070 * the page from the list that we just put it on again
2071 * because the number of objects in the slab may have
2076 old.freelist = page->freelist;
2077 old.counters = page->counters;
2078 VM_BUG_ON(!old.frozen);
2080 /* Determine target state of the slab */
2081 new.counters = old.counters;
2084 set_freepointer(s, freelist, old.freelist);
2085 new.freelist = freelist;
2087 new.freelist = old.freelist;
2091 if (!new.inuse && n->nr_partial >= s->min_partial)
2093 else if (new.freelist) {
2098 * Taking the spinlock removes the possiblity
2099 * that acquire_slab() will see a slab page that
2102 spin_lock(&n->list_lock);
2106 if (kmem_cache_debug(s) && !lock) {
2109 * This also ensures that the scanning of full
2110 * slabs from diagnostic functions will not see
2113 spin_lock(&n->list_lock);
2121 remove_partial(n, page);
2123 else if (l == M_FULL)
2125 remove_full(s, n, page);
2127 if (m == M_PARTIAL) {
2129 add_partial(n, page, tail);
2132 } else if (m == M_FULL) {
2134 stat(s, DEACTIVATE_FULL);
2135 add_full(s, n, page);
2141 if (!__cmpxchg_double_slab(s, page,
2142 old.freelist, old.counters,
2143 new.freelist, new.counters,
2148 spin_unlock(&n->list_lock);
2151 stat(s, DEACTIVATE_EMPTY);
2152 discard_slab(s, page);
2161 * Unfreeze all the cpu partial slabs.
2163 * This function must be called with interrupts disabled
2164 * for the cpu using c (or some other guarantee must be there
2165 * to guarantee no concurrent accesses).
2167 static void unfreeze_partials(struct kmem_cache *s,
2168 struct kmem_cache_cpu *c)
2170 #ifdef CONFIG_SLUB_CPU_PARTIAL
2171 struct kmem_cache_node *n = NULL, *n2 = NULL;
2172 struct page *page, *discard_page = NULL;
2174 while ((page = c->partial)) {
2178 c->partial = page->next;
2180 n2 = get_node(s, page_to_nid(page));
2183 spin_unlock(&n->list_lock);
2186 spin_lock(&n->list_lock);
2191 old.freelist = page->freelist;
2192 old.counters = page->counters;
2193 VM_BUG_ON(!old.frozen);
2195 new.counters = old.counters;
2196 new.freelist = old.freelist;
2200 } while (!__cmpxchg_double_slab(s, page,
2201 old.freelist, old.counters,
2202 new.freelist, new.counters,
2203 "unfreezing slab"));
2205 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2206 page->next = discard_page;
2207 discard_page = page;
2209 add_partial(n, page, DEACTIVATE_TO_TAIL);
2210 stat(s, FREE_ADD_PARTIAL);
2215 spin_unlock(&n->list_lock);
2217 while (discard_page) {
2218 page = discard_page;
2219 discard_page = discard_page->next;
2221 stat(s, DEACTIVATE_EMPTY);
2222 discard_slab(s, page);
2229 * Put a page that was just frozen (in __slab_free) into a partial page
2230 * slot if available.
2232 * If we did not find a slot then simply move all the partials to the
2233 * per node partial list.
2235 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2237 #ifdef CONFIG_SLUB_CPU_PARTIAL
2238 struct page *oldpage;
2246 oldpage = this_cpu_read(s->cpu_slab->partial);
2249 pobjects = oldpage->pobjects;
2250 pages = oldpage->pages;
2251 if (drain && pobjects > s->cpu_partial) {
2252 unsigned long flags;
2254 * partial array is full. Move the existing
2255 * set to the per node partial list.
2257 local_irq_save(flags);
2258 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2259 local_irq_restore(flags);
2263 stat(s, CPU_PARTIAL_DRAIN);
2268 pobjects += page->objects - page->inuse;
2270 page->pages = pages;
2271 page->pobjects = pobjects;
2272 page->next = oldpage;
2274 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2276 if (unlikely(!s->cpu_partial)) {
2277 unsigned long flags;
2279 local_irq_save(flags);
2280 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2281 local_irq_restore(flags);
2287 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2289 stat(s, CPUSLAB_FLUSH);
2290 deactivate_slab(s, c->page, c->freelist, c);
2292 c->tid = next_tid(c->tid);
2298 * Called from IPI handler with interrupts disabled.
2300 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2302 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2308 unfreeze_partials(s, c);
2312 static void flush_cpu_slab(void *d)
2314 struct kmem_cache *s = d;
2316 __flush_cpu_slab(s, smp_processor_id());
2319 static bool has_cpu_slab(int cpu, void *info)
2321 struct kmem_cache *s = info;
2322 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2324 return c->page || slub_percpu_partial(c);
2327 static void flush_all(struct kmem_cache *s)
2329 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2333 * Use the cpu notifier to insure that the cpu slabs are flushed when
2336 static int slub_cpu_dead(unsigned int cpu)
2338 struct kmem_cache *s;
2339 unsigned long flags;
2341 mutex_lock(&slab_mutex);
2342 list_for_each_entry(s, &slab_caches, list) {
2343 local_irq_save(flags);
2344 __flush_cpu_slab(s, cpu);
2345 local_irq_restore(flags);
2347 mutex_unlock(&slab_mutex);
2352 * Check if the objects in a per cpu structure fit numa
2353 * locality expectations.
2355 static inline int node_match(struct page *page, int node)
2358 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2364 #ifdef CONFIG_SLUB_DEBUG
2365 static int count_free(struct page *page)
2367 return page->objects - page->inuse;
2370 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2372 return atomic_long_read(&n->total_objects);
2374 #endif /* CONFIG_SLUB_DEBUG */
2376 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2377 static unsigned long count_partial(struct kmem_cache_node *n,
2378 int (*get_count)(struct page *))
2380 unsigned long flags;
2381 unsigned long x = 0;
2384 spin_lock_irqsave(&n->list_lock, flags);
2385 list_for_each_entry(page, &n->partial, lru)
2386 x += get_count(page);
2387 spin_unlock_irqrestore(&n->list_lock, flags);
2390 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2392 static noinline void
2393 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2395 #ifdef CONFIG_SLUB_DEBUG
2396 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2397 DEFAULT_RATELIMIT_BURST);
2399 struct kmem_cache_node *n;
2401 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2404 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2405 nid, gfpflags, &gfpflags);
2406 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2407 s->name, s->object_size, s->size, oo_order(s->oo),
2410 if (oo_order(s->min) > get_order(s->object_size))
2411 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2414 for_each_kmem_cache_node(s, node, n) {
2415 unsigned long nr_slabs;
2416 unsigned long nr_objs;
2417 unsigned long nr_free;
2419 nr_free = count_partial(n, count_free);
2420 nr_slabs = node_nr_slabs(n);
2421 nr_objs = node_nr_objs(n);
2423 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2424 node, nr_slabs, nr_objs, nr_free);
2429 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2430 int node, struct kmem_cache_cpu **pc)
2433 struct kmem_cache_cpu *c = *pc;
2436 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
2438 freelist = get_partial(s, flags, node, c);
2443 page = new_slab(s, flags, node);
2445 c = raw_cpu_ptr(s->cpu_slab);
2450 * No other reference to the page yet so we can
2451 * muck around with it freely without cmpxchg
2453 freelist = page->freelist;
2454 page->freelist = NULL;
2456 stat(s, ALLOC_SLAB);
2465 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2467 if (unlikely(PageSlabPfmemalloc(page)))
2468 return gfp_pfmemalloc_allowed(gfpflags);
2474 * Check the page->freelist of a page and either transfer the freelist to the
2475 * per cpu freelist or deactivate the page.
2477 * The page is still frozen if the return value is not NULL.
2479 * If this function returns NULL then the page has been unfrozen.
2481 * This function must be called with interrupt disabled.
2483 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2486 unsigned long counters;
2490 freelist = page->freelist;
2491 counters = page->counters;
2493 new.counters = counters;
2494 VM_BUG_ON(!new.frozen);
2496 new.inuse = page->objects;
2497 new.frozen = freelist != NULL;
2499 } while (!__cmpxchg_double_slab(s, page,
2508 * Slow path. The lockless freelist is empty or we need to perform
2511 * Processing is still very fast if new objects have been freed to the
2512 * regular freelist. In that case we simply take over the regular freelist
2513 * as the lockless freelist and zap the regular freelist.
2515 * If that is not working then we fall back to the partial lists. We take the
2516 * first element of the freelist as the object to allocate now and move the
2517 * rest of the freelist to the lockless freelist.
2519 * And if we were unable to get a new slab from the partial slab lists then
2520 * we need to allocate a new slab. This is the slowest path since it involves
2521 * a call to the page allocator and the setup of a new slab.
2523 * Version of __slab_alloc to use when we know that interrupts are
2524 * already disabled (which is the case for bulk allocation).
2526 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2527 unsigned long addr, struct kmem_cache_cpu *c)
2535 * if the node is not online or has no normal memory, just
2536 * ignore the node constraint
2538 if (unlikely(node != NUMA_NO_NODE &&
2539 !node_state(node, N_NORMAL_MEMORY)))
2540 node = NUMA_NO_NODE;
2545 if (unlikely(!node_match(page, node))) {
2547 * same as above but node_match() being false already
2548 * implies node != NUMA_NO_NODE
2550 if (!node_state(node, N_NORMAL_MEMORY)) {
2551 node = NUMA_NO_NODE;
2554 stat(s, ALLOC_NODE_MISMATCH);
2555 deactivate_slab(s, page, c->freelist, c);
2561 * By rights, we should be searching for a slab page that was
2562 * PFMEMALLOC but right now, we are losing the pfmemalloc
2563 * information when the page leaves the per-cpu allocator
2565 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2566 deactivate_slab(s, page, c->freelist, c);
2570 /* must check again c->freelist in case of cpu migration or IRQ */
2571 freelist = c->freelist;
2575 freelist = get_freelist(s, page);
2579 stat(s, DEACTIVATE_BYPASS);
2583 stat(s, ALLOC_REFILL);
2587 * freelist is pointing to the list of objects to be used.
2588 * page is pointing to the page from which the objects are obtained.
2589 * That page must be frozen for per cpu allocations to work.
2591 VM_BUG_ON(!c->page->frozen);
2592 c->freelist = get_freepointer(s, freelist);
2593 c->tid = next_tid(c->tid);
2598 if (slub_percpu_partial(c)) {
2599 page = c->page = slub_percpu_partial(c);
2600 slub_set_percpu_partial(c, page);
2601 stat(s, CPU_PARTIAL_ALLOC);
2605 freelist = new_slab_objects(s, gfpflags, node, &c);
2607 if (unlikely(!freelist)) {
2608 slab_out_of_memory(s, gfpflags, node);
2613 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2616 /* Only entered in the debug case */
2617 if (kmem_cache_debug(s) &&
2618 !alloc_debug_processing(s, page, freelist, addr))
2619 goto new_slab; /* Slab failed checks. Next slab needed */
2621 deactivate_slab(s, page, get_freepointer(s, freelist), c);
2626 * Another one that disabled interrupt and compensates for possible
2627 * cpu changes by refetching the per cpu area pointer.
2629 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2630 unsigned long addr, struct kmem_cache_cpu *c)
2633 unsigned long flags;
2635 local_irq_save(flags);
2636 #ifdef CONFIG_PREEMPT
2638 * We may have been preempted and rescheduled on a different
2639 * cpu before disabling interrupts. Need to reload cpu area
2642 c = this_cpu_ptr(s->cpu_slab);
2645 p = ___slab_alloc(s, gfpflags, node, addr, c);
2646 local_irq_restore(flags);
2651 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2652 * have the fastpath folded into their functions. So no function call
2653 * overhead for requests that can be satisfied on the fastpath.
2655 * The fastpath works by first checking if the lockless freelist can be used.
2656 * If not then __slab_alloc is called for slow processing.
2658 * Otherwise we can simply pick the next object from the lockless free list.
2660 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2661 gfp_t gfpflags, int node, unsigned long addr)
2664 struct kmem_cache_cpu *c;
2668 s = slab_pre_alloc_hook(s, gfpflags);
2673 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2674 * enabled. We may switch back and forth between cpus while
2675 * reading from one cpu area. That does not matter as long
2676 * as we end up on the original cpu again when doing the cmpxchg.
2678 * We should guarantee that tid and kmem_cache are retrieved on
2679 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2680 * to check if it is matched or not.
2683 tid = this_cpu_read(s->cpu_slab->tid);
2684 c = raw_cpu_ptr(s->cpu_slab);
2685 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2686 unlikely(tid != READ_ONCE(c->tid)));
2689 * Irqless object alloc/free algorithm used here depends on sequence
2690 * of fetching cpu_slab's data. tid should be fetched before anything
2691 * on c to guarantee that object and page associated with previous tid
2692 * won't be used with current tid. If we fetch tid first, object and
2693 * page could be one associated with next tid and our alloc/free
2694 * request will be failed. In this case, we will retry. So, no problem.
2699 * The transaction ids are globally unique per cpu and per operation on
2700 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2701 * occurs on the right processor and that there was no operation on the
2702 * linked list in between.
2705 object = c->freelist;
2707 if (unlikely(!object || !node_match(page, node))) {
2708 object = __slab_alloc(s, gfpflags, node, addr, c);
2709 stat(s, ALLOC_SLOWPATH);
2711 void *next_object = get_freepointer_safe(s, object);
2714 * The cmpxchg will only match if there was no additional
2715 * operation and if we are on the right processor.
2717 * The cmpxchg does the following atomically (without lock
2719 * 1. Relocate first pointer to the current per cpu area.
2720 * 2. Verify that tid and freelist have not been changed
2721 * 3. If they were not changed replace tid and freelist
2723 * Since this is without lock semantics the protection is only
2724 * against code executing on this cpu *not* from access by
2727 if (unlikely(!this_cpu_cmpxchg_double(
2728 s->cpu_slab->freelist, s->cpu_slab->tid,
2730 next_object, next_tid(tid)))) {
2732 note_cmpxchg_failure("slab_alloc", s, tid);
2735 prefetch_freepointer(s, next_object);
2736 stat(s, ALLOC_FASTPATH);
2739 if (unlikely(gfpflags & __GFP_ZERO) && object)
2740 memset(object, 0, s->object_size);
2742 slab_post_alloc_hook(s, gfpflags, 1, &object);
2747 static __always_inline void *slab_alloc(struct kmem_cache *s,
2748 gfp_t gfpflags, unsigned long addr)
2750 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2753 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2755 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2757 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2762 EXPORT_SYMBOL(kmem_cache_alloc);
2764 #ifdef CONFIG_TRACING
2765 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2767 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2768 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2769 kasan_kmalloc(s, ret, size, gfpflags);
2772 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2776 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2778 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2780 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2781 s->object_size, s->size, gfpflags, node);
2785 EXPORT_SYMBOL(kmem_cache_alloc_node);
2787 #ifdef CONFIG_TRACING
2788 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2790 int node, size_t size)
2792 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2794 trace_kmalloc_node(_RET_IP_, ret,
2795 size, s->size, gfpflags, node);
2797 kasan_kmalloc(s, ret, size, gfpflags);
2800 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2805 * Slow path handling. This may still be called frequently since objects
2806 * have a longer lifetime than the cpu slabs in most processing loads.
2808 * So we still attempt to reduce cache line usage. Just take the slab
2809 * lock and free the item. If there is no additional partial page
2810 * handling required then we can return immediately.
2812 static void __slab_free(struct kmem_cache *s, struct page *page,
2813 void *head, void *tail, int cnt,
2820 unsigned long counters;
2821 struct kmem_cache_node *n = NULL;
2822 unsigned long uninitialized_var(flags);
2824 stat(s, FREE_SLOWPATH);
2826 if (kmem_cache_debug(s) &&
2827 !free_debug_processing(s, page, head, tail, cnt, addr))
2832 spin_unlock_irqrestore(&n->list_lock, flags);
2835 prior = page->freelist;
2836 counters = page->counters;
2837 set_freepointer(s, tail, prior);
2838 new.counters = counters;
2839 was_frozen = new.frozen;
2841 if ((!new.inuse || !prior) && !was_frozen) {
2843 if (kmem_cache_has_cpu_partial(s) && !prior) {
2846 * Slab was on no list before and will be
2848 * We can defer the list move and instead
2853 } else { /* Needs to be taken off a list */
2855 n = get_node(s, page_to_nid(page));
2857 * Speculatively acquire the list_lock.
2858 * If the cmpxchg does not succeed then we may
2859 * drop the list_lock without any processing.
2861 * Otherwise the list_lock will synchronize with
2862 * other processors updating the list of slabs.
2864 spin_lock_irqsave(&n->list_lock, flags);
2869 } while (!cmpxchg_double_slab(s, page,
2877 * If we just froze the page then put it onto the
2878 * per cpu partial list.
2880 if (new.frozen && !was_frozen) {
2881 put_cpu_partial(s, page, 1);
2882 stat(s, CPU_PARTIAL_FREE);
2885 * The list lock was not taken therefore no list
2886 * activity can be necessary.
2889 stat(s, FREE_FROZEN);
2893 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2897 * Objects left in the slab. If it was not on the partial list before
2900 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2901 if (kmem_cache_debug(s))
2902 remove_full(s, n, page);
2903 add_partial(n, page, DEACTIVATE_TO_TAIL);
2904 stat(s, FREE_ADD_PARTIAL);
2906 spin_unlock_irqrestore(&n->list_lock, flags);
2912 * Slab on the partial list.
2914 remove_partial(n, page);
2915 stat(s, FREE_REMOVE_PARTIAL);
2917 /* Slab must be on the full list */
2918 remove_full(s, n, page);
2921 spin_unlock_irqrestore(&n->list_lock, flags);
2923 discard_slab(s, page);
2927 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2928 * can perform fastpath freeing without additional function calls.
2930 * The fastpath is only possible if we are freeing to the current cpu slab
2931 * of this processor. This typically the case if we have just allocated
2934 * If fastpath is not possible then fall back to __slab_free where we deal
2935 * with all sorts of special processing.
2937 * Bulk free of a freelist with several objects (all pointing to the
2938 * same page) possible by specifying head and tail ptr, plus objects
2939 * count (cnt). Bulk free indicated by tail pointer being set.
2941 static __always_inline void do_slab_free(struct kmem_cache *s,
2942 struct page *page, void *head, void *tail,
2943 int cnt, unsigned long addr)
2945 void *tail_obj = tail ? : head;
2946 struct kmem_cache_cpu *c;
2950 * Determine the currently cpus per cpu slab.
2951 * The cpu may change afterward. However that does not matter since
2952 * data is retrieved via this pointer. If we are on the same cpu
2953 * during the cmpxchg then the free will succeed.
2956 tid = this_cpu_read(s->cpu_slab->tid);
2957 c = raw_cpu_ptr(s->cpu_slab);
2958 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2959 unlikely(tid != READ_ONCE(c->tid)));
2961 /* Same with comment on barrier() in slab_alloc_node() */
2964 if (likely(page == c->page)) {
2965 void **freelist = READ_ONCE(c->freelist);
2967 set_freepointer(s, tail_obj, freelist);
2969 if (unlikely(!this_cpu_cmpxchg_double(
2970 s->cpu_slab->freelist, s->cpu_slab->tid,
2972 head, next_tid(tid)))) {
2974 note_cmpxchg_failure("slab_free", s, tid);
2977 stat(s, FREE_FASTPATH);
2979 __slab_free(s, page, head, tail_obj, cnt, addr);
2983 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
2984 void *head, void *tail, int cnt,
2988 * With KASAN enabled slab_free_freelist_hook modifies the freelist
2989 * to remove objects, whose reuse must be delayed.
2991 if (slab_free_freelist_hook(s, &head, &tail))
2992 do_slab_free(s, page, head, tail, cnt, addr);
2996 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
2998 do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr);
3002 void kmem_cache_free(struct kmem_cache *s, void *x)
3004 s = cache_from_obj(s, x);
3007 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
3008 trace_kmem_cache_free(_RET_IP_, x);
3010 EXPORT_SYMBOL(kmem_cache_free);
3012 struct detached_freelist {
3017 struct kmem_cache *s;
3021 * This function progressively scans the array with free objects (with
3022 * a limited look ahead) and extract objects belonging to the same
3023 * page. It builds a detached freelist directly within the given
3024 * page/objects. This can happen without any need for
3025 * synchronization, because the objects are owned by running process.
3026 * The freelist is build up as a single linked list in the objects.
3027 * The idea is, that this detached freelist can then be bulk
3028 * transferred to the real freelist(s), but only requiring a single
3029 * synchronization primitive. Look ahead in the array is limited due
3030 * to performance reasons.
3033 int build_detached_freelist(struct kmem_cache *s, size_t size,
3034 void **p, struct detached_freelist *df)
3036 size_t first_skipped_index = 0;
3041 /* Always re-init detached_freelist */
3046 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3047 } while (!object && size);
3052 page = virt_to_head_page(object);
3054 /* Handle kalloc'ed objects */
3055 if (unlikely(!PageSlab(page))) {
3056 BUG_ON(!PageCompound(page));
3058 __free_pages(page, compound_order(page));
3059 p[size] = NULL; /* mark object processed */
3062 /* Derive kmem_cache from object */
3063 df->s = page->slab_cache;
3065 df->s = cache_from_obj(s, object); /* Support for memcg */
3068 /* Start new detached freelist */
3070 set_freepointer(df->s, object, NULL);
3072 df->freelist = object;
3073 p[size] = NULL; /* mark object processed */
3079 continue; /* Skip processed objects */
3081 /* df->page is always set at this point */
3082 if (df->page == virt_to_head_page(object)) {
3083 /* Opportunity build freelist */
3084 set_freepointer(df->s, object, df->freelist);
3085 df->freelist = object;
3087 p[size] = NULL; /* mark object processed */
3092 /* Limit look ahead search */
3096 if (!first_skipped_index)
3097 first_skipped_index = size + 1;
3100 return first_skipped_index;
3103 /* Note that interrupts must be enabled when calling this function. */
3104 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3110 struct detached_freelist df;
3112 size = build_detached_freelist(s, size, p, &df);
3116 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
3117 } while (likely(size));
3119 EXPORT_SYMBOL(kmem_cache_free_bulk);
3121 /* Note that interrupts must be enabled when calling this function. */
3122 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3125 struct kmem_cache_cpu *c;
3128 /* memcg and kmem_cache debug support */
3129 s = slab_pre_alloc_hook(s, flags);
3133 * Drain objects in the per cpu slab, while disabling local
3134 * IRQs, which protects against PREEMPT and interrupts
3135 * handlers invoking normal fastpath.
3137 local_irq_disable();
3138 c = this_cpu_ptr(s->cpu_slab);
3140 for (i = 0; i < size; i++) {
3141 void *object = c->freelist;
3143 if (unlikely(!object)) {
3145 * We may have removed an object from c->freelist using
3146 * the fastpath in the previous iteration; in that case,
3147 * c->tid has not been bumped yet.
3148 * Since ___slab_alloc() may reenable interrupts while
3149 * allocating memory, we should bump c->tid now.
3151 c->tid = next_tid(c->tid);
3154 * Invoking slow path likely have side-effect
3155 * of re-populating per CPU c->freelist
3157 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3159 if (unlikely(!p[i]))
3162 c = this_cpu_ptr(s->cpu_slab);
3163 continue; /* goto for-loop */
3165 c->freelist = get_freepointer(s, object);
3168 c->tid = next_tid(c->tid);
3171 /* Clear memory outside IRQ disabled fastpath loop */
3172 if (unlikely(flags & __GFP_ZERO)) {
3175 for (j = 0; j < i; j++)
3176 memset(p[j], 0, s->object_size);
3179 /* memcg and kmem_cache debug support */
3180 slab_post_alloc_hook(s, flags, size, p);
3184 slab_post_alloc_hook(s, flags, i, p);
3185 __kmem_cache_free_bulk(s, i, p);
3188 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3192 * Object placement in a slab is made very easy because we always start at
3193 * offset 0. If we tune the size of the object to the alignment then we can
3194 * get the required alignment by putting one properly sized object after
3197 * Notice that the allocation order determines the sizes of the per cpu
3198 * caches. Each processor has always one slab available for allocations.
3199 * Increasing the allocation order reduces the number of times that slabs
3200 * must be moved on and off the partial lists and is therefore a factor in
3205 * Mininum / Maximum order of slab pages. This influences locking overhead
3206 * and slab fragmentation. A higher order reduces the number of partial slabs
3207 * and increases the number of allocations possible without having to
3208 * take the list_lock.
3210 static unsigned int slub_min_order;
3211 static unsigned int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3212 static unsigned int slub_min_objects;
3215 * Calculate the order of allocation given an slab object size.
3217 * The order of allocation has significant impact on performance and other
3218 * system components. Generally order 0 allocations should be preferred since
3219 * order 0 does not cause fragmentation in the page allocator. Larger objects
3220 * be problematic to put into order 0 slabs because there may be too much
3221 * unused space left. We go to a higher order if more than 1/16th of the slab
3224 * In order to reach satisfactory performance we must ensure that a minimum
3225 * number of objects is in one slab. Otherwise we may generate too much
3226 * activity on the partial lists which requires taking the list_lock. This is
3227 * less a concern for large slabs though which are rarely used.
3229 * slub_max_order specifies the order where we begin to stop considering the
3230 * number of objects in a slab as critical. If we reach slub_max_order then
3231 * we try to keep the page order as low as possible. So we accept more waste
3232 * of space in favor of a small page order.
3234 * Higher order allocations also allow the placement of more objects in a
3235 * slab and thereby reduce object handling overhead. If the user has
3236 * requested a higher mininum order then we start with that one instead of
3237 * the smallest order which will fit the object.
3239 static inline unsigned int slab_order(unsigned int size,
3240 unsigned int min_objects, unsigned int max_order,
3241 unsigned int fract_leftover)
3243 unsigned int min_order = slub_min_order;
3246 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
3247 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3249 for (order = max(min_order, (unsigned int)get_order(min_objects * size));
3250 order <= max_order; order++) {
3252 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
3255 rem = slab_size % size;
3257 if (rem <= slab_size / fract_leftover)
3264 static inline int calculate_order(unsigned int size)
3267 unsigned int min_objects;
3268 unsigned int max_objects;
3271 * Attempt to find best configuration for a slab. This
3272 * works by first attempting to generate a layout with
3273 * the best configuration and backing off gradually.
3275 * First we increase the acceptable waste in a slab. Then
3276 * we reduce the minimum objects required in a slab.
3278 min_objects = slub_min_objects;
3280 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3281 max_objects = order_objects(slub_max_order, size);
3282 min_objects = min(min_objects, max_objects);
3284 while (min_objects > 1) {
3285 unsigned int fraction;
3288 while (fraction >= 4) {
3289 order = slab_order(size, min_objects,
3290 slub_max_order, fraction);
3291 if (order <= slub_max_order)
3299 * We were unable to place multiple objects in a slab. Now
3300 * lets see if we can place a single object there.
3302 order = slab_order(size, 1, slub_max_order, 1);
3303 if (order <= slub_max_order)
3307 * Doh this slab cannot be placed using slub_max_order.
3309 order = slab_order(size, 1, MAX_ORDER, 1);
3310 if (order < MAX_ORDER)
3316 init_kmem_cache_node(struct kmem_cache_node *n)
3319 spin_lock_init(&n->list_lock);
3320 INIT_LIST_HEAD(&n->partial);
3321 #ifdef CONFIG_SLUB_DEBUG
3322 atomic_long_set(&n->nr_slabs, 0);
3323 atomic_long_set(&n->total_objects, 0);
3324 INIT_LIST_HEAD(&n->full);
3328 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3330 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3331 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3334 * Must align to double word boundary for the double cmpxchg
3335 * instructions to work; see __pcpu_double_call_return_bool().
3337 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3338 2 * sizeof(void *));
3343 init_kmem_cache_cpus(s);
3348 static struct kmem_cache *kmem_cache_node;
3351 * No kmalloc_node yet so do it by hand. We know that this is the first
3352 * slab on the node for this slabcache. There are no concurrent accesses
3355 * Note that this function only works on the kmem_cache_node
3356 * when allocating for the kmem_cache_node. This is used for bootstrapping
3357 * memory on a fresh node that has no slab structures yet.
3359 static void early_kmem_cache_node_alloc(int node)
3362 struct kmem_cache_node *n;
3364 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3366 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3369 if (page_to_nid(page) != node) {
3370 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3371 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3376 page->freelist = get_freepointer(kmem_cache_node, n);
3379 kmem_cache_node->node[node] = n;
3380 #ifdef CONFIG_SLUB_DEBUG
3381 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3382 init_tracking(kmem_cache_node, n);
3384 kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node),
3386 init_kmem_cache_node(n);
3387 inc_slabs_node(kmem_cache_node, node, page->objects);
3390 * No locks need to be taken here as it has just been
3391 * initialized and there is no concurrent access.
3393 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3396 static void free_kmem_cache_nodes(struct kmem_cache *s)
3399 struct kmem_cache_node *n;
3401 for_each_kmem_cache_node(s, node, n) {
3402 s->node[node] = NULL;
3403 kmem_cache_free(kmem_cache_node, n);
3407 void __kmem_cache_release(struct kmem_cache *s)
3409 cache_random_seq_destroy(s);
3410 free_percpu(s->cpu_slab);
3411 free_kmem_cache_nodes(s);
3414 static int init_kmem_cache_nodes(struct kmem_cache *s)
3418 for_each_node_state(node, N_NORMAL_MEMORY) {
3419 struct kmem_cache_node *n;
3421 if (slab_state == DOWN) {
3422 early_kmem_cache_node_alloc(node);
3425 n = kmem_cache_alloc_node(kmem_cache_node,
3429 free_kmem_cache_nodes(s);
3433 init_kmem_cache_node(n);
3439 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3441 if (min < MIN_PARTIAL)
3443 else if (min > MAX_PARTIAL)
3445 s->min_partial = min;
3448 static void set_cpu_partial(struct kmem_cache *s)
3450 #ifdef CONFIG_SLUB_CPU_PARTIAL
3452 * cpu_partial determined the maximum number of objects kept in the
3453 * per cpu partial lists of a processor.
3455 * Per cpu partial lists mainly contain slabs that just have one
3456 * object freed. If they are used for allocation then they can be
3457 * filled up again with minimal effort. The slab will never hit the
3458 * per node partial lists and therefore no locking will be required.
3460 * This setting also determines
3462 * A) The number of objects from per cpu partial slabs dumped to the
3463 * per node list when we reach the limit.
3464 * B) The number of objects in cpu partial slabs to extract from the
3465 * per node list when we run out of per cpu objects. We only fetch
3466 * 50% to keep some capacity around for frees.
3468 if (!kmem_cache_has_cpu_partial(s))
3470 else if (s->size >= PAGE_SIZE)
3472 else if (s->size >= 1024)
3474 else if (s->size >= 256)
3475 s->cpu_partial = 13;
3477 s->cpu_partial = 30;
3482 * calculate_sizes() determines the order and the distribution of data within
3485 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3487 slab_flags_t flags = s->flags;
3488 unsigned int size = s->object_size;
3492 * Round up object size to the next word boundary. We can only
3493 * place the free pointer at word boundaries and this determines
3494 * the possible location of the free pointer.
3496 size = ALIGN(size, sizeof(void *));
3498 #ifdef CONFIG_SLUB_DEBUG
3500 * Determine if we can poison the object itself. If the user of
3501 * the slab may touch the object after free or before allocation
3502 * then we should never poison the object itself.
3504 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
3506 s->flags |= __OBJECT_POISON;
3508 s->flags &= ~__OBJECT_POISON;
3512 * If we are Redzoning then check if there is some space between the
3513 * end of the object and the free pointer. If not then add an
3514 * additional word to have some bytes to store Redzone information.
3516 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3517 size += sizeof(void *);
3521 * With that we have determined the number of bytes in actual use
3522 * by the object. This is the potential offset to the free pointer.
3526 if (((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
3529 * Relocate free pointer after the object if it is not
3530 * permitted to overwrite the first word of the object on
3533 * This is the case if we do RCU, have a constructor or
3534 * destructor or are poisoning the objects.
3537 size += sizeof(void *);
3540 #ifdef CONFIG_SLUB_DEBUG
3541 if (flags & SLAB_STORE_USER)
3543 * Need to store information about allocs and frees after
3546 size += 2 * sizeof(struct track);
3549 kasan_cache_create(s, &size, &s->flags);
3550 #ifdef CONFIG_SLUB_DEBUG
3551 if (flags & SLAB_RED_ZONE) {
3553 * Add some empty padding so that we can catch
3554 * overwrites from earlier objects rather than let
3555 * tracking information or the free pointer be
3556 * corrupted if a user writes before the start
3559 size += sizeof(void *);
3561 s->red_left_pad = sizeof(void *);
3562 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3563 size += s->red_left_pad;
3568 * SLUB stores one object immediately after another beginning from
3569 * offset 0. In order to align the objects we have to simply size
3570 * each object to conform to the alignment.
3572 size = ALIGN(size, s->align);
3574 if (forced_order >= 0)
3575 order = forced_order;
3577 order = calculate_order(size);
3584 s->allocflags |= __GFP_COMP;
3586 if (s->flags & SLAB_CACHE_DMA)
3587 s->allocflags |= GFP_DMA;
3589 if (s->flags & SLAB_CACHE_DMA32)
3590 s->allocflags |= GFP_DMA32;
3592 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3593 s->allocflags |= __GFP_RECLAIMABLE;
3596 * Determine the number of objects per slab
3598 s->oo = oo_make(order, size);
3599 s->min = oo_make(get_order(size), size);
3600 if (oo_objects(s->oo) > oo_objects(s->max))
3603 return !!oo_objects(s->oo);
3606 static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
3608 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3609 #ifdef CONFIG_SLAB_FREELIST_HARDENED
3610 s->random = get_random_long();
3613 if (!calculate_sizes(s, -1))
3615 if (disable_higher_order_debug) {
3617 * Disable debugging flags that store metadata if the min slab
3620 if (get_order(s->size) > get_order(s->object_size)) {
3621 s->flags &= ~DEBUG_METADATA_FLAGS;
3623 if (!calculate_sizes(s, -1))
3628 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3629 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3630 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
3631 /* Enable fast mode */
3632 s->flags |= __CMPXCHG_DOUBLE;
3636 * The larger the object size is, the more pages we want on the partial
3637 * list to avoid pounding the page allocator excessively.
3639 set_min_partial(s, ilog2(s->size) / 2);
3644 s->remote_node_defrag_ratio = 1000;
3647 /* Initialize the pre-computed randomized freelist if slab is up */
3648 if (slab_state >= UP) {
3649 if (init_cache_random_seq(s))
3653 if (!init_kmem_cache_nodes(s))
3656 if (alloc_kmem_cache_cpus(s))
3659 free_kmem_cache_nodes(s);
3661 if (flags & SLAB_PANIC)
3662 panic("Cannot create slab %s size=%u realsize=%u order=%u offset=%u flags=%lx\n",
3663 s->name, s->size, s->size,
3664 oo_order(s->oo), s->offset, (unsigned long)flags);
3668 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3671 #ifdef CONFIG_SLUB_DEBUG
3672 void *addr = page_address(page);
3674 unsigned long *map = kcalloc(BITS_TO_LONGS(page->objects),
3679 slab_err(s, page, text, s->name);
3682 get_map(s, page, map);
3683 for_each_object(p, s, addr, page->objects) {
3685 if (!test_bit(slab_index(p, s, addr), map)) {
3686 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3687 print_tracking(s, p);
3696 * Attempt to free all partial slabs on a node.
3697 * This is called from __kmem_cache_shutdown(). We must take list_lock
3698 * because sysfs file might still access partial list after the shutdowning.
3700 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3703 struct page *page, *h;
3705 BUG_ON(irqs_disabled());
3706 spin_lock_irq(&n->list_lock);
3707 list_for_each_entry_safe(page, h, &n->partial, lru) {
3709 remove_partial(n, page);
3710 list_add(&page->lru, &discard);
3712 list_slab_objects(s, page,
3713 "Objects remaining in %s on __kmem_cache_shutdown()");
3716 spin_unlock_irq(&n->list_lock);
3718 list_for_each_entry_safe(page, h, &discard, lru)
3719 discard_slab(s, page);
3722 bool __kmem_cache_empty(struct kmem_cache *s)
3725 struct kmem_cache_node *n;
3727 for_each_kmem_cache_node(s, node, n)
3728 if (n->nr_partial || slabs_node(s, node))
3734 * Release all resources used by a slab cache.
3736 int __kmem_cache_shutdown(struct kmem_cache *s)
3739 struct kmem_cache_node *n;
3742 /* Attempt to free all objects */
3743 for_each_kmem_cache_node(s, node, n) {
3745 if (n->nr_partial || slabs_node(s, node))
3748 sysfs_slab_remove(s);
3752 /********************************************************************
3754 *******************************************************************/
3756 static int __init setup_slub_min_order(char *str)
3758 get_option(&str, (int *)&slub_min_order);
3763 __setup("slub_min_order=", setup_slub_min_order);
3765 static int __init setup_slub_max_order(char *str)
3767 get_option(&str, (int *)&slub_max_order);
3768 slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1);
3773 __setup("slub_max_order=", setup_slub_max_order);
3775 static int __init setup_slub_min_objects(char *str)
3777 get_option(&str, (int *)&slub_min_objects);
3782 __setup("slub_min_objects=", setup_slub_min_objects);
3784 void *__kmalloc(size_t size, gfp_t flags)
3786 struct kmem_cache *s;
3789 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3790 return kmalloc_large(size, flags);
3792 s = kmalloc_slab(size, flags);
3794 if (unlikely(ZERO_OR_NULL_PTR(s)))
3797 ret = slab_alloc(s, flags, _RET_IP_);
3799 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3801 kasan_kmalloc(s, ret, size, flags);
3805 EXPORT_SYMBOL(__kmalloc);
3808 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3813 flags |= __GFP_COMP;
3814 page = alloc_pages_node(node, flags, get_order(size));
3816 ptr = page_address(page);
3818 kmalloc_large_node_hook(ptr, size, flags);
3822 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3824 struct kmem_cache *s;
3827 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3828 ret = kmalloc_large_node(size, flags, node);
3830 trace_kmalloc_node(_RET_IP_, ret,
3831 size, PAGE_SIZE << get_order(size),
3837 s = kmalloc_slab(size, flags);
3839 if (unlikely(ZERO_OR_NULL_PTR(s)))
3842 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3844 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3846 kasan_kmalloc(s, ret, size, flags);
3850 EXPORT_SYMBOL(__kmalloc_node);
3853 #ifdef CONFIG_HARDENED_USERCOPY
3855 * Rejects incorrectly sized objects and objects that are to be copied
3856 * to/from userspace but do not fall entirely within the containing slab
3857 * cache's usercopy region.
3859 * Returns NULL if check passes, otherwise const char * to name of cache
3860 * to indicate an error.
3862 void __check_heap_object(const void *ptr, unsigned long n, struct page *page,
3865 struct kmem_cache *s;
3866 unsigned int offset;
3869 /* Find object and usable object size. */
3870 s = page->slab_cache;
3872 /* Reject impossible pointers. */
3873 if (ptr < page_address(page))
3874 usercopy_abort("SLUB object not in SLUB page?!", NULL,
3877 /* Find offset within object. */
3878 offset = (ptr - page_address(page)) % s->size;
3880 /* Adjust for redzone and reject if within the redzone. */
3881 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE) {
3882 if (offset < s->red_left_pad)
3883 usercopy_abort("SLUB object in left red zone",
3884 s->name, to_user, offset, n);
3885 offset -= s->red_left_pad;
3888 /* Allow address range falling entirely within usercopy region. */
3889 if (offset >= s->useroffset &&
3890 offset - s->useroffset <= s->usersize &&
3891 n <= s->useroffset - offset + s->usersize)
3895 * If the copy is still within the allocated object, produce
3896 * a warning instead of rejecting the copy. This is intended
3897 * to be a temporary method to find any missing usercopy
3900 object_size = slab_ksize(s);
3901 if (usercopy_fallback &&
3902 offset <= object_size && n <= object_size - offset) {
3903 usercopy_warn("SLUB object", s->name, to_user, offset, n);
3907 usercopy_abort("SLUB object", s->name, to_user, offset, n);
3909 #endif /* CONFIG_HARDENED_USERCOPY */
3911 static size_t __ksize(const void *object)
3915 if (unlikely(object == ZERO_SIZE_PTR))
3918 page = virt_to_head_page(object);
3920 if (unlikely(!PageSlab(page))) {
3921 WARN_ON(!PageCompound(page));
3922 return PAGE_SIZE << compound_order(page);
3925 return slab_ksize(page->slab_cache);
3928 size_t ksize(const void *object)
3930 size_t size = __ksize(object);
3931 /* We assume that ksize callers could use whole allocated area,
3932 * so we need to unpoison this area.
3934 kasan_unpoison_shadow(object, size);
3937 EXPORT_SYMBOL(ksize);
3939 void kfree(const void *x)
3942 void *object = (void *)x;
3944 trace_kfree(_RET_IP_, x);
3946 if (unlikely(ZERO_OR_NULL_PTR(x)))
3949 page = virt_to_head_page(x);
3950 if (unlikely(!PageSlab(page))) {
3951 BUG_ON(!PageCompound(page));
3953 __free_pages(page, compound_order(page));
3956 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
3958 EXPORT_SYMBOL(kfree);
3960 #define SHRINK_PROMOTE_MAX 32
3963 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3964 * up most to the head of the partial lists. New allocations will then
3965 * fill those up and thus they can be removed from the partial lists.
3967 * The slabs with the least items are placed last. This results in them
3968 * being allocated from last increasing the chance that the last objects
3969 * are freed in them.
3971 int __kmem_cache_shrink(struct kmem_cache *s)
3975 struct kmem_cache_node *n;
3978 struct list_head discard;
3979 struct list_head promote[SHRINK_PROMOTE_MAX];
3980 unsigned long flags;
3984 for_each_kmem_cache_node(s, node, n) {
3985 INIT_LIST_HEAD(&discard);
3986 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3987 INIT_LIST_HEAD(promote + i);
3989 spin_lock_irqsave(&n->list_lock, flags);
3992 * Build lists of slabs to discard or promote.
3994 * Note that concurrent frees may occur while we hold the
3995 * list_lock. page->inuse here is the upper limit.
3997 list_for_each_entry_safe(page, t, &n->partial, lru) {
3998 int free = page->objects - page->inuse;
4000 /* Do not reread page->inuse */
4003 /* We do not keep full slabs on the list */
4006 if (free == page->objects) {
4007 list_move(&page->lru, &discard);
4009 } else if (free <= SHRINK_PROMOTE_MAX)
4010 list_move(&page->lru, promote + free - 1);
4014 * Promote the slabs filled up most to the head of the
4017 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
4018 list_splice(promote + i, &n->partial);
4020 spin_unlock_irqrestore(&n->list_lock, flags);
4022 /* Release empty slabs */
4023 list_for_each_entry_safe(page, t, &discard, lru)
4024 discard_slab(s, page);
4026 if (slabs_node(s, node))
4034 static void kmemcg_cache_deact_after_rcu(struct kmem_cache *s)
4037 * Called with all the locks held after a sched RCU grace period.
4038 * Even if @s becomes empty after shrinking, we can't know that @s
4039 * doesn't have allocations already in-flight and thus can't
4040 * destroy @s until the associated memcg is released.
4042 * However, let's remove the sysfs files for empty caches here.
4043 * Each cache has a lot of interface files which aren't
4044 * particularly useful for empty draining caches; otherwise, we can
4045 * easily end up with millions of unnecessary sysfs files on
4046 * systems which have a lot of memory and transient cgroups.
4048 if (!__kmem_cache_shrink(s))
4049 sysfs_slab_remove(s);
4052 void __kmemcg_cache_deactivate(struct kmem_cache *s)
4055 * Disable empty slabs caching. Used to avoid pinning offline
4056 * memory cgroups by kmem pages that can be freed.
4058 slub_set_cpu_partial(s, 0);
4062 * s->cpu_partial is checked locklessly (see put_cpu_partial), so
4063 * we have to make sure the change is visible before shrinking.
4065 slab_deactivate_memcg_cache_rcu_sched(s, kmemcg_cache_deact_after_rcu);
4069 static int slab_mem_going_offline_callback(void *arg)
4071 struct kmem_cache *s;
4073 mutex_lock(&slab_mutex);
4074 list_for_each_entry(s, &slab_caches, list)
4075 __kmem_cache_shrink(s);
4076 mutex_unlock(&slab_mutex);
4081 static void slab_mem_offline_callback(void *arg)
4083 struct kmem_cache_node *n;
4084 struct kmem_cache *s;
4085 struct memory_notify *marg = arg;
4088 offline_node = marg->status_change_nid_normal;
4091 * If the node still has available memory. we need kmem_cache_node
4094 if (offline_node < 0)
4097 mutex_lock(&slab_mutex);
4098 list_for_each_entry(s, &slab_caches, list) {
4099 n = get_node(s, offline_node);
4102 * if n->nr_slabs > 0, slabs still exist on the node
4103 * that is going down. We were unable to free them,
4104 * and offline_pages() function shouldn't call this
4105 * callback. So, we must fail.
4107 BUG_ON(slabs_node(s, offline_node));
4109 s->node[offline_node] = NULL;
4110 kmem_cache_free(kmem_cache_node, n);
4113 mutex_unlock(&slab_mutex);
4116 static int slab_mem_going_online_callback(void *arg)
4118 struct kmem_cache_node *n;
4119 struct kmem_cache *s;
4120 struct memory_notify *marg = arg;
4121 int nid = marg->status_change_nid_normal;
4125 * If the node's memory is already available, then kmem_cache_node is
4126 * already created. Nothing to do.
4132 * We are bringing a node online. No memory is available yet. We must
4133 * allocate a kmem_cache_node structure in order to bring the node
4136 mutex_lock(&slab_mutex);
4137 list_for_each_entry(s, &slab_caches, list) {
4139 * XXX: kmem_cache_alloc_node will fallback to other nodes
4140 * since memory is not yet available from the node that
4143 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4148 init_kmem_cache_node(n);
4152 mutex_unlock(&slab_mutex);
4156 static int slab_memory_callback(struct notifier_block *self,
4157 unsigned long action, void *arg)
4162 case MEM_GOING_ONLINE:
4163 ret = slab_mem_going_online_callback(arg);
4165 case MEM_GOING_OFFLINE:
4166 ret = slab_mem_going_offline_callback(arg);
4169 case MEM_CANCEL_ONLINE:
4170 slab_mem_offline_callback(arg);
4173 case MEM_CANCEL_OFFLINE:
4177 ret = notifier_from_errno(ret);
4183 static struct notifier_block slab_memory_callback_nb = {
4184 .notifier_call = slab_memory_callback,
4185 .priority = SLAB_CALLBACK_PRI,
4188 /********************************************************************
4189 * Basic setup of slabs
4190 *******************************************************************/
4193 * Used for early kmem_cache structures that were allocated using
4194 * the page allocator. Allocate them properly then fix up the pointers
4195 * that may be pointing to the wrong kmem_cache structure.
4198 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4201 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4202 struct kmem_cache_node *n;
4204 memcpy(s, static_cache, kmem_cache->object_size);
4207 * This runs very early, and only the boot processor is supposed to be
4208 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4211 __flush_cpu_slab(s, smp_processor_id());
4212 for_each_kmem_cache_node(s, node, n) {
4215 list_for_each_entry(p, &n->partial, lru)
4218 #ifdef CONFIG_SLUB_DEBUG
4219 list_for_each_entry(p, &n->full, lru)
4223 slab_init_memcg_params(s);
4224 list_add(&s->list, &slab_caches);
4225 memcg_link_cache(s);
4229 void __init kmem_cache_init(void)
4231 static __initdata struct kmem_cache boot_kmem_cache,
4232 boot_kmem_cache_node;
4234 if (debug_guardpage_minorder())
4237 kmem_cache_node = &boot_kmem_cache_node;
4238 kmem_cache = &boot_kmem_cache;
4240 create_boot_cache(kmem_cache_node, "kmem_cache_node",
4241 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
4243 register_hotmemory_notifier(&slab_memory_callback_nb);
4245 /* Able to allocate the per node structures */
4246 slab_state = PARTIAL;
4248 create_boot_cache(kmem_cache, "kmem_cache",
4249 offsetof(struct kmem_cache, node) +
4250 nr_node_ids * sizeof(struct kmem_cache_node *),
4251 SLAB_HWCACHE_ALIGN, 0, 0);
4253 kmem_cache = bootstrap(&boot_kmem_cache);
4254 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
4256 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4257 setup_kmalloc_cache_index_table();
4258 create_kmalloc_caches(0);
4260 /* Setup random freelists for each cache */
4261 init_freelist_randomization();
4263 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
4266 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%d\n",
4268 slub_min_order, slub_max_order, slub_min_objects,
4269 nr_cpu_ids, nr_node_ids);
4272 void __init kmem_cache_init_late(void)
4277 __kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
4278 slab_flags_t flags, void (*ctor)(void *))
4280 struct kmem_cache *s, *c;
4282 s = find_mergeable(size, align, flags, name, ctor);
4287 * Adjust the object sizes so that we clear
4288 * the complete object on kzalloc.
4290 s->object_size = max(s->object_size, size);
4291 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
4293 for_each_memcg_cache(c, s) {
4294 c->object_size = s->object_size;
4295 c->inuse = max(c->inuse, ALIGN(size, sizeof(void *)));
4298 if (sysfs_slab_alias(s, name)) {
4307 int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
4311 err = kmem_cache_open(s, flags);
4315 /* Mutex is not taken during early boot */
4316 if (slab_state <= UP)
4319 memcg_propagate_slab_attrs(s);
4320 err = sysfs_slab_add(s);
4322 __kmem_cache_release(s);
4327 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4329 struct kmem_cache *s;
4332 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4333 return kmalloc_large(size, gfpflags);
4335 s = kmalloc_slab(size, gfpflags);
4337 if (unlikely(ZERO_OR_NULL_PTR(s)))
4340 ret = slab_alloc(s, gfpflags, caller);
4342 /* Honor the call site pointer we received. */
4343 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4349 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4350 int node, unsigned long caller)
4352 struct kmem_cache *s;
4355 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4356 ret = kmalloc_large_node(size, gfpflags, node);
4358 trace_kmalloc_node(caller, ret,
4359 size, PAGE_SIZE << get_order(size),
4365 s = kmalloc_slab(size, gfpflags);
4367 if (unlikely(ZERO_OR_NULL_PTR(s)))
4370 ret = slab_alloc_node(s, gfpflags, node, caller);
4372 /* Honor the call site pointer we received. */
4373 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4380 static int count_inuse(struct page *page)
4385 static int count_total(struct page *page)
4387 return page->objects;
4391 #ifdef CONFIG_SLUB_DEBUG
4392 static int validate_slab(struct kmem_cache *s, struct page *page,
4396 void *addr = page_address(page);
4398 if (!check_slab(s, page) ||
4399 !on_freelist(s, page, NULL))
4402 /* Now we know that a valid freelist exists */
4403 bitmap_zero(map, page->objects);
4405 get_map(s, page, map);
4406 for_each_object(p, s, addr, page->objects) {
4407 if (test_bit(slab_index(p, s, addr), map))
4408 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4412 for_each_object(p, s, addr, page->objects)
4413 if (!test_bit(slab_index(p, s, addr), map))
4414 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4419 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4423 validate_slab(s, page, map);
4427 static int validate_slab_node(struct kmem_cache *s,
4428 struct kmem_cache_node *n, unsigned long *map)
4430 unsigned long count = 0;
4432 unsigned long flags;
4434 spin_lock_irqsave(&n->list_lock, flags);
4436 list_for_each_entry(page, &n->partial, lru) {
4437 validate_slab_slab(s, page, map);
4440 if (count != n->nr_partial)
4441 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4442 s->name, count, n->nr_partial);
4444 if (!(s->flags & SLAB_STORE_USER))
4447 list_for_each_entry(page, &n->full, lru) {
4448 validate_slab_slab(s, page, map);
4451 if (count != atomic_long_read(&n->nr_slabs))
4452 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4453 s->name, count, atomic_long_read(&n->nr_slabs));
4456 spin_unlock_irqrestore(&n->list_lock, flags);
4460 static long validate_slab_cache(struct kmem_cache *s)
4463 unsigned long count = 0;
4464 unsigned long *map = kmalloc_array(BITS_TO_LONGS(oo_objects(s->max)),
4465 sizeof(unsigned long),
4467 struct kmem_cache_node *n;
4473 for_each_kmem_cache_node(s, node, n)
4474 count += validate_slab_node(s, n, map);
4479 * Generate lists of code addresses where slabcache objects are allocated
4484 unsigned long count;
4491 DECLARE_BITMAP(cpus, NR_CPUS);
4497 unsigned long count;
4498 struct location *loc;
4501 static void free_loc_track(struct loc_track *t)
4504 free_pages((unsigned long)t->loc,
4505 get_order(sizeof(struct location) * t->max));
4508 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4513 order = get_order(sizeof(struct location) * max);
4515 l = (void *)__get_free_pages(flags, order);
4520 memcpy(l, t->loc, sizeof(struct location) * t->count);
4528 static int add_location(struct loc_track *t, struct kmem_cache *s,
4529 const struct track *track)
4531 long start, end, pos;
4533 unsigned long caddr;
4534 unsigned long age = jiffies - track->when;
4540 pos = start + (end - start + 1) / 2;
4543 * There is nothing at "end". If we end up there
4544 * we need to add something to before end.
4549 caddr = t->loc[pos].addr;
4550 if (track->addr == caddr) {
4556 if (age < l->min_time)
4558 if (age > l->max_time)
4561 if (track->pid < l->min_pid)
4562 l->min_pid = track->pid;
4563 if (track->pid > l->max_pid)
4564 l->max_pid = track->pid;
4566 cpumask_set_cpu(track->cpu,
4567 to_cpumask(l->cpus));
4569 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4573 if (track->addr < caddr)
4580 * Not found. Insert new tracking element.
4582 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4588 (t->count - pos) * sizeof(struct location));
4591 l->addr = track->addr;
4595 l->min_pid = track->pid;
4596 l->max_pid = track->pid;
4597 cpumask_clear(to_cpumask(l->cpus));
4598 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4599 nodes_clear(l->nodes);
4600 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4604 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4605 struct page *page, enum track_item alloc,
4608 void *addr = page_address(page);
4611 bitmap_zero(map, page->objects);
4612 get_map(s, page, map);
4614 for_each_object(p, s, addr, page->objects)
4615 if (!test_bit(slab_index(p, s, addr), map))
4616 add_location(t, s, get_track(s, p, alloc));
4619 static int list_locations(struct kmem_cache *s, char *buf,
4620 enum track_item alloc)
4624 struct loc_track t = { 0, 0, NULL };
4626 unsigned long *map = kmalloc_array(BITS_TO_LONGS(oo_objects(s->max)),
4627 sizeof(unsigned long),
4629 struct kmem_cache_node *n;
4631 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4634 return sprintf(buf, "Out of memory\n");
4636 /* Push back cpu slabs */
4639 for_each_kmem_cache_node(s, node, n) {
4640 unsigned long flags;
4643 if (!atomic_long_read(&n->nr_slabs))
4646 spin_lock_irqsave(&n->list_lock, flags);
4647 list_for_each_entry(page, &n->partial, lru)
4648 process_slab(&t, s, page, alloc, map);
4649 list_for_each_entry(page, &n->full, lru)
4650 process_slab(&t, s, page, alloc, map);
4651 spin_unlock_irqrestore(&n->list_lock, flags);
4654 for (i = 0; i < t.count; i++) {
4655 struct location *l = &t.loc[i];
4657 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4659 len += sprintf(buf + len, "%7ld ", l->count);
4662 len += sprintf(buf + len, "%pS", (void *)l->addr);
4664 len += sprintf(buf + len, "<not-available>");
4666 if (l->sum_time != l->min_time) {
4667 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4669 (long)div_u64(l->sum_time, l->count),
4672 len += sprintf(buf + len, " age=%ld",
4675 if (l->min_pid != l->max_pid)
4676 len += sprintf(buf + len, " pid=%ld-%ld",
4677 l->min_pid, l->max_pid);
4679 len += sprintf(buf + len, " pid=%ld",
4682 if (num_online_cpus() > 1 &&
4683 !cpumask_empty(to_cpumask(l->cpus)) &&
4684 len < PAGE_SIZE - 60)
4685 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4687 cpumask_pr_args(to_cpumask(l->cpus)));
4689 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4690 len < PAGE_SIZE - 60)
4691 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4693 nodemask_pr_args(&l->nodes));
4695 len += sprintf(buf + len, "\n");
4701 len += sprintf(buf, "No data\n");
4706 #ifdef SLUB_RESILIENCY_TEST
4707 static void __init resiliency_test(void)
4711 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4713 pr_err("SLUB resiliency testing\n");
4714 pr_err("-----------------------\n");
4715 pr_err("A. Corruption after allocation\n");
4717 p = kzalloc(16, GFP_KERNEL);
4719 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4722 validate_slab_cache(kmalloc_caches[4]);
4724 /* Hmmm... The next two are dangerous */
4725 p = kzalloc(32, GFP_KERNEL);
4726 p[32 + sizeof(void *)] = 0x34;
4727 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4729 pr_err("If allocated object is overwritten then not detectable\n\n");
4731 validate_slab_cache(kmalloc_caches[5]);
4732 p = kzalloc(64, GFP_KERNEL);
4733 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4735 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4737 pr_err("If allocated object is overwritten then not detectable\n\n");
4738 validate_slab_cache(kmalloc_caches[6]);
4740 pr_err("\nB. Corruption after free\n");
4741 p = kzalloc(128, GFP_KERNEL);
4744 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4745 validate_slab_cache(kmalloc_caches[7]);
4747 p = kzalloc(256, GFP_KERNEL);
4750 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4751 validate_slab_cache(kmalloc_caches[8]);
4753 p = kzalloc(512, GFP_KERNEL);
4756 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4757 validate_slab_cache(kmalloc_caches[9]);
4761 static void resiliency_test(void) {};
4766 enum slab_stat_type {
4767 SL_ALL, /* All slabs */
4768 SL_PARTIAL, /* Only partially allocated slabs */
4769 SL_CPU, /* Only slabs used for cpu caches */
4770 SL_OBJECTS, /* Determine allocated objects not slabs */
4771 SL_TOTAL /* Determine object capacity not slabs */
4774 #define SO_ALL (1 << SL_ALL)
4775 #define SO_PARTIAL (1 << SL_PARTIAL)
4776 #define SO_CPU (1 << SL_CPU)
4777 #define SO_OBJECTS (1 << SL_OBJECTS)
4778 #define SO_TOTAL (1 << SL_TOTAL)
4781 static bool memcg_sysfs_enabled = IS_ENABLED(CONFIG_SLUB_MEMCG_SYSFS_ON);
4783 static int __init setup_slub_memcg_sysfs(char *str)
4787 if (get_option(&str, &v) > 0)
4788 memcg_sysfs_enabled = v;
4793 __setup("slub_memcg_sysfs=", setup_slub_memcg_sysfs);
4796 static ssize_t show_slab_objects(struct kmem_cache *s,
4797 char *buf, unsigned long flags)
4799 unsigned long total = 0;
4802 unsigned long *nodes;
4804 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
4808 if (flags & SO_CPU) {
4811 for_each_possible_cpu(cpu) {
4812 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4817 page = READ_ONCE(c->page);
4821 node = page_to_nid(page);
4822 if (flags & SO_TOTAL)
4824 else if (flags & SO_OBJECTS)
4832 page = slub_percpu_partial_read_once(c);
4834 node = page_to_nid(page);
4835 if (flags & SO_TOTAL)
4837 else if (flags & SO_OBJECTS)
4848 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
4849 * already held which will conflict with an existing lock order:
4851 * mem_hotplug_lock->slab_mutex->kernfs_mutex
4853 * We don't really need mem_hotplug_lock (to hold off
4854 * slab_mem_going_offline_callback) here because slab's memory hot
4855 * unplug code doesn't destroy the kmem_cache->node[] data.
4858 #ifdef CONFIG_SLUB_DEBUG
4859 if (flags & SO_ALL) {
4860 struct kmem_cache_node *n;
4862 for_each_kmem_cache_node(s, node, n) {
4864 if (flags & SO_TOTAL)
4865 x = atomic_long_read(&n->total_objects);
4866 else if (flags & SO_OBJECTS)
4867 x = atomic_long_read(&n->total_objects) -
4868 count_partial(n, count_free);
4870 x = atomic_long_read(&n->nr_slabs);
4877 if (flags & SO_PARTIAL) {
4878 struct kmem_cache_node *n;
4880 for_each_kmem_cache_node(s, node, n) {
4881 if (flags & SO_TOTAL)
4882 x = count_partial(n, count_total);
4883 else if (flags & SO_OBJECTS)
4884 x = count_partial(n, count_inuse);
4891 x = sprintf(buf, "%lu", total);
4893 for (node = 0; node < nr_node_ids; node++)
4895 x += sprintf(buf + x, " N%d=%lu",
4899 return x + sprintf(buf + x, "\n");
4902 #ifdef CONFIG_SLUB_DEBUG
4903 static int any_slab_objects(struct kmem_cache *s)
4906 struct kmem_cache_node *n;
4908 for_each_kmem_cache_node(s, node, n)
4909 if (atomic_long_read(&n->total_objects))
4916 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4917 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4919 struct slab_attribute {
4920 struct attribute attr;
4921 ssize_t (*show)(struct kmem_cache *s, char *buf);
4922 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4925 #define SLAB_ATTR_RO(_name) \
4926 static struct slab_attribute _name##_attr = \
4927 __ATTR(_name, 0400, _name##_show, NULL)
4929 #define SLAB_ATTR(_name) \
4930 static struct slab_attribute _name##_attr = \
4931 __ATTR(_name, 0600, _name##_show, _name##_store)
4933 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4935 return sprintf(buf, "%u\n", s->size);
4937 SLAB_ATTR_RO(slab_size);
4939 static ssize_t align_show(struct kmem_cache *s, char *buf)
4941 return sprintf(buf, "%u\n", s->align);
4943 SLAB_ATTR_RO(align);
4945 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4947 return sprintf(buf, "%u\n", s->object_size);
4949 SLAB_ATTR_RO(object_size);
4951 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4953 return sprintf(buf, "%u\n", oo_objects(s->oo));
4955 SLAB_ATTR_RO(objs_per_slab);
4957 static ssize_t order_store(struct kmem_cache *s,
4958 const char *buf, size_t length)
4963 err = kstrtouint(buf, 10, &order);
4967 if (order > slub_max_order || order < slub_min_order)
4970 calculate_sizes(s, order);
4974 static ssize_t order_show(struct kmem_cache *s, char *buf)
4976 return sprintf(buf, "%u\n", oo_order(s->oo));
4980 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4982 return sprintf(buf, "%lu\n", s->min_partial);
4985 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4991 err = kstrtoul(buf, 10, &min);
4995 set_min_partial(s, min);
4998 SLAB_ATTR(min_partial);
5000 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
5002 return sprintf(buf, "%u\n", slub_cpu_partial(s));
5005 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
5008 unsigned int objects;
5011 err = kstrtouint(buf, 10, &objects);
5014 if (objects && !kmem_cache_has_cpu_partial(s))
5017 slub_set_cpu_partial(s, objects);
5021 SLAB_ATTR(cpu_partial);
5023 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
5027 return sprintf(buf, "%pS\n", s->ctor);
5031 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
5033 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
5035 SLAB_ATTR_RO(aliases);
5037 static ssize_t partial_show(struct kmem_cache *s, char *buf)
5039 return show_slab_objects(s, buf, SO_PARTIAL);
5041 SLAB_ATTR_RO(partial);
5043 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
5045 return show_slab_objects(s, buf, SO_CPU);
5047 SLAB_ATTR_RO(cpu_slabs);
5049 static ssize_t objects_show(struct kmem_cache *s, char *buf)
5051 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5053 SLAB_ATTR_RO(objects);
5055 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5057 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5059 SLAB_ATTR_RO(objects_partial);
5061 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5068 for_each_online_cpu(cpu) {
5071 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5074 pages += page->pages;
5075 objects += page->pobjects;
5079 len = sprintf(buf, "%d(%d)", objects, pages);
5082 for_each_online_cpu(cpu) {
5085 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5087 if (page && len < PAGE_SIZE - 20)
5088 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
5089 page->pobjects, page->pages);
5092 return len + sprintf(buf + len, "\n");
5094 SLAB_ATTR_RO(slabs_cpu_partial);
5096 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5098 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5101 static ssize_t reclaim_account_store(struct kmem_cache *s,
5102 const char *buf, size_t length)
5104 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
5106 s->flags |= SLAB_RECLAIM_ACCOUNT;
5109 SLAB_ATTR(reclaim_account);
5111 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5113 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5115 SLAB_ATTR_RO(hwcache_align);
5117 #ifdef CONFIG_ZONE_DMA
5118 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5120 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5122 SLAB_ATTR_RO(cache_dma);
5125 static ssize_t usersize_show(struct kmem_cache *s, char *buf)
5127 return sprintf(buf, "%u\n", s->usersize);
5129 SLAB_ATTR_RO(usersize);
5131 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5133 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
5135 SLAB_ATTR_RO(destroy_by_rcu);
5137 #ifdef CONFIG_SLUB_DEBUG
5138 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5140 return show_slab_objects(s, buf, SO_ALL);
5142 SLAB_ATTR_RO(slabs);
5144 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5146 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5148 SLAB_ATTR_RO(total_objects);
5150 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5152 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5155 static ssize_t sanity_checks_store(struct kmem_cache *s,
5156 const char *buf, size_t length)
5158 s->flags &= ~SLAB_CONSISTENCY_CHECKS;
5159 if (buf[0] == '1') {
5160 s->flags &= ~__CMPXCHG_DOUBLE;
5161 s->flags |= SLAB_CONSISTENCY_CHECKS;
5165 SLAB_ATTR(sanity_checks);
5167 static ssize_t trace_show(struct kmem_cache *s, char *buf)
5169 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5172 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
5176 * Tracing a merged cache is going to give confusing results
5177 * as well as cause other issues like converting a mergeable
5178 * cache into an umergeable one.
5180 if (s->refcount > 1)
5183 s->flags &= ~SLAB_TRACE;
5184 if (buf[0] == '1') {
5185 s->flags &= ~__CMPXCHG_DOUBLE;
5186 s->flags |= SLAB_TRACE;
5192 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5194 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5197 static ssize_t red_zone_store(struct kmem_cache *s,
5198 const char *buf, size_t length)
5200 if (any_slab_objects(s))
5203 s->flags &= ~SLAB_RED_ZONE;
5204 if (buf[0] == '1') {
5205 s->flags |= SLAB_RED_ZONE;
5207 calculate_sizes(s, -1);
5210 SLAB_ATTR(red_zone);
5212 static ssize_t poison_show(struct kmem_cache *s, char *buf)
5214 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
5217 static ssize_t poison_store(struct kmem_cache *s,
5218 const char *buf, size_t length)
5220 if (any_slab_objects(s))
5223 s->flags &= ~SLAB_POISON;
5224 if (buf[0] == '1') {
5225 s->flags |= SLAB_POISON;
5227 calculate_sizes(s, -1);
5232 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5234 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5237 static ssize_t store_user_store(struct kmem_cache *s,
5238 const char *buf, size_t length)
5240 if (any_slab_objects(s))
5243 s->flags &= ~SLAB_STORE_USER;
5244 if (buf[0] == '1') {
5245 s->flags &= ~__CMPXCHG_DOUBLE;
5246 s->flags |= SLAB_STORE_USER;
5248 calculate_sizes(s, -1);
5251 SLAB_ATTR(store_user);
5253 static ssize_t validate_show(struct kmem_cache *s, char *buf)
5258 static ssize_t validate_store(struct kmem_cache *s,
5259 const char *buf, size_t length)
5263 if (buf[0] == '1') {
5264 ret = validate_slab_cache(s);
5270 SLAB_ATTR(validate);
5272 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
5274 if (!(s->flags & SLAB_STORE_USER))
5276 return list_locations(s, buf, TRACK_ALLOC);
5278 SLAB_ATTR_RO(alloc_calls);
5280 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
5282 if (!(s->flags & SLAB_STORE_USER))
5284 return list_locations(s, buf, TRACK_FREE);
5286 SLAB_ATTR_RO(free_calls);
5287 #endif /* CONFIG_SLUB_DEBUG */
5289 #ifdef CONFIG_FAILSLAB
5290 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5292 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5295 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
5298 if (s->refcount > 1)
5301 s->flags &= ~SLAB_FAILSLAB;
5303 s->flags |= SLAB_FAILSLAB;
5306 SLAB_ATTR(failslab);
5309 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5314 static ssize_t shrink_store(struct kmem_cache *s,
5315 const char *buf, size_t length)
5318 kmem_cache_shrink(s);
5326 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5328 return sprintf(buf, "%u\n", s->remote_node_defrag_ratio / 10);
5331 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5332 const char *buf, size_t length)
5337 err = kstrtouint(buf, 10, &ratio);
5343 s->remote_node_defrag_ratio = ratio * 10;
5347 SLAB_ATTR(remote_node_defrag_ratio);
5350 #ifdef CONFIG_SLUB_STATS
5351 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5353 unsigned long sum = 0;
5356 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
5361 for_each_online_cpu(cpu) {
5362 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5368 len = sprintf(buf, "%lu", sum);
5371 for_each_online_cpu(cpu) {
5372 if (data[cpu] && len < PAGE_SIZE - 20)
5373 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5377 return len + sprintf(buf + len, "\n");
5380 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5384 for_each_online_cpu(cpu)
5385 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5388 #define STAT_ATTR(si, text) \
5389 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5391 return show_stat(s, buf, si); \
5393 static ssize_t text##_store(struct kmem_cache *s, \
5394 const char *buf, size_t length) \
5396 if (buf[0] != '0') \
5398 clear_stat(s, si); \
5403 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5404 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5405 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5406 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5407 STAT_ATTR(FREE_FROZEN, free_frozen);
5408 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5409 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5410 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5411 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5412 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5413 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5414 STAT_ATTR(FREE_SLAB, free_slab);
5415 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5416 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5417 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5418 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5419 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5420 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5421 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5422 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5423 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5424 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5425 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5426 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5427 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5428 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5431 static struct attribute *slab_attrs[] = {
5432 &slab_size_attr.attr,
5433 &object_size_attr.attr,
5434 &objs_per_slab_attr.attr,
5436 &min_partial_attr.attr,
5437 &cpu_partial_attr.attr,
5439 &objects_partial_attr.attr,
5441 &cpu_slabs_attr.attr,
5445 &hwcache_align_attr.attr,
5446 &reclaim_account_attr.attr,
5447 &destroy_by_rcu_attr.attr,
5449 &slabs_cpu_partial_attr.attr,
5450 #ifdef CONFIG_SLUB_DEBUG
5451 &total_objects_attr.attr,
5453 &sanity_checks_attr.attr,
5455 &red_zone_attr.attr,
5457 &store_user_attr.attr,
5458 &validate_attr.attr,
5459 &alloc_calls_attr.attr,
5460 &free_calls_attr.attr,
5462 #ifdef CONFIG_ZONE_DMA
5463 &cache_dma_attr.attr,
5466 &remote_node_defrag_ratio_attr.attr,
5468 #ifdef CONFIG_SLUB_STATS
5469 &alloc_fastpath_attr.attr,
5470 &alloc_slowpath_attr.attr,
5471 &free_fastpath_attr.attr,
5472 &free_slowpath_attr.attr,
5473 &free_frozen_attr.attr,
5474 &free_add_partial_attr.attr,
5475 &free_remove_partial_attr.attr,
5476 &alloc_from_partial_attr.attr,
5477 &alloc_slab_attr.attr,
5478 &alloc_refill_attr.attr,
5479 &alloc_node_mismatch_attr.attr,
5480 &free_slab_attr.attr,
5481 &cpuslab_flush_attr.attr,
5482 &deactivate_full_attr.attr,
5483 &deactivate_empty_attr.attr,
5484 &deactivate_to_head_attr.attr,
5485 &deactivate_to_tail_attr.attr,
5486 &deactivate_remote_frees_attr.attr,
5487 &deactivate_bypass_attr.attr,
5488 &order_fallback_attr.attr,
5489 &cmpxchg_double_fail_attr.attr,
5490 &cmpxchg_double_cpu_fail_attr.attr,
5491 &cpu_partial_alloc_attr.attr,
5492 &cpu_partial_free_attr.attr,
5493 &cpu_partial_node_attr.attr,
5494 &cpu_partial_drain_attr.attr,
5496 #ifdef CONFIG_FAILSLAB
5497 &failslab_attr.attr,
5499 &usersize_attr.attr,
5504 static const struct attribute_group slab_attr_group = {
5505 .attrs = slab_attrs,
5508 static ssize_t slab_attr_show(struct kobject *kobj,
5509 struct attribute *attr,
5512 struct slab_attribute *attribute;
5513 struct kmem_cache *s;
5516 attribute = to_slab_attr(attr);
5519 if (!attribute->show)
5522 err = attribute->show(s, buf);
5527 static ssize_t slab_attr_store(struct kobject *kobj,
5528 struct attribute *attr,
5529 const char *buf, size_t len)
5531 struct slab_attribute *attribute;
5532 struct kmem_cache *s;
5535 attribute = to_slab_attr(attr);
5538 if (!attribute->store)
5541 err = attribute->store(s, buf, len);
5543 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5544 struct kmem_cache *c;
5546 mutex_lock(&slab_mutex);
5547 if (s->max_attr_size < len)
5548 s->max_attr_size = len;
5551 * This is a best effort propagation, so this function's return
5552 * value will be determined by the parent cache only. This is
5553 * basically because not all attributes will have a well
5554 * defined semantics for rollbacks - most of the actions will
5555 * have permanent effects.
5557 * Returning the error value of any of the children that fail
5558 * is not 100 % defined, in the sense that users seeing the
5559 * error code won't be able to know anything about the state of
5562 * Only returning the error code for the parent cache at least
5563 * has well defined semantics. The cache being written to
5564 * directly either failed or succeeded, in which case we loop
5565 * through the descendants with best-effort propagation.
5567 for_each_memcg_cache(c, s)
5568 attribute->store(c, buf, len);
5569 mutex_unlock(&slab_mutex);
5575 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5579 char *buffer = NULL;
5580 struct kmem_cache *root_cache;
5582 if (is_root_cache(s))
5585 root_cache = s->memcg_params.root_cache;
5588 * This mean this cache had no attribute written. Therefore, no point
5589 * in copying default values around
5591 if (!root_cache->max_attr_size)
5594 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5597 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5600 if (!attr || !attr->store || !attr->show)
5604 * It is really bad that we have to allocate here, so we will
5605 * do it only as a fallback. If we actually allocate, though,
5606 * we can just use the allocated buffer until the end.
5608 * Most of the slub attributes will tend to be very small in
5609 * size, but sysfs allows buffers up to a page, so they can
5610 * theoretically happen.
5614 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf) &&
5615 !IS_ENABLED(CONFIG_SLUB_STATS))
5618 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5619 if (WARN_ON(!buffer))
5624 len = attr->show(root_cache, buf);
5626 attr->store(s, buf, len);
5630 free_page((unsigned long)buffer);
5634 static void kmem_cache_release(struct kobject *k)
5636 slab_kmem_cache_release(to_slab(k));
5639 static const struct sysfs_ops slab_sysfs_ops = {
5640 .show = slab_attr_show,
5641 .store = slab_attr_store,
5644 static struct kobj_type slab_ktype = {
5645 .sysfs_ops = &slab_sysfs_ops,
5646 .release = kmem_cache_release,
5649 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5651 struct kobj_type *ktype = get_ktype(kobj);
5653 if (ktype == &slab_ktype)
5658 static const struct kset_uevent_ops slab_uevent_ops = {
5659 .filter = uevent_filter,
5662 static struct kset *slab_kset;
5664 static inline struct kset *cache_kset(struct kmem_cache *s)
5667 if (!is_root_cache(s))
5668 return s->memcg_params.root_cache->memcg_kset;
5673 #define ID_STR_LENGTH 64
5675 /* Create a unique string id for a slab cache:
5677 * Format :[flags-]size
5679 static char *create_unique_id(struct kmem_cache *s)
5681 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5688 * First flags affecting slabcache operations. We will only
5689 * get here for aliasable slabs so we do not need to support
5690 * too many flags. The flags here must cover all flags that
5691 * are matched during merging to guarantee that the id is
5694 if (s->flags & SLAB_CACHE_DMA)
5696 if (s->flags & SLAB_CACHE_DMA32)
5698 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5700 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5702 if (s->flags & SLAB_ACCOUNT)
5706 p += sprintf(p, "%07u", s->size);
5708 BUG_ON(p > name + ID_STR_LENGTH - 1);
5712 static void sysfs_slab_remove_workfn(struct work_struct *work)
5714 struct kmem_cache *s =
5715 container_of(work, struct kmem_cache, kobj_remove_work);
5717 if (!s->kobj.state_in_sysfs)
5719 * For a memcg cache, this may be called during
5720 * deactivation and again on shutdown. Remove only once.
5721 * A cache is never shut down before deactivation is
5722 * complete, so no need to worry about synchronization.
5727 kset_unregister(s->memcg_kset);
5729 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5731 kobject_put(&s->kobj);
5734 static int sysfs_slab_add(struct kmem_cache *s)
5738 struct kset *kset = cache_kset(s);
5739 int unmergeable = slab_unmergeable(s);
5741 INIT_WORK(&s->kobj_remove_work, sysfs_slab_remove_workfn);
5744 kobject_init(&s->kobj, &slab_ktype);
5748 if (!unmergeable && disable_higher_order_debug &&
5749 (slub_debug & DEBUG_METADATA_FLAGS))
5754 * Slabcache can never be merged so we can use the name proper.
5755 * This is typically the case for debug situations. In that
5756 * case we can catch duplicate names easily.
5758 sysfs_remove_link(&slab_kset->kobj, s->name);
5762 * Create a unique name for the slab as a target
5765 name = create_unique_id(s);
5768 s->kobj.kset = kset;
5769 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5773 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5778 if (is_root_cache(s) && memcg_sysfs_enabled) {
5779 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5780 if (!s->memcg_kset) {
5787 kobject_uevent(&s->kobj, KOBJ_ADD);
5789 /* Setup first alias */
5790 sysfs_slab_alias(s, s->name);
5797 kobject_del(&s->kobj);
5801 static void sysfs_slab_remove(struct kmem_cache *s)
5803 if (slab_state < FULL)
5805 * Sysfs has not been setup yet so no need to remove the
5810 kobject_get(&s->kobj);
5811 schedule_work(&s->kobj_remove_work);
5814 void sysfs_slab_unlink(struct kmem_cache *s)
5816 if (slab_state >= FULL)
5817 kobject_del(&s->kobj);
5820 void sysfs_slab_release(struct kmem_cache *s)
5822 if (slab_state >= FULL)
5823 kobject_put(&s->kobj);
5827 * Need to buffer aliases during bootup until sysfs becomes
5828 * available lest we lose that information.
5830 struct saved_alias {
5831 struct kmem_cache *s;
5833 struct saved_alias *next;
5836 static struct saved_alias *alias_list;
5838 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5840 struct saved_alias *al;
5842 if (slab_state == FULL) {
5844 * If we have a leftover link then remove it.
5846 sysfs_remove_link(&slab_kset->kobj, name);
5847 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5850 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5856 al->next = alias_list;
5861 static int __init slab_sysfs_init(void)
5863 struct kmem_cache *s;
5866 mutex_lock(&slab_mutex);
5868 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5870 mutex_unlock(&slab_mutex);
5871 pr_err("Cannot register slab subsystem.\n");
5877 list_for_each_entry(s, &slab_caches, list) {
5878 err = sysfs_slab_add(s);
5880 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5884 while (alias_list) {
5885 struct saved_alias *al = alias_list;
5887 alias_list = alias_list->next;
5888 err = sysfs_slab_alias(al->s, al->name);
5890 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5895 mutex_unlock(&slab_mutex);
5900 __initcall(slab_sysfs_init);
5901 #endif /* CONFIG_SYSFS */
5904 * The /proc/slabinfo ABI
5906 #ifdef CONFIG_SLUB_DEBUG
5907 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5909 unsigned long nr_slabs = 0;
5910 unsigned long nr_objs = 0;
5911 unsigned long nr_free = 0;
5913 struct kmem_cache_node *n;
5915 for_each_kmem_cache_node(s, node, n) {
5916 nr_slabs += node_nr_slabs(n);
5917 nr_objs += node_nr_objs(n);
5918 nr_free += count_partial(n, count_free);
5921 sinfo->active_objs = nr_objs - nr_free;
5922 sinfo->num_objs = nr_objs;
5923 sinfo->active_slabs = nr_slabs;
5924 sinfo->num_slabs = nr_slabs;
5925 sinfo->objects_per_slab = oo_objects(s->oo);
5926 sinfo->cache_order = oo_order(s->oo);
5929 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5933 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5934 size_t count, loff_t *ppos)
5938 #endif /* CONFIG_SLUB_DEBUG */