1 // SPDX-License-Identifier: GPL-2.0
3 * SLUB: A slab allocator that limits cache line use instead of queuing
4 * objects in per cpu and per node lists.
6 * The allocator synchronizes using per slab locks or atomic operatios
7 * and only uses a centralized lock to manage a pool of partial slabs.
9 * (C) 2007 SGI, Christoph Lameter
10 * (C) 2011 Linux Foundation, Christoph Lameter
14 #include <linux/swap.h> /* struct reclaim_state */
15 #include <linux/module.h>
16 #include <linux/bit_spinlock.h>
17 #include <linux/interrupt.h>
18 #include <linux/bitops.h>
19 #include <linux/slab.h>
21 #include <linux/proc_fs.h>
22 #include <linux/seq_file.h>
23 #include <linux/kasan.h>
24 #include <linux/cpu.h>
25 #include <linux/cpuset.h>
26 #include <linux/mempolicy.h>
27 #include <linux/ctype.h>
28 #include <linux/debugobjects.h>
29 #include <linux/kallsyms.h>
30 #include <linux/memory.h>
31 #include <linux/math64.h>
32 #include <linux/fault-inject.h>
33 #include <linux/stacktrace.h>
34 #include <linux/prefetch.h>
35 #include <linux/memcontrol.h>
36 #include <linux/random.h>
38 #include <trace/events/kmem.h>
44 * 1. slab_mutex (Global Mutex)
46 * 3. slab_lock(page) (Only on some arches and for debugging)
50 * The role of the slab_mutex is to protect the list of all the slabs
51 * and to synchronize major metadata changes to slab cache structures.
53 * The slab_lock is only used for debugging and on arches that do not
54 * have the ability to do a cmpxchg_double. It only protects:
55 * A. page->freelist -> List of object free in a page
56 * B. page->inuse -> Number of objects in use
57 * C. page->objects -> Number of objects in page
58 * D. page->frozen -> frozen state
60 * If a slab is frozen then it is exempt from list management. It is not
61 * on any list. The processor that froze the slab is the one who can
62 * perform list operations on the page. Other processors may put objects
63 * onto the freelist but the processor that froze the slab is the only
64 * one that can retrieve the objects from the page's freelist.
66 * The list_lock protects the partial and full list on each node and
67 * the partial slab counter. If taken then no new slabs may be added or
68 * removed from the lists nor make the number of partial slabs be modified.
69 * (Note that the total number of slabs is an atomic value that may be
70 * modified without taking the list lock).
72 * The list_lock is a centralized lock and thus we avoid taking it as
73 * much as possible. As long as SLUB does not have to handle partial
74 * slabs, operations can continue without any centralized lock. F.e.
75 * allocating a long series of objects that fill up slabs does not require
77 * Interrupts are disabled during allocation and deallocation in order to
78 * make the slab allocator safe to use in the context of an irq. In addition
79 * interrupts are disabled to ensure that the processor does not change
80 * while handling per_cpu slabs, due to kernel preemption.
82 * SLUB assigns one slab for allocation to each processor.
83 * Allocations only occur from these slabs called cpu slabs.
85 * Slabs with free elements are kept on a partial list and during regular
86 * operations no list for full slabs is used. If an object in a full slab is
87 * freed then the slab will show up again on the partial lists.
88 * We track full slabs for debugging purposes though because otherwise we
89 * cannot scan all objects.
91 * Slabs are freed when they become empty. Teardown and setup is
92 * minimal so we rely on the page allocators per cpu caches for
93 * fast frees and allocs.
95 * Overloading of page flags that are otherwise used for LRU management.
97 * PageActive The slab is frozen and exempt from list processing.
98 * This means that the slab is dedicated to a purpose
99 * such as satisfying allocations for a specific
100 * processor. Objects may be freed in the slab while
101 * it is frozen but slab_free will then skip the usual
102 * list operations. It is up to the processor holding
103 * the slab to integrate the slab into the slab lists
104 * when the slab is no longer needed.
106 * One use of this flag is to mark slabs that are
107 * used for allocations. Then such a slab becomes a cpu
108 * slab. The cpu slab may be equipped with an additional
109 * freelist that allows lockless access to
110 * free objects in addition to the regular freelist
111 * that requires the slab lock.
113 * PageError Slab requires special handling due to debug
114 * options set. This moves slab handling out of
115 * the fast path and disables lockless freelists.
118 static inline int kmem_cache_debug(struct kmem_cache *s)
120 #ifdef CONFIG_SLUB_DEBUG
121 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
127 void *fixup_red_left(struct kmem_cache *s, void *p)
129 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE)
130 p += s->red_left_pad;
135 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
137 #ifdef CONFIG_SLUB_CPU_PARTIAL
138 return !kmem_cache_debug(s);
145 * Issues still to be resolved:
147 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
149 * - Variable sizing of the per node arrays
152 /* Enable to test recovery from slab corruption on boot */
153 #undef SLUB_RESILIENCY_TEST
155 /* Enable to log cmpxchg failures */
156 #undef SLUB_DEBUG_CMPXCHG
159 * Mininum number of partial slabs. These will be left on the partial
160 * lists even if they are empty. kmem_cache_shrink may reclaim them.
162 #define MIN_PARTIAL 5
165 * Maximum number of desirable partial slabs.
166 * The existence of more partial slabs makes kmem_cache_shrink
167 * sort the partial list by the number of objects in use.
169 #define MAX_PARTIAL 10
171 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
172 SLAB_POISON | SLAB_STORE_USER)
175 * These debug flags cannot use CMPXCHG because there might be consistency
176 * issues when checking or reading debug information
178 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
183 * Debugging flags that require metadata to be stored in the slab. These get
184 * disabled when slub_debug=O is used and a cache's min order increases with
187 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
190 #define OO_MASK ((1 << OO_SHIFT) - 1)
191 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
193 /* Internal SLUB flags */
195 #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
196 /* Use cmpxchg_double */
197 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
200 * Tracking user of a slab.
202 #define TRACK_ADDRS_COUNT 16
204 unsigned long addr; /* Called from address */
205 #ifdef CONFIG_STACKTRACE
206 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
208 int cpu; /* Was running on cpu */
209 int pid; /* Pid context */
210 unsigned long when; /* When did the operation occur */
213 enum track_item { TRACK_ALLOC, TRACK_FREE };
216 static int sysfs_slab_add(struct kmem_cache *);
217 static int sysfs_slab_alias(struct kmem_cache *, const char *);
218 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
219 static void sysfs_slab_remove(struct kmem_cache *s);
221 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
222 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
224 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
225 static inline void sysfs_slab_remove(struct kmem_cache *s) { }
228 static inline void stat(const struct kmem_cache *s, enum stat_item si)
230 #ifdef CONFIG_SLUB_STATS
232 * The rmw is racy on a preemptible kernel but this is acceptable, so
233 * avoid this_cpu_add()'s irq-disable overhead.
235 raw_cpu_inc(s->cpu_slab->stat[si]);
239 /********************************************************************
240 * Core slab cache functions
241 *******************************************************************/
244 * Returns freelist pointer (ptr). With hardening, this is obfuscated
245 * with an XOR of the address where the pointer is held and a per-cache
248 static inline void *freelist_ptr(const struct kmem_cache *s, void *ptr,
249 unsigned long ptr_addr)
251 #ifdef CONFIG_SLAB_FREELIST_HARDENED
252 return (void *)((unsigned long)ptr ^ s->random ^ ptr_addr);
258 /* Returns the freelist pointer recorded at location ptr_addr. */
259 static inline void *freelist_dereference(const struct kmem_cache *s,
262 return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr),
263 (unsigned long)ptr_addr);
266 static inline void *get_freepointer(struct kmem_cache *s, void *object)
268 return freelist_dereference(s, object + s->offset);
271 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
273 prefetch(object + s->offset);
276 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
278 unsigned long freepointer_addr;
281 if (!debug_pagealloc_enabled())
282 return get_freepointer(s, object);
284 freepointer_addr = (unsigned long)object + s->offset;
285 probe_kernel_read(&p, (void **)freepointer_addr, sizeof(p));
286 return freelist_ptr(s, p, freepointer_addr);
289 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
291 unsigned long freeptr_addr = (unsigned long)object + s->offset;
293 #ifdef CONFIG_SLAB_FREELIST_HARDENED
294 BUG_ON(object == fp); /* naive detection of double free or corruption */
297 *(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr);
300 /* Loop over all objects in a slab */
301 #define for_each_object(__p, __s, __addr, __objects) \
302 for (__p = fixup_red_left(__s, __addr); \
303 __p < (__addr) + (__objects) * (__s)->size; \
306 #define for_each_object_idx(__p, __idx, __s, __addr, __objects) \
307 for (__p = fixup_red_left(__s, __addr), __idx = 1; \
308 __idx <= __objects; \
309 __p += (__s)->size, __idx++)
311 /* Determine object index from a given position */
312 static inline unsigned int slab_index(void *p, struct kmem_cache *s, void *addr)
314 return (p - addr) / s->size;
317 static inline unsigned int order_objects(unsigned int order, unsigned int size)
319 return ((unsigned int)PAGE_SIZE << order) / size;
322 static inline struct kmem_cache_order_objects oo_make(unsigned int order,
325 struct kmem_cache_order_objects x = {
326 (order << OO_SHIFT) + order_objects(order, size)
332 static inline unsigned int oo_order(struct kmem_cache_order_objects x)
334 return x.x >> OO_SHIFT;
337 static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
339 return x.x & OO_MASK;
343 * Per slab locking using the pagelock
345 static __always_inline void slab_lock(struct page *page)
347 VM_BUG_ON_PAGE(PageTail(page), page);
348 bit_spin_lock(PG_locked, &page->flags);
351 static __always_inline void slab_unlock(struct page *page)
353 VM_BUG_ON_PAGE(PageTail(page), page);
354 __bit_spin_unlock(PG_locked, &page->flags);
357 /* Interrupts must be disabled (for the fallback code to work right) */
358 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
359 void *freelist_old, unsigned long counters_old,
360 void *freelist_new, unsigned long counters_new,
363 VM_BUG_ON(!irqs_disabled());
364 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
365 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
366 if (s->flags & __CMPXCHG_DOUBLE) {
367 if (cmpxchg_double(&page->freelist, &page->counters,
368 freelist_old, counters_old,
369 freelist_new, counters_new))
375 if (page->freelist == freelist_old &&
376 page->counters == counters_old) {
377 page->freelist = freelist_new;
378 page->counters = counters_new;
386 stat(s, CMPXCHG_DOUBLE_FAIL);
388 #ifdef SLUB_DEBUG_CMPXCHG
389 pr_info("%s %s: cmpxchg double redo ", n, s->name);
395 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
396 void *freelist_old, unsigned long counters_old,
397 void *freelist_new, unsigned long counters_new,
400 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
401 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
402 if (s->flags & __CMPXCHG_DOUBLE) {
403 if (cmpxchg_double(&page->freelist, &page->counters,
404 freelist_old, counters_old,
405 freelist_new, counters_new))
412 local_irq_save(flags);
414 if (page->freelist == freelist_old &&
415 page->counters == counters_old) {
416 page->freelist = freelist_new;
417 page->counters = counters_new;
419 local_irq_restore(flags);
423 local_irq_restore(flags);
427 stat(s, CMPXCHG_DOUBLE_FAIL);
429 #ifdef SLUB_DEBUG_CMPXCHG
430 pr_info("%s %s: cmpxchg double redo ", n, s->name);
436 #ifdef CONFIG_SLUB_DEBUG
438 * Determine a map of object in use on a page.
440 * Node listlock must be held to guarantee that the page does
441 * not vanish from under us.
443 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
446 void *addr = page_address(page);
448 for (p = page->freelist; p; p = get_freepointer(s, p))
449 set_bit(slab_index(p, s, addr), map);
452 static inline unsigned int size_from_object(struct kmem_cache *s)
454 if (s->flags & SLAB_RED_ZONE)
455 return s->size - s->red_left_pad;
460 static inline void *restore_red_left(struct kmem_cache *s, void *p)
462 if (s->flags & SLAB_RED_ZONE)
463 p -= s->red_left_pad;
471 #if defined(CONFIG_SLUB_DEBUG_ON)
472 static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
474 static slab_flags_t slub_debug;
477 static char *slub_debug_slabs;
478 static int disable_higher_order_debug;
481 * slub is about to manipulate internal object metadata. This memory lies
482 * outside the range of the allocated object, so accessing it would normally
483 * be reported by kasan as a bounds error. metadata_access_enable() is used
484 * to tell kasan that these accesses are OK.
486 static inline void metadata_access_enable(void)
488 kasan_disable_current();
491 static inline void metadata_access_disable(void)
493 kasan_enable_current();
500 /* Verify that a pointer has an address that is valid within a slab page */
501 static inline int check_valid_pointer(struct kmem_cache *s,
502 struct page *page, void *object)
509 base = page_address(page);
510 object = restore_red_left(s, object);
511 if (object < base || object >= base + page->objects * s->size ||
512 (object - base) % s->size) {
519 static void print_section(char *level, char *text, u8 *addr,
522 metadata_access_enable();
523 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
525 metadata_access_disable();
528 static struct track *get_track(struct kmem_cache *s, void *object,
529 enum track_item alloc)
534 p = object + s->offset + sizeof(void *);
536 p = object + s->inuse;
541 static void set_track(struct kmem_cache *s, void *object,
542 enum track_item alloc, unsigned long addr)
544 struct track *p = get_track(s, object, alloc);
547 #ifdef CONFIG_STACKTRACE
548 struct stack_trace trace;
551 trace.nr_entries = 0;
552 trace.max_entries = TRACK_ADDRS_COUNT;
553 trace.entries = p->addrs;
555 metadata_access_enable();
556 save_stack_trace(&trace);
557 metadata_access_disable();
559 /* See rant in lockdep.c */
560 if (trace.nr_entries != 0 &&
561 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
564 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
568 p->cpu = smp_processor_id();
569 p->pid = current->pid;
572 memset(p, 0, sizeof(struct track));
575 static void init_tracking(struct kmem_cache *s, void *object)
577 if (!(s->flags & SLAB_STORE_USER))
580 set_track(s, object, TRACK_FREE, 0UL);
581 set_track(s, object, TRACK_ALLOC, 0UL);
584 static void print_track(const char *s, struct track *t, unsigned long pr_time)
589 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
590 s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
591 #ifdef CONFIG_STACKTRACE
594 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
596 pr_err("\t%pS\n", (void *)t->addrs[i]);
603 static void print_tracking(struct kmem_cache *s, void *object)
605 unsigned long pr_time = jiffies;
606 if (!(s->flags & SLAB_STORE_USER))
609 print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
610 print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
613 static void print_page_info(struct page *page)
615 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
616 page, page->objects, page->inuse, page->freelist, page->flags);
620 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
622 struct va_format vaf;
628 pr_err("=============================================================================\n");
629 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
630 pr_err("-----------------------------------------------------------------------------\n\n");
632 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
636 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
638 struct va_format vaf;
644 pr_err("FIX %s: %pV\n", s->name, &vaf);
648 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
650 unsigned int off; /* Offset of last byte */
651 u8 *addr = page_address(page);
653 print_tracking(s, p);
655 print_page_info(page);
657 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
658 p, p - addr, get_freepointer(s, p));
660 if (s->flags & SLAB_RED_ZONE)
661 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
663 else if (p > addr + 16)
664 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
666 print_section(KERN_ERR, "Object ", p,
667 min_t(unsigned int, s->object_size, PAGE_SIZE));
668 if (s->flags & SLAB_RED_ZONE)
669 print_section(KERN_ERR, "Redzone ", p + s->object_size,
670 s->inuse - s->object_size);
673 off = s->offset + sizeof(void *);
677 if (s->flags & SLAB_STORE_USER)
678 off += 2 * sizeof(struct track);
680 off += kasan_metadata_size(s);
682 if (off != size_from_object(s))
683 /* Beginning of the filler is the free pointer */
684 print_section(KERN_ERR, "Padding ", p + off,
685 size_from_object(s) - off);
690 void object_err(struct kmem_cache *s, struct page *page,
691 u8 *object, char *reason)
693 slab_bug(s, "%s", reason);
694 print_trailer(s, page, object);
697 static __printf(3, 4) void slab_err(struct kmem_cache *s, struct page *page,
698 const char *fmt, ...)
704 vsnprintf(buf, sizeof(buf), fmt, args);
706 slab_bug(s, "%s", buf);
707 print_page_info(page);
711 static void init_object(struct kmem_cache *s, void *object, u8 val)
715 if (s->flags & SLAB_RED_ZONE)
716 memset(p - s->red_left_pad, val, s->red_left_pad);
718 if (s->flags & __OBJECT_POISON) {
719 memset(p, POISON_FREE, s->object_size - 1);
720 p[s->object_size - 1] = POISON_END;
723 if (s->flags & SLAB_RED_ZONE)
724 memset(p + s->object_size, val, s->inuse - s->object_size);
727 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
728 void *from, void *to)
730 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
731 memset(from, data, to - from);
734 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
735 u8 *object, char *what,
736 u8 *start, unsigned int value, unsigned int bytes)
741 metadata_access_enable();
742 fault = memchr_inv(start, value, bytes);
743 metadata_access_disable();
748 while (end > fault && end[-1] == value)
751 slab_bug(s, "%s overwritten", what);
752 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
753 fault, end - 1, fault[0], value);
754 print_trailer(s, page, object);
756 restore_bytes(s, what, value, fault, end);
764 * Bytes of the object to be managed.
765 * If the freepointer may overlay the object then the free
766 * pointer is the first word of the object.
768 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
771 * object + s->object_size
772 * Padding to reach word boundary. This is also used for Redzoning.
773 * Padding is extended by another word if Redzoning is enabled and
774 * object_size == inuse.
776 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
777 * 0xcc (RED_ACTIVE) for objects in use.
780 * Meta data starts here.
782 * A. Free pointer (if we cannot overwrite object on free)
783 * B. Tracking data for SLAB_STORE_USER
784 * C. Padding to reach required alignment boundary or at mininum
785 * one word if debugging is on to be able to detect writes
786 * before the word boundary.
788 * Padding is done using 0x5a (POISON_INUSE)
791 * Nothing is used beyond s->size.
793 * If slabcaches are merged then the object_size and inuse boundaries are mostly
794 * ignored. And therefore no slab options that rely on these boundaries
795 * may be used with merged slabcaches.
798 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
800 unsigned long off = s->inuse; /* The end of info */
803 /* Freepointer is placed after the object. */
804 off += sizeof(void *);
806 if (s->flags & SLAB_STORE_USER)
807 /* We also have user information there */
808 off += 2 * sizeof(struct track);
810 off += kasan_metadata_size(s);
812 if (size_from_object(s) == off)
815 return check_bytes_and_report(s, page, p, "Object padding",
816 p + off, POISON_INUSE, size_from_object(s) - off);
819 /* Check the pad bytes at the end of a slab page */
820 static int slab_pad_check(struct kmem_cache *s, struct page *page)
829 if (!(s->flags & SLAB_POISON))
832 start = page_address(page);
833 length = PAGE_SIZE << compound_order(page);
834 end = start + length;
835 remainder = length % s->size;
839 pad = end - remainder;
840 metadata_access_enable();
841 fault = memchr_inv(pad, POISON_INUSE, remainder);
842 metadata_access_disable();
845 while (end > fault && end[-1] == POISON_INUSE)
848 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
849 print_section(KERN_ERR, "Padding ", pad, remainder);
851 restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
855 static int check_object(struct kmem_cache *s, struct page *page,
856 void *object, u8 val)
859 u8 *endobject = object + s->object_size;
861 if (s->flags & SLAB_RED_ZONE) {
862 if (!check_bytes_and_report(s, page, object, "Redzone",
863 object - s->red_left_pad, val, s->red_left_pad))
866 if (!check_bytes_and_report(s, page, object, "Redzone",
867 endobject, val, s->inuse - s->object_size))
870 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
871 check_bytes_and_report(s, page, p, "Alignment padding",
872 endobject, POISON_INUSE,
873 s->inuse - s->object_size);
877 if (s->flags & SLAB_POISON) {
878 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
879 (!check_bytes_and_report(s, page, p, "Poison", p,
880 POISON_FREE, s->object_size - 1) ||
881 !check_bytes_and_report(s, page, p, "Poison",
882 p + s->object_size - 1, POISON_END, 1)))
885 * check_pad_bytes cleans up on its own.
887 check_pad_bytes(s, page, p);
890 if (!s->offset && val == SLUB_RED_ACTIVE)
892 * Object and freepointer overlap. Cannot check
893 * freepointer while object is allocated.
897 /* Check free pointer validity */
898 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
899 object_err(s, page, p, "Freepointer corrupt");
901 * No choice but to zap it and thus lose the remainder
902 * of the free objects in this slab. May cause
903 * another error because the object count is now wrong.
905 set_freepointer(s, p, NULL);
911 static int check_slab(struct kmem_cache *s, struct page *page)
915 VM_BUG_ON(!irqs_disabled());
917 if (!PageSlab(page)) {
918 slab_err(s, page, "Not a valid slab page");
922 maxobj = order_objects(compound_order(page), s->size);
923 if (page->objects > maxobj) {
924 slab_err(s, page, "objects %u > max %u",
925 page->objects, maxobj);
928 if (page->inuse > page->objects) {
929 slab_err(s, page, "inuse %u > max %u",
930 page->inuse, page->objects);
933 /* Slab_pad_check fixes things up after itself */
934 slab_pad_check(s, page);
939 * Determine if a certain object on a page is on the freelist. Must hold the
940 * slab lock to guarantee that the chains are in a consistent state.
942 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
950 while (fp && nr <= page->objects) {
953 if (!check_valid_pointer(s, page, fp)) {
955 object_err(s, page, object,
956 "Freechain corrupt");
957 set_freepointer(s, object, NULL);
959 slab_err(s, page, "Freepointer corrupt");
960 page->freelist = NULL;
961 page->inuse = page->objects;
962 slab_fix(s, "Freelist cleared");
968 fp = get_freepointer(s, object);
972 max_objects = order_objects(compound_order(page), s->size);
973 if (max_objects > MAX_OBJS_PER_PAGE)
974 max_objects = MAX_OBJS_PER_PAGE;
976 if (page->objects != max_objects) {
977 slab_err(s, page, "Wrong number of objects. Found %d but should be %d",
978 page->objects, max_objects);
979 page->objects = max_objects;
980 slab_fix(s, "Number of objects adjusted.");
982 if (page->inuse != page->objects - nr) {
983 slab_err(s, page, "Wrong object count. Counter is %d but counted were %d",
984 page->inuse, page->objects - nr);
985 page->inuse = page->objects - nr;
986 slab_fix(s, "Object count adjusted.");
988 return search == NULL;
991 static void trace(struct kmem_cache *s, struct page *page, void *object,
994 if (s->flags & SLAB_TRACE) {
995 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
997 alloc ? "alloc" : "free",
1002 print_section(KERN_INFO, "Object ", (void *)object,
1010 * Tracking of fully allocated slabs for debugging purposes.
1012 static void add_full(struct kmem_cache *s,
1013 struct kmem_cache_node *n, struct page *page)
1015 if (!(s->flags & SLAB_STORE_USER))
1018 lockdep_assert_held(&n->list_lock);
1019 list_add(&page->lru, &n->full);
1022 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
1024 if (!(s->flags & SLAB_STORE_USER))
1027 lockdep_assert_held(&n->list_lock);
1028 list_del(&page->lru);
1031 /* Tracking of the number of slabs for debugging purposes */
1032 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1034 struct kmem_cache_node *n = get_node(s, node);
1036 return atomic_long_read(&n->nr_slabs);
1039 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1041 return atomic_long_read(&n->nr_slabs);
1044 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1046 struct kmem_cache_node *n = get_node(s, node);
1049 * May be called early in order to allocate a slab for the
1050 * kmem_cache_node structure. Solve the chicken-egg
1051 * dilemma by deferring the increment of the count during
1052 * bootstrap (see early_kmem_cache_node_alloc).
1055 atomic_long_inc(&n->nr_slabs);
1056 atomic_long_add(objects, &n->total_objects);
1059 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1061 struct kmem_cache_node *n = get_node(s, node);
1063 atomic_long_dec(&n->nr_slabs);
1064 atomic_long_sub(objects, &n->total_objects);
1067 /* Object debug checks for alloc/free paths */
1068 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1071 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1074 init_object(s, object, SLUB_RED_INACTIVE);
1075 init_tracking(s, object);
1078 static inline int alloc_consistency_checks(struct kmem_cache *s,
1080 void *object, unsigned long addr)
1082 if (!check_slab(s, page))
1085 if (!check_valid_pointer(s, page, object)) {
1086 object_err(s, page, object, "Freelist Pointer check fails");
1090 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1096 static noinline int alloc_debug_processing(struct kmem_cache *s,
1098 void *object, unsigned long addr)
1100 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1101 if (!alloc_consistency_checks(s, page, object, addr))
1105 /* Success perform special debug activities for allocs */
1106 if (s->flags & SLAB_STORE_USER)
1107 set_track(s, object, TRACK_ALLOC, addr);
1108 trace(s, page, object, 1);
1109 init_object(s, object, SLUB_RED_ACTIVE);
1113 if (PageSlab(page)) {
1115 * If this is a slab page then lets do the best we can
1116 * to avoid issues in the future. Marking all objects
1117 * as used avoids touching the remaining objects.
1119 slab_fix(s, "Marking all objects used");
1120 page->inuse = page->objects;
1121 page->freelist = NULL;
1126 static inline int free_consistency_checks(struct kmem_cache *s,
1127 struct page *page, void *object, unsigned long addr)
1129 if (!check_valid_pointer(s, page, object)) {
1130 slab_err(s, page, "Invalid object pointer 0x%p", object);
1134 if (on_freelist(s, page, object)) {
1135 object_err(s, page, object, "Object already free");
1139 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1142 if (unlikely(s != page->slab_cache)) {
1143 if (!PageSlab(page)) {
1144 slab_err(s, page, "Attempt to free object(0x%p) outside of slab",
1146 } else if (!page->slab_cache) {
1147 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1151 object_err(s, page, object,
1152 "page slab pointer corrupt.");
1158 /* Supports checking bulk free of a constructed freelist */
1159 static noinline int free_debug_processing(
1160 struct kmem_cache *s, struct page *page,
1161 void *head, void *tail, int bulk_cnt,
1164 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1165 void *object = head;
1167 unsigned long uninitialized_var(flags);
1170 spin_lock_irqsave(&n->list_lock, flags);
1173 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1174 if (!check_slab(s, page))
1181 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1182 if (!free_consistency_checks(s, page, object, addr))
1186 if (s->flags & SLAB_STORE_USER)
1187 set_track(s, object, TRACK_FREE, addr);
1188 trace(s, page, object, 0);
1189 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1190 init_object(s, object, SLUB_RED_INACTIVE);
1192 /* Reached end of constructed freelist yet? */
1193 if (object != tail) {
1194 object = get_freepointer(s, object);
1200 if (cnt != bulk_cnt)
1201 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1205 spin_unlock_irqrestore(&n->list_lock, flags);
1207 slab_fix(s, "Object at 0x%p not freed", object);
1211 static int __init setup_slub_debug(char *str)
1213 slub_debug = DEBUG_DEFAULT_FLAGS;
1214 if (*str++ != '=' || !*str)
1216 * No options specified. Switch on full debugging.
1222 * No options but restriction on slabs. This means full
1223 * debugging for slabs matching a pattern.
1230 * Switch off all debugging measures.
1235 * Determine which debug features should be switched on
1237 for (; *str && *str != ','; str++) {
1238 switch (tolower(*str)) {
1240 slub_debug |= SLAB_CONSISTENCY_CHECKS;
1243 slub_debug |= SLAB_RED_ZONE;
1246 slub_debug |= SLAB_POISON;
1249 slub_debug |= SLAB_STORE_USER;
1252 slub_debug |= SLAB_TRACE;
1255 slub_debug |= SLAB_FAILSLAB;
1259 * Avoid enabling debugging on caches if its minimum
1260 * order would increase as a result.
1262 disable_higher_order_debug = 1;
1265 pr_err("slub_debug option '%c' unknown. skipped\n",
1272 slub_debug_slabs = str + 1;
1277 __setup("slub_debug", setup_slub_debug);
1279 slab_flags_t kmem_cache_flags(unsigned int object_size,
1280 slab_flags_t flags, const char *name,
1281 void (*ctor)(void *))
1284 * Enable debugging if selected on the kernel commandline.
1286 if (slub_debug && (!slub_debug_slabs || (name &&
1287 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1288 flags |= slub_debug;
1292 #else /* !CONFIG_SLUB_DEBUG */
1293 static inline void setup_object_debug(struct kmem_cache *s,
1294 struct page *page, void *object) {}
1296 static inline int alloc_debug_processing(struct kmem_cache *s,
1297 struct page *page, void *object, unsigned long addr) { return 0; }
1299 static inline int free_debug_processing(
1300 struct kmem_cache *s, struct page *page,
1301 void *head, void *tail, int bulk_cnt,
1302 unsigned long addr) { return 0; }
1304 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1306 static inline int check_object(struct kmem_cache *s, struct page *page,
1307 void *object, u8 val) { return 1; }
1308 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1309 struct page *page) {}
1310 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1311 struct page *page) {}
1312 slab_flags_t kmem_cache_flags(unsigned int object_size,
1313 slab_flags_t flags, const char *name,
1314 void (*ctor)(void *))
1318 #define slub_debug 0
1320 #define disable_higher_order_debug 0
1322 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1324 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1326 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1328 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1331 #endif /* CONFIG_SLUB_DEBUG */
1334 * Hooks for other subsystems that check memory allocations. In a typical
1335 * production configuration these hooks all should produce no code at all.
1337 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1339 kmemleak_alloc(ptr, size, 1, flags);
1340 kasan_kmalloc_large(ptr, size, flags);
1343 static __always_inline void kfree_hook(void *x)
1346 kasan_kfree_large(x, _RET_IP_);
1349 static __always_inline bool slab_free_hook(struct kmem_cache *s, void *x)
1351 kmemleak_free_recursive(x, s->flags);
1354 * Trouble is that we may no longer disable interrupts in the fast path
1355 * So in order to make the debug calls that expect irqs to be
1356 * disabled we need to disable interrupts temporarily.
1358 #ifdef CONFIG_LOCKDEP
1360 unsigned long flags;
1362 local_irq_save(flags);
1363 debug_check_no_locks_freed(x, s->object_size);
1364 local_irq_restore(flags);
1367 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1368 debug_check_no_obj_freed(x, s->object_size);
1370 /* KASAN might put x into memory quarantine, delaying its reuse */
1371 return kasan_slab_free(s, x, _RET_IP_);
1374 static inline bool slab_free_freelist_hook(struct kmem_cache *s,
1375 void **head, void **tail)
1378 * Compiler cannot detect this function can be removed if slab_free_hook()
1379 * evaluates to nothing. Thus, catch all relevant config debug options here.
1381 #if defined(CONFIG_LOCKDEP) || \
1382 defined(CONFIG_DEBUG_KMEMLEAK) || \
1383 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1384 defined(CONFIG_KASAN)
1388 void *old_tail = *tail ? *tail : *head;
1390 /* Head and tail of the reconstructed freelist */
1396 next = get_freepointer(s, object);
1397 /* If object's reuse doesn't have to be delayed */
1398 if (!slab_free_hook(s, object)) {
1399 /* Move object to the new freelist */
1400 set_freepointer(s, object, *head);
1405 } while (object != old_tail);
1410 return *head != NULL;
1416 static void setup_object(struct kmem_cache *s, struct page *page,
1419 setup_object_debug(s, page, object);
1420 kasan_init_slab_obj(s, object);
1421 if (unlikely(s->ctor)) {
1422 kasan_unpoison_object_data(s, object);
1424 kasan_poison_object_data(s, object);
1429 * Slab allocation and freeing
1431 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1432 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1435 unsigned int order = oo_order(oo);
1437 if (node == NUMA_NO_NODE)
1438 page = alloc_pages(flags, order);
1440 page = __alloc_pages_node(node, flags, order);
1442 if (page && memcg_charge_slab(page, flags, order, s)) {
1443 __free_pages(page, order);
1450 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1451 /* Pre-initialize the random sequence cache */
1452 static int init_cache_random_seq(struct kmem_cache *s)
1454 unsigned int count = oo_objects(s->oo);
1457 /* Bailout if already initialised */
1461 err = cache_random_seq_create(s, count, GFP_KERNEL);
1463 pr_err("SLUB: Unable to initialize free list for %s\n",
1468 /* Transform to an offset on the set of pages */
1469 if (s->random_seq) {
1472 for (i = 0; i < count; i++)
1473 s->random_seq[i] *= s->size;
1478 /* Initialize each random sequence freelist per cache */
1479 static void __init init_freelist_randomization(void)
1481 struct kmem_cache *s;
1483 mutex_lock(&slab_mutex);
1485 list_for_each_entry(s, &slab_caches, list)
1486 init_cache_random_seq(s);
1488 mutex_unlock(&slab_mutex);
1491 /* Get the next entry on the pre-computed freelist randomized */
1492 static void *next_freelist_entry(struct kmem_cache *s, struct page *page,
1493 unsigned long *pos, void *start,
1494 unsigned long page_limit,
1495 unsigned long freelist_count)
1500 * If the target page allocation failed, the number of objects on the
1501 * page might be smaller than the usual size defined by the cache.
1504 idx = s->random_seq[*pos];
1506 if (*pos >= freelist_count)
1508 } while (unlikely(idx >= page_limit));
1510 return (char *)start + idx;
1513 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1514 static bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1519 unsigned long idx, pos, page_limit, freelist_count;
1521 if (page->objects < 2 || !s->random_seq)
1524 freelist_count = oo_objects(s->oo);
1525 pos = get_random_int() % freelist_count;
1527 page_limit = page->objects * s->size;
1528 start = fixup_red_left(s, page_address(page));
1530 /* First entry is used as the base of the freelist */
1531 cur = next_freelist_entry(s, page, &pos, start, page_limit,
1533 page->freelist = cur;
1535 for (idx = 1; idx < page->objects; idx++) {
1536 setup_object(s, page, cur);
1537 next = next_freelist_entry(s, page, &pos, start, page_limit,
1539 set_freepointer(s, cur, next);
1542 setup_object(s, page, cur);
1543 set_freepointer(s, cur, NULL);
1548 static inline int init_cache_random_seq(struct kmem_cache *s)
1552 static inline void init_freelist_randomization(void) { }
1553 static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1557 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1559 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1562 struct kmem_cache_order_objects oo = s->oo;
1568 flags &= gfp_allowed_mask;
1570 if (gfpflags_allow_blocking(flags))
1573 flags |= s->allocflags;
1576 * Let the initial higher-order allocation fail under memory pressure
1577 * so we fall-back to the minimum order allocation.
1579 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1580 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1581 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
1583 page = alloc_slab_page(s, alloc_gfp, node, oo);
1584 if (unlikely(!page)) {
1588 * Allocation may have failed due to fragmentation.
1589 * Try a lower order alloc if possible
1591 page = alloc_slab_page(s, alloc_gfp, node, oo);
1592 if (unlikely(!page))
1594 stat(s, ORDER_FALLBACK);
1597 page->objects = oo_objects(oo);
1599 order = compound_order(page);
1600 page->slab_cache = s;
1601 __SetPageSlab(page);
1602 if (page_is_pfmemalloc(page))
1603 SetPageSlabPfmemalloc(page);
1605 start = page_address(page);
1607 if (unlikely(s->flags & SLAB_POISON))
1608 memset(start, POISON_INUSE, PAGE_SIZE << order);
1610 kasan_poison_slab(page);
1612 shuffle = shuffle_freelist(s, page);
1615 for_each_object_idx(p, idx, s, start, page->objects) {
1616 setup_object(s, page, p);
1617 if (likely(idx < page->objects))
1618 set_freepointer(s, p, p + s->size);
1620 set_freepointer(s, p, NULL);
1622 page->freelist = fixup_red_left(s, start);
1625 page->inuse = page->objects;
1629 if (gfpflags_allow_blocking(flags))
1630 local_irq_disable();
1634 mod_lruvec_page_state(page,
1635 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1636 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1639 inc_slabs_node(s, page_to_nid(page), page->objects);
1644 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1646 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1647 gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK;
1648 flags &= ~GFP_SLAB_BUG_MASK;
1649 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
1650 invalid_mask, &invalid_mask, flags, &flags);
1654 return allocate_slab(s,
1655 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1658 static void __free_slab(struct kmem_cache *s, struct page *page)
1660 int order = compound_order(page);
1661 int pages = 1 << order;
1663 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1666 slab_pad_check(s, page);
1667 for_each_object(p, s, page_address(page),
1669 check_object(s, page, p, SLUB_RED_INACTIVE);
1672 mod_lruvec_page_state(page,
1673 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1674 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1677 __ClearPageSlabPfmemalloc(page);
1678 __ClearPageSlab(page);
1680 page->mapping = NULL;
1681 if (current->reclaim_state)
1682 current->reclaim_state->reclaimed_slab += pages;
1683 memcg_uncharge_slab(page, order, s);
1684 __free_pages(page, order);
1687 static void rcu_free_slab(struct rcu_head *h)
1689 struct page *page = container_of(h, struct page, rcu_head);
1691 __free_slab(page->slab_cache, page);
1694 static void free_slab(struct kmem_cache *s, struct page *page)
1696 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) {
1697 call_rcu(&page->rcu_head, rcu_free_slab);
1699 __free_slab(s, page);
1702 static void discard_slab(struct kmem_cache *s, struct page *page)
1704 dec_slabs_node(s, page_to_nid(page), page->objects);
1709 * Management of partially allocated slabs.
1712 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1715 if (tail == DEACTIVATE_TO_TAIL)
1716 list_add_tail(&page->lru, &n->partial);
1718 list_add(&page->lru, &n->partial);
1721 static inline void add_partial(struct kmem_cache_node *n,
1722 struct page *page, int tail)
1724 lockdep_assert_held(&n->list_lock);
1725 __add_partial(n, page, tail);
1728 static inline void remove_partial(struct kmem_cache_node *n,
1731 lockdep_assert_held(&n->list_lock);
1732 list_del(&page->lru);
1737 * Remove slab from the partial list, freeze it and
1738 * return the pointer to the freelist.
1740 * Returns a list of objects or NULL if it fails.
1742 static inline void *acquire_slab(struct kmem_cache *s,
1743 struct kmem_cache_node *n, struct page *page,
1744 int mode, int *objects)
1747 unsigned long counters;
1750 lockdep_assert_held(&n->list_lock);
1753 * Zap the freelist and set the frozen bit.
1754 * The old freelist is the list of objects for the
1755 * per cpu allocation list.
1757 freelist = page->freelist;
1758 counters = page->counters;
1759 new.counters = counters;
1760 *objects = new.objects - new.inuse;
1762 new.inuse = page->objects;
1763 new.freelist = NULL;
1765 new.freelist = freelist;
1768 VM_BUG_ON(new.frozen);
1771 if (!__cmpxchg_double_slab(s, page,
1773 new.freelist, new.counters,
1777 remove_partial(n, page);
1782 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1783 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1786 * Try to allocate a partial slab from a specific node.
1788 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1789 struct kmem_cache_cpu *c, gfp_t flags)
1791 struct page *page, *page2;
1792 void *object = NULL;
1793 unsigned int available = 0;
1797 * Racy check. If we mistakenly see no partial slabs then we
1798 * just allocate an empty slab. If we mistakenly try to get a
1799 * partial slab and there is none available then get_partials()
1802 if (!n || !n->nr_partial)
1805 spin_lock(&n->list_lock);
1806 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1809 if (!pfmemalloc_match(page, flags))
1812 t = acquire_slab(s, n, page, object == NULL, &objects);
1816 available += objects;
1819 stat(s, ALLOC_FROM_PARTIAL);
1822 put_cpu_partial(s, page, 0);
1823 stat(s, CPU_PARTIAL_NODE);
1825 if (!kmem_cache_has_cpu_partial(s)
1826 || available > slub_cpu_partial(s) / 2)
1830 spin_unlock(&n->list_lock);
1835 * Get a page from somewhere. Search in increasing NUMA distances.
1837 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1838 struct kmem_cache_cpu *c)
1841 struct zonelist *zonelist;
1844 enum zone_type high_zoneidx = gfp_zone(flags);
1846 unsigned int cpuset_mems_cookie;
1849 * The defrag ratio allows a configuration of the tradeoffs between
1850 * inter node defragmentation and node local allocations. A lower
1851 * defrag_ratio increases the tendency to do local allocations
1852 * instead of attempting to obtain partial slabs from other nodes.
1854 * If the defrag_ratio is set to 0 then kmalloc() always
1855 * returns node local objects. If the ratio is higher then kmalloc()
1856 * may return off node objects because partial slabs are obtained
1857 * from other nodes and filled up.
1859 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
1860 * (which makes defrag_ratio = 1000) then every (well almost)
1861 * allocation will first attempt to defrag slab caches on other nodes.
1862 * This means scanning over all nodes to look for partial slabs which
1863 * may be expensive if we do it every time we are trying to find a slab
1864 * with available objects.
1866 if (!s->remote_node_defrag_ratio ||
1867 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1871 cpuset_mems_cookie = read_mems_allowed_begin();
1872 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1873 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1874 struct kmem_cache_node *n;
1876 n = get_node(s, zone_to_nid(zone));
1878 if (n && cpuset_zone_allowed(zone, flags) &&
1879 n->nr_partial > s->min_partial) {
1880 object = get_partial_node(s, n, c, flags);
1883 * Don't check read_mems_allowed_retry()
1884 * here - if mems_allowed was updated in
1885 * parallel, that was a harmless race
1886 * between allocation and the cpuset
1893 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1899 * Get a partial page, lock it and return it.
1901 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1902 struct kmem_cache_cpu *c)
1905 int searchnode = node;
1907 if (node == NUMA_NO_NODE)
1908 searchnode = numa_mem_id();
1909 else if (!node_present_pages(node))
1910 searchnode = node_to_mem_node(node);
1912 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1913 if (object || node != NUMA_NO_NODE)
1916 return get_any_partial(s, flags, c);
1919 #ifdef CONFIG_PREEMPT
1921 * Calculate the next globally unique transaction for disambiguiation
1922 * during cmpxchg. The transactions start with the cpu number and are then
1923 * incremented by CONFIG_NR_CPUS.
1925 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1928 * No preemption supported therefore also no need to check for
1934 static inline unsigned long next_tid(unsigned long tid)
1936 return tid + TID_STEP;
1939 static inline unsigned int tid_to_cpu(unsigned long tid)
1941 return tid % TID_STEP;
1944 static inline unsigned long tid_to_event(unsigned long tid)
1946 return tid / TID_STEP;
1949 static inline unsigned int init_tid(int cpu)
1954 static inline void note_cmpxchg_failure(const char *n,
1955 const struct kmem_cache *s, unsigned long tid)
1957 #ifdef SLUB_DEBUG_CMPXCHG
1958 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1960 pr_info("%s %s: cmpxchg redo ", n, s->name);
1962 #ifdef CONFIG_PREEMPT
1963 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1964 pr_warn("due to cpu change %d -> %d\n",
1965 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1968 if (tid_to_event(tid) != tid_to_event(actual_tid))
1969 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1970 tid_to_event(tid), tid_to_event(actual_tid));
1972 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1973 actual_tid, tid, next_tid(tid));
1975 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1978 static void init_kmem_cache_cpus(struct kmem_cache *s)
1982 for_each_possible_cpu(cpu)
1983 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1987 * Remove the cpu slab
1989 static void deactivate_slab(struct kmem_cache *s, struct page *page,
1990 void *freelist, struct kmem_cache_cpu *c)
1992 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1993 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1995 enum slab_modes l = M_NONE, m = M_NONE;
1997 int tail = DEACTIVATE_TO_HEAD;
2001 if (page->freelist) {
2002 stat(s, DEACTIVATE_REMOTE_FREES);
2003 tail = DEACTIVATE_TO_TAIL;
2007 * Stage one: Free all available per cpu objects back
2008 * to the page freelist while it is still frozen. Leave the
2011 * There is no need to take the list->lock because the page
2014 while (freelist && (nextfree = get_freepointer(s, freelist))) {
2016 unsigned long counters;
2019 prior = page->freelist;
2020 counters = page->counters;
2021 set_freepointer(s, freelist, prior);
2022 new.counters = counters;
2024 VM_BUG_ON(!new.frozen);
2026 } while (!__cmpxchg_double_slab(s, page,
2028 freelist, new.counters,
2029 "drain percpu freelist"));
2031 freelist = nextfree;
2035 * Stage two: Ensure that the page is unfrozen while the
2036 * list presence reflects the actual number of objects
2039 * We setup the list membership and then perform a cmpxchg
2040 * with the count. If there is a mismatch then the page
2041 * is not unfrozen but the page is on the wrong list.
2043 * Then we restart the process which may have to remove
2044 * the page from the list that we just put it on again
2045 * because the number of objects in the slab may have
2050 old.freelist = page->freelist;
2051 old.counters = page->counters;
2052 VM_BUG_ON(!old.frozen);
2054 /* Determine target state of the slab */
2055 new.counters = old.counters;
2058 set_freepointer(s, freelist, old.freelist);
2059 new.freelist = freelist;
2061 new.freelist = old.freelist;
2065 if (!new.inuse && n->nr_partial >= s->min_partial)
2067 else if (new.freelist) {
2072 * Taking the spinlock removes the possiblity
2073 * that acquire_slab() will see a slab page that
2076 spin_lock(&n->list_lock);
2080 if (kmem_cache_debug(s) && !lock) {
2083 * This also ensures that the scanning of full
2084 * slabs from diagnostic functions will not see
2087 spin_lock(&n->list_lock);
2095 remove_partial(n, page);
2097 else if (l == M_FULL)
2099 remove_full(s, n, page);
2101 if (m == M_PARTIAL) {
2103 add_partial(n, page, tail);
2106 } else if (m == M_FULL) {
2108 stat(s, DEACTIVATE_FULL);
2109 add_full(s, n, page);
2115 if (!__cmpxchg_double_slab(s, page,
2116 old.freelist, old.counters,
2117 new.freelist, new.counters,
2122 spin_unlock(&n->list_lock);
2125 stat(s, DEACTIVATE_EMPTY);
2126 discard_slab(s, page);
2135 * Unfreeze all the cpu partial slabs.
2137 * This function must be called with interrupts disabled
2138 * for the cpu using c (or some other guarantee must be there
2139 * to guarantee no concurrent accesses).
2141 static void unfreeze_partials(struct kmem_cache *s,
2142 struct kmem_cache_cpu *c)
2144 #ifdef CONFIG_SLUB_CPU_PARTIAL
2145 struct kmem_cache_node *n = NULL, *n2 = NULL;
2146 struct page *page, *discard_page = NULL;
2148 while ((page = c->partial)) {
2152 c->partial = page->next;
2154 n2 = get_node(s, page_to_nid(page));
2157 spin_unlock(&n->list_lock);
2160 spin_lock(&n->list_lock);
2165 old.freelist = page->freelist;
2166 old.counters = page->counters;
2167 VM_BUG_ON(!old.frozen);
2169 new.counters = old.counters;
2170 new.freelist = old.freelist;
2174 } while (!__cmpxchg_double_slab(s, page,
2175 old.freelist, old.counters,
2176 new.freelist, new.counters,
2177 "unfreezing slab"));
2179 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2180 page->next = discard_page;
2181 discard_page = page;
2183 add_partial(n, page, DEACTIVATE_TO_TAIL);
2184 stat(s, FREE_ADD_PARTIAL);
2189 spin_unlock(&n->list_lock);
2191 while (discard_page) {
2192 page = discard_page;
2193 discard_page = discard_page->next;
2195 stat(s, DEACTIVATE_EMPTY);
2196 discard_slab(s, page);
2203 * Put a page that was just frozen (in __slab_free) into a partial page
2204 * slot if available.
2206 * If we did not find a slot then simply move all the partials to the
2207 * per node partial list.
2209 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2211 #ifdef CONFIG_SLUB_CPU_PARTIAL
2212 struct page *oldpage;
2220 oldpage = this_cpu_read(s->cpu_slab->partial);
2223 pobjects = oldpage->pobjects;
2224 pages = oldpage->pages;
2225 if (drain && pobjects > s->cpu_partial) {
2226 unsigned long flags;
2228 * partial array is full. Move the existing
2229 * set to the per node partial list.
2231 local_irq_save(flags);
2232 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2233 local_irq_restore(flags);
2237 stat(s, CPU_PARTIAL_DRAIN);
2242 pobjects += page->objects - page->inuse;
2244 page->pages = pages;
2245 page->pobjects = pobjects;
2246 page->next = oldpage;
2248 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2250 if (unlikely(!s->cpu_partial)) {
2251 unsigned long flags;
2253 local_irq_save(flags);
2254 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2255 local_irq_restore(flags);
2261 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2263 stat(s, CPUSLAB_FLUSH);
2264 deactivate_slab(s, c->page, c->freelist, c);
2266 c->tid = next_tid(c->tid);
2272 * Called from IPI handler with interrupts disabled.
2274 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2276 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2282 unfreeze_partials(s, c);
2286 static void flush_cpu_slab(void *d)
2288 struct kmem_cache *s = d;
2290 __flush_cpu_slab(s, smp_processor_id());
2293 static bool has_cpu_slab(int cpu, void *info)
2295 struct kmem_cache *s = info;
2296 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2298 return c->page || slub_percpu_partial(c);
2301 static void flush_all(struct kmem_cache *s)
2303 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2307 * Use the cpu notifier to insure that the cpu slabs are flushed when
2310 static int slub_cpu_dead(unsigned int cpu)
2312 struct kmem_cache *s;
2313 unsigned long flags;
2315 mutex_lock(&slab_mutex);
2316 list_for_each_entry(s, &slab_caches, list) {
2317 local_irq_save(flags);
2318 __flush_cpu_slab(s, cpu);
2319 local_irq_restore(flags);
2321 mutex_unlock(&slab_mutex);
2326 * Check if the objects in a per cpu structure fit numa
2327 * locality expectations.
2329 static inline int node_match(struct page *page, int node)
2332 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2338 #ifdef CONFIG_SLUB_DEBUG
2339 static int count_free(struct page *page)
2341 return page->objects - page->inuse;
2344 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2346 return atomic_long_read(&n->total_objects);
2348 #endif /* CONFIG_SLUB_DEBUG */
2350 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2351 static unsigned long count_partial(struct kmem_cache_node *n,
2352 int (*get_count)(struct page *))
2354 unsigned long flags;
2355 unsigned long x = 0;
2358 spin_lock_irqsave(&n->list_lock, flags);
2359 list_for_each_entry(page, &n->partial, lru)
2360 x += get_count(page);
2361 spin_unlock_irqrestore(&n->list_lock, flags);
2364 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2366 static noinline void
2367 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2369 #ifdef CONFIG_SLUB_DEBUG
2370 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2371 DEFAULT_RATELIMIT_BURST);
2373 struct kmem_cache_node *n;
2375 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2378 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2379 nid, gfpflags, &gfpflags);
2380 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2381 s->name, s->object_size, s->size, oo_order(s->oo),
2384 if (oo_order(s->min) > get_order(s->object_size))
2385 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2388 for_each_kmem_cache_node(s, node, n) {
2389 unsigned long nr_slabs;
2390 unsigned long nr_objs;
2391 unsigned long nr_free;
2393 nr_free = count_partial(n, count_free);
2394 nr_slabs = node_nr_slabs(n);
2395 nr_objs = node_nr_objs(n);
2397 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2398 node, nr_slabs, nr_objs, nr_free);
2403 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2404 int node, struct kmem_cache_cpu **pc)
2407 struct kmem_cache_cpu *c = *pc;
2410 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
2412 freelist = get_partial(s, flags, node, c);
2417 page = new_slab(s, flags, node);
2419 c = raw_cpu_ptr(s->cpu_slab);
2424 * No other reference to the page yet so we can
2425 * muck around with it freely without cmpxchg
2427 freelist = page->freelist;
2428 page->freelist = NULL;
2430 stat(s, ALLOC_SLAB);
2439 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2441 if (unlikely(PageSlabPfmemalloc(page)))
2442 return gfp_pfmemalloc_allowed(gfpflags);
2448 * Check the page->freelist of a page and either transfer the freelist to the
2449 * per cpu freelist or deactivate the page.
2451 * The page is still frozen if the return value is not NULL.
2453 * If this function returns NULL then the page has been unfrozen.
2455 * This function must be called with interrupt disabled.
2457 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2460 unsigned long counters;
2464 freelist = page->freelist;
2465 counters = page->counters;
2467 new.counters = counters;
2468 VM_BUG_ON(!new.frozen);
2470 new.inuse = page->objects;
2471 new.frozen = freelist != NULL;
2473 } while (!__cmpxchg_double_slab(s, page,
2482 * Slow path. The lockless freelist is empty or we need to perform
2485 * Processing is still very fast if new objects have been freed to the
2486 * regular freelist. In that case we simply take over the regular freelist
2487 * as the lockless freelist and zap the regular freelist.
2489 * If that is not working then we fall back to the partial lists. We take the
2490 * first element of the freelist as the object to allocate now and move the
2491 * rest of the freelist to the lockless freelist.
2493 * And if we were unable to get a new slab from the partial slab lists then
2494 * we need to allocate a new slab. This is the slowest path since it involves
2495 * a call to the page allocator and the setup of a new slab.
2497 * Version of __slab_alloc to use when we know that interrupts are
2498 * already disabled (which is the case for bulk allocation).
2500 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2501 unsigned long addr, struct kmem_cache_cpu *c)
2511 if (unlikely(!node_match(page, node))) {
2512 int searchnode = node;
2514 if (node != NUMA_NO_NODE && !node_present_pages(node))
2515 searchnode = node_to_mem_node(node);
2517 if (unlikely(!node_match(page, searchnode))) {
2518 stat(s, ALLOC_NODE_MISMATCH);
2519 deactivate_slab(s, page, c->freelist, c);
2525 * By rights, we should be searching for a slab page that was
2526 * PFMEMALLOC but right now, we are losing the pfmemalloc
2527 * information when the page leaves the per-cpu allocator
2529 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2530 deactivate_slab(s, page, c->freelist, c);
2534 /* must check again c->freelist in case of cpu migration or IRQ */
2535 freelist = c->freelist;
2539 freelist = get_freelist(s, page);
2543 stat(s, DEACTIVATE_BYPASS);
2547 stat(s, ALLOC_REFILL);
2551 * freelist is pointing to the list of objects to be used.
2552 * page is pointing to the page from which the objects are obtained.
2553 * That page must be frozen for per cpu allocations to work.
2555 VM_BUG_ON(!c->page->frozen);
2556 c->freelist = get_freepointer(s, freelist);
2557 c->tid = next_tid(c->tid);
2562 if (slub_percpu_partial(c)) {
2563 page = c->page = slub_percpu_partial(c);
2564 slub_set_percpu_partial(c, page);
2565 stat(s, CPU_PARTIAL_ALLOC);
2569 freelist = new_slab_objects(s, gfpflags, node, &c);
2571 if (unlikely(!freelist)) {
2572 slab_out_of_memory(s, gfpflags, node);
2577 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2580 /* Only entered in the debug case */
2581 if (kmem_cache_debug(s) &&
2582 !alloc_debug_processing(s, page, freelist, addr))
2583 goto new_slab; /* Slab failed checks. Next slab needed */
2585 deactivate_slab(s, page, get_freepointer(s, freelist), c);
2590 * Another one that disabled interrupt and compensates for possible
2591 * cpu changes by refetching the per cpu area pointer.
2593 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2594 unsigned long addr, struct kmem_cache_cpu *c)
2597 unsigned long flags;
2599 local_irq_save(flags);
2600 #ifdef CONFIG_PREEMPT
2602 * We may have been preempted and rescheduled on a different
2603 * cpu before disabling interrupts. Need to reload cpu area
2606 c = this_cpu_ptr(s->cpu_slab);
2609 p = ___slab_alloc(s, gfpflags, node, addr, c);
2610 local_irq_restore(flags);
2615 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2616 * have the fastpath folded into their functions. So no function call
2617 * overhead for requests that can be satisfied on the fastpath.
2619 * The fastpath works by first checking if the lockless freelist can be used.
2620 * If not then __slab_alloc is called for slow processing.
2622 * Otherwise we can simply pick the next object from the lockless free list.
2624 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2625 gfp_t gfpflags, int node, unsigned long addr)
2628 struct kmem_cache_cpu *c;
2632 s = slab_pre_alloc_hook(s, gfpflags);
2637 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2638 * enabled. We may switch back and forth between cpus while
2639 * reading from one cpu area. That does not matter as long
2640 * as we end up on the original cpu again when doing the cmpxchg.
2642 * We should guarantee that tid and kmem_cache are retrieved on
2643 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2644 * to check if it is matched or not.
2647 tid = this_cpu_read(s->cpu_slab->tid);
2648 c = raw_cpu_ptr(s->cpu_slab);
2649 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2650 unlikely(tid != READ_ONCE(c->tid)));
2653 * Irqless object alloc/free algorithm used here depends on sequence
2654 * of fetching cpu_slab's data. tid should be fetched before anything
2655 * on c to guarantee that object and page associated with previous tid
2656 * won't be used with current tid. If we fetch tid first, object and
2657 * page could be one associated with next tid and our alloc/free
2658 * request will be failed. In this case, we will retry. So, no problem.
2663 * The transaction ids are globally unique per cpu and per operation on
2664 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2665 * occurs on the right processor and that there was no operation on the
2666 * linked list in between.
2669 object = c->freelist;
2671 if (unlikely(!object || !node_match(page, node))) {
2672 object = __slab_alloc(s, gfpflags, node, addr, c);
2673 stat(s, ALLOC_SLOWPATH);
2675 void *next_object = get_freepointer_safe(s, object);
2678 * The cmpxchg will only match if there was no additional
2679 * operation and if we are on the right processor.
2681 * The cmpxchg does the following atomically (without lock
2683 * 1. Relocate first pointer to the current per cpu area.
2684 * 2. Verify that tid and freelist have not been changed
2685 * 3. If they were not changed replace tid and freelist
2687 * Since this is without lock semantics the protection is only
2688 * against code executing on this cpu *not* from access by
2691 if (unlikely(!this_cpu_cmpxchg_double(
2692 s->cpu_slab->freelist, s->cpu_slab->tid,
2694 next_object, next_tid(tid)))) {
2696 note_cmpxchg_failure("slab_alloc", s, tid);
2699 prefetch_freepointer(s, next_object);
2700 stat(s, ALLOC_FASTPATH);
2703 if (unlikely(gfpflags & __GFP_ZERO) && object)
2704 memset(object, 0, s->object_size);
2706 slab_post_alloc_hook(s, gfpflags, 1, &object);
2711 static __always_inline void *slab_alloc(struct kmem_cache *s,
2712 gfp_t gfpflags, unsigned long addr)
2714 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2717 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2719 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2721 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2726 EXPORT_SYMBOL(kmem_cache_alloc);
2728 #ifdef CONFIG_TRACING
2729 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2731 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2732 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2733 kasan_kmalloc(s, ret, size, gfpflags);
2736 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2740 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2742 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2744 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2745 s->object_size, s->size, gfpflags, node);
2749 EXPORT_SYMBOL(kmem_cache_alloc_node);
2751 #ifdef CONFIG_TRACING
2752 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2754 int node, size_t size)
2756 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2758 trace_kmalloc_node(_RET_IP_, ret,
2759 size, s->size, gfpflags, node);
2761 kasan_kmalloc(s, ret, size, gfpflags);
2764 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2769 * Slow path handling. This may still be called frequently since objects
2770 * have a longer lifetime than the cpu slabs in most processing loads.
2772 * So we still attempt to reduce cache line usage. Just take the slab
2773 * lock and free the item. If there is no additional partial page
2774 * handling required then we can return immediately.
2776 static void __slab_free(struct kmem_cache *s, struct page *page,
2777 void *head, void *tail, int cnt,
2784 unsigned long counters;
2785 struct kmem_cache_node *n = NULL;
2786 unsigned long uninitialized_var(flags);
2788 stat(s, FREE_SLOWPATH);
2790 if (kmem_cache_debug(s) &&
2791 !free_debug_processing(s, page, head, tail, cnt, addr))
2796 spin_unlock_irqrestore(&n->list_lock, flags);
2799 prior = page->freelist;
2800 counters = page->counters;
2801 set_freepointer(s, tail, prior);
2802 new.counters = counters;
2803 was_frozen = new.frozen;
2805 if ((!new.inuse || !prior) && !was_frozen) {
2807 if (kmem_cache_has_cpu_partial(s) && !prior) {
2810 * Slab was on no list before and will be
2812 * We can defer the list move and instead
2817 } else { /* Needs to be taken off a list */
2819 n = get_node(s, page_to_nid(page));
2821 * Speculatively acquire the list_lock.
2822 * If the cmpxchg does not succeed then we may
2823 * drop the list_lock without any processing.
2825 * Otherwise the list_lock will synchronize with
2826 * other processors updating the list of slabs.
2828 spin_lock_irqsave(&n->list_lock, flags);
2833 } while (!cmpxchg_double_slab(s, page,
2841 * If we just froze the page then put it onto the
2842 * per cpu partial list.
2844 if (new.frozen && !was_frozen) {
2845 put_cpu_partial(s, page, 1);
2846 stat(s, CPU_PARTIAL_FREE);
2849 * The list lock was not taken therefore no list
2850 * activity can be necessary.
2853 stat(s, FREE_FROZEN);
2857 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2861 * Objects left in the slab. If it was not on the partial list before
2864 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2865 if (kmem_cache_debug(s))
2866 remove_full(s, n, page);
2867 add_partial(n, page, DEACTIVATE_TO_TAIL);
2868 stat(s, FREE_ADD_PARTIAL);
2870 spin_unlock_irqrestore(&n->list_lock, flags);
2876 * Slab on the partial list.
2878 remove_partial(n, page);
2879 stat(s, FREE_REMOVE_PARTIAL);
2881 /* Slab must be on the full list */
2882 remove_full(s, n, page);
2885 spin_unlock_irqrestore(&n->list_lock, flags);
2887 discard_slab(s, page);
2891 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2892 * can perform fastpath freeing without additional function calls.
2894 * The fastpath is only possible if we are freeing to the current cpu slab
2895 * of this processor. This typically the case if we have just allocated
2898 * If fastpath is not possible then fall back to __slab_free where we deal
2899 * with all sorts of special processing.
2901 * Bulk free of a freelist with several objects (all pointing to the
2902 * same page) possible by specifying head and tail ptr, plus objects
2903 * count (cnt). Bulk free indicated by tail pointer being set.
2905 static __always_inline void do_slab_free(struct kmem_cache *s,
2906 struct page *page, void *head, void *tail,
2907 int cnt, unsigned long addr)
2909 void *tail_obj = tail ? : head;
2910 struct kmem_cache_cpu *c;
2914 * Determine the currently cpus per cpu slab.
2915 * The cpu may change afterward. However that does not matter since
2916 * data is retrieved via this pointer. If we are on the same cpu
2917 * during the cmpxchg then the free will succeed.
2920 tid = this_cpu_read(s->cpu_slab->tid);
2921 c = raw_cpu_ptr(s->cpu_slab);
2922 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2923 unlikely(tid != READ_ONCE(c->tid)));
2925 /* Same with comment on barrier() in slab_alloc_node() */
2928 if (likely(page == c->page)) {
2929 set_freepointer(s, tail_obj, c->freelist);
2931 if (unlikely(!this_cpu_cmpxchg_double(
2932 s->cpu_slab->freelist, s->cpu_slab->tid,
2934 head, next_tid(tid)))) {
2936 note_cmpxchg_failure("slab_free", s, tid);
2939 stat(s, FREE_FASTPATH);
2941 __slab_free(s, page, head, tail_obj, cnt, addr);
2945 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
2946 void *head, void *tail, int cnt,
2950 * With KASAN enabled slab_free_freelist_hook modifies the freelist
2951 * to remove objects, whose reuse must be delayed.
2953 if (slab_free_freelist_hook(s, &head, &tail))
2954 do_slab_free(s, page, head, tail, cnt, addr);
2958 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
2960 do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr);
2964 void kmem_cache_free(struct kmem_cache *s, void *x)
2966 s = cache_from_obj(s, x);
2969 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
2970 trace_kmem_cache_free(_RET_IP_, x);
2972 EXPORT_SYMBOL(kmem_cache_free);
2974 struct detached_freelist {
2979 struct kmem_cache *s;
2983 * This function progressively scans the array with free objects (with
2984 * a limited look ahead) and extract objects belonging to the same
2985 * page. It builds a detached freelist directly within the given
2986 * page/objects. This can happen without any need for
2987 * synchronization, because the objects are owned by running process.
2988 * The freelist is build up as a single linked list in the objects.
2989 * The idea is, that this detached freelist can then be bulk
2990 * transferred to the real freelist(s), but only requiring a single
2991 * synchronization primitive. Look ahead in the array is limited due
2992 * to performance reasons.
2995 int build_detached_freelist(struct kmem_cache *s, size_t size,
2996 void **p, struct detached_freelist *df)
2998 size_t first_skipped_index = 0;
3003 /* Always re-init detached_freelist */
3008 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3009 } while (!object && size);
3014 page = virt_to_head_page(object);
3016 /* Handle kalloc'ed objects */
3017 if (unlikely(!PageSlab(page))) {
3018 BUG_ON(!PageCompound(page));
3020 __free_pages(page, compound_order(page));
3021 p[size] = NULL; /* mark object processed */
3024 /* Derive kmem_cache from object */
3025 df->s = page->slab_cache;
3027 df->s = cache_from_obj(s, object); /* Support for memcg */
3030 /* Start new detached freelist */
3032 set_freepointer(df->s, object, NULL);
3034 df->freelist = object;
3035 p[size] = NULL; /* mark object processed */
3041 continue; /* Skip processed objects */
3043 /* df->page is always set at this point */
3044 if (df->page == virt_to_head_page(object)) {
3045 /* Opportunity build freelist */
3046 set_freepointer(df->s, object, df->freelist);
3047 df->freelist = object;
3049 p[size] = NULL; /* mark object processed */
3054 /* Limit look ahead search */
3058 if (!first_skipped_index)
3059 first_skipped_index = size + 1;
3062 return first_skipped_index;
3065 /* Note that interrupts must be enabled when calling this function. */
3066 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3072 struct detached_freelist df;
3074 size = build_detached_freelist(s, size, p, &df);
3078 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
3079 } while (likely(size));
3081 EXPORT_SYMBOL(kmem_cache_free_bulk);
3083 /* Note that interrupts must be enabled when calling this function. */
3084 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3087 struct kmem_cache_cpu *c;
3090 /* memcg and kmem_cache debug support */
3091 s = slab_pre_alloc_hook(s, flags);
3095 * Drain objects in the per cpu slab, while disabling local
3096 * IRQs, which protects against PREEMPT and interrupts
3097 * handlers invoking normal fastpath.
3099 local_irq_disable();
3100 c = this_cpu_ptr(s->cpu_slab);
3102 for (i = 0; i < size; i++) {
3103 void *object = c->freelist;
3105 if (unlikely(!object)) {
3107 * Invoking slow path likely have side-effect
3108 * of re-populating per CPU c->freelist
3110 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3112 if (unlikely(!p[i]))
3115 c = this_cpu_ptr(s->cpu_slab);
3116 continue; /* goto for-loop */
3118 c->freelist = get_freepointer(s, object);
3121 c->tid = next_tid(c->tid);
3124 /* Clear memory outside IRQ disabled fastpath loop */
3125 if (unlikely(flags & __GFP_ZERO)) {
3128 for (j = 0; j < i; j++)
3129 memset(p[j], 0, s->object_size);
3132 /* memcg and kmem_cache debug support */
3133 slab_post_alloc_hook(s, flags, size, p);
3137 slab_post_alloc_hook(s, flags, i, p);
3138 __kmem_cache_free_bulk(s, i, p);
3141 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3145 * Object placement in a slab is made very easy because we always start at
3146 * offset 0. If we tune the size of the object to the alignment then we can
3147 * get the required alignment by putting one properly sized object after
3150 * Notice that the allocation order determines the sizes of the per cpu
3151 * caches. Each processor has always one slab available for allocations.
3152 * Increasing the allocation order reduces the number of times that slabs
3153 * must be moved on and off the partial lists and is therefore a factor in
3158 * Mininum / Maximum order of slab pages. This influences locking overhead
3159 * and slab fragmentation. A higher order reduces the number of partial slabs
3160 * and increases the number of allocations possible without having to
3161 * take the list_lock.
3163 static unsigned int slub_min_order;
3164 static unsigned int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3165 static unsigned int slub_min_objects;
3168 * Calculate the order of allocation given an slab object size.
3170 * The order of allocation has significant impact on performance and other
3171 * system components. Generally order 0 allocations should be preferred since
3172 * order 0 does not cause fragmentation in the page allocator. Larger objects
3173 * be problematic to put into order 0 slabs because there may be too much
3174 * unused space left. We go to a higher order if more than 1/16th of the slab
3177 * In order to reach satisfactory performance we must ensure that a minimum
3178 * number of objects is in one slab. Otherwise we may generate too much
3179 * activity on the partial lists which requires taking the list_lock. This is
3180 * less a concern for large slabs though which are rarely used.
3182 * slub_max_order specifies the order where we begin to stop considering the
3183 * number of objects in a slab as critical. If we reach slub_max_order then
3184 * we try to keep the page order as low as possible. So we accept more waste
3185 * of space in favor of a small page order.
3187 * Higher order allocations also allow the placement of more objects in a
3188 * slab and thereby reduce object handling overhead. If the user has
3189 * requested a higher mininum order then we start with that one instead of
3190 * the smallest order which will fit the object.
3192 static inline unsigned int slab_order(unsigned int size,
3193 unsigned int min_objects, unsigned int max_order,
3194 unsigned int fract_leftover)
3196 unsigned int min_order = slub_min_order;
3199 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
3200 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3202 for (order = max(min_order, (unsigned int)get_order(min_objects * size));
3203 order <= max_order; order++) {
3205 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
3208 rem = slab_size % size;
3210 if (rem <= slab_size / fract_leftover)
3217 static inline int calculate_order(unsigned int size)
3220 unsigned int min_objects;
3221 unsigned int max_objects;
3224 * Attempt to find best configuration for a slab. This
3225 * works by first attempting to generate a layout with
3226 * the best configuration and backing off gradually.
3228 * First we increase the acceptable waste in a slab. Then
3229 * we reduce the minimum objects required in a slab.
3231 min_objects = slub_min_objects;
3233 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3234 max_objects = order_objects(slub_max_order, size);
3235 min_objects = min(min_objects, max_objects);
3237 while (min_objects > 1) {
3238 unsigned int fraction;
3241 while (fraction >= 4) {
3242 order = slab_order(size, min_objects,
3243 slub_max_order, fraction);
3244 if (order <= slub_max_order)
3252 * We were unable to place multiple objects in a slab. Now
3253 * lets see if we can place a single object there.
3255 order = slab_order(size, 1, slub_max_order, 1);
3256 if (order <= slub_max_order)
3260 * Doh this slab cannot be placed using slub_max_order.
3262 order = slab_order(size, 1, MAX_ORDER, 1);
3263 if (order < MAX_ORDER)
3269 init_kmem_cache_node(struct kmem_cache_node *n)
3272 spin_lock_init(&n->list_lock);
3273 INIT_LIST_HEAD(&n->partial);
3274 #ifdef CONFIG_SLUB_DEBUG
3275 atomic_long_set(&n->nr_slabs, 0);
3276 atomic_long_set(&n->total_objects, 0);
3277 INIT_LIST_HEAD(&n->full);
3281 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3283 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3284 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3287 * Must align to double word boundary for the double cmpxchg
3288 * instructions to work; see __pcpu_double_call_return_bool().
3290 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3291 2 * sizeof(void *));
3296 init_kmem_cache_cpus(s);
3301 static struct kmem_cache *kmem_cache_node;
3304 * No kmalloc_node yet so do it by hand. We know that this is the first
3305 * slab on the node for this slabcache. There are no concurrent accesses
3308 * Note that this function only works on the kmem_cache_node
3309 * when allocating for the kmem_cache_node. This is used for bootstrapping
3310 * memory on a fresh node that has no slab structures yet.
3312 static void early_kmem_cache_node_alloc(int node)
3315 struct kmem_cache_node *n;
3317 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3319 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3322 if (page_to_nid(page) != node) {
3323 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3324 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3329 page->freelist = get_freepointer(kmem_cache_node, n);
3332 kmem_cache_node->node[node] = n;
3333 #ifdef CONFIG_SLUB_DEBUG
3334 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3335 init_tracking(kmem_cache_node, n);
3337 kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node),
3339 init_kmem_cache_node(n);
3340 inc_slabs_node(kmem_cache_node, node, page->objects);
3343 * No locks need to be taken here as it has just been
3344 * initialized and there is no concurrent access.
3346 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3349 static void free_kmem_cache_nodes(struct kmem_cache *s)
3352 struct kmem_cache_node *n;
3354 for_each_kmem_cache_node(s, node, n) {
3355 s->node[node] = NULL;
3356 kmem_cache_free(kmem_cache_node, n);
3360 void __kmem_cache_release(struct kmem_cache *s)
3362 cache_random_seq_destroy(s);
3363 free_percpu(s->cpu_slab);
3364 free_kmem_cache_nodes(s);
3367 static int init_kmem_cache_nodes(struct kmem_cache *s)
3371 for_each_node_state(node, N_NORMAL_MEMORY) {
3372 struct kmem_cache_node *n;
3374 if (slab_state == DOWN) {
3375 early_kmem_cache_node_alloc(node);
3378 n = kmem_cache_alloc_node(kmem_cache_node,
3382 free_kmem_cache_nodes(s);
3386 init_kmem_cache_node(n);
3392 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3394 if (min < MIN_PARTIAL)
3396 else if (min > MAX_PARTIAL)
3398 s->min_partial = min;
3401 static void set_cpu_partial(struct kmem_cache *s)
3403 #ifdef CONFIG_SLUB_CPU_PARTIAL
3405 * cpu_partial determined the maximum number of objects kept in the
3406 * per cpu partial lists of a processor.
3408 * Per cpu partial lists mainly contain slabs that just have one
3409 * object freed. If they are used for allocation then they can be
3410 * filled up again with minimal effort. The slab will never hit the
3411 * per node partial lists and therefore no locking will be required.
3413 * This setting also determines
3415 * A) The number of objects from per cpu partial slabs dumped to the
3416 * per node list when we reach the limit.
3417 * B) The number of objects in cpu partial slabs to extract from the
3418 * per node list when we run out of per cpu objects. We only fetch
3419 * 50% to keep some capacity around for frees.
3421 if (!kmem_cache_has_cpu_partial(s))
3423 else if (s->size >= PAGE_SIZE)
3425 else if (s->size >= 1024)
3427 else if (s->size >= 256)
3428 s->cpu_partial = 13;
3430 s->cpu_partial = 30;
3435 * calculate_sizes() determines the order and the distribution of data within
3438 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3440 slab_flags_t flags = s->flags;
3441 unsigned int size = s->object_size;
3445 * Round up object size to the next word boundary. We can only
3446 * place the free pointer at word boundaries and this determines
3447 * the possible location of the free pointer.
3449 size = ALIGN(size, sizeof(void *));
3451 #ifdef CONFIG_SLUB_DEBUG
3453 * Determine if we can poison the object itself. If the user of
3454 * the slab may touch the object after free or before allocation
3455 * then we should never poison the object itself.
3457 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
3459 s->flags |= __OBJECT_POISON;
3461 s->flags &= ~__OBJECT_POISON;
3465 * If we are Redzoning then check if there is some space between the
3466 * end of the object and the free pointer. If not then add an
3467 * additional word to have some bytes to store Redzone information.
3469 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3470 size += sizeof(void *);
3474 * With that we have determined the number of bytes in actual use
3475 * by the object. This is the potential offset to the free pointer.
3479 if (((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
3482 * Relocate free pointer after the object if it is not
3483 * permitted to overwrite the first word of the object on
3486 * This is the case if we do RCU, have a constructor or
3487 * destructor or are poisoning the objects.
3490 size += sizeof(void *);
3493 #ifdef CONFIG_SLUB_DEBUG
3494 if (flags & SLAB_STORE_USER)
3496 * Need to store information about allocs and frees after
3499 size += 2 * sizeof(struct track);
3502 kasan_cache_create(s, &size, &s->flags);
3503 #ifdef CONFIG_SLUB_DEBUG
3504 if (flags & SLAB_RED_ZONE) {
3506 * Add some empty padding so that we can catch
3507 * overwrites from earlier objects rather than let
3508 * tracking information or the free pointer be
3509 * corrupted if a user writes before the start
3512 size += sizeof(void *);
3514 s->red_left_pad = sizeof(void *);
3515 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3516 size += s->red_left_pad;
3521 * SLUB stores one object immediately after another beginning from
3522 * offset 0. In order to align the objects we have to simply size
3523 * each object to conform to the alignment.
3525 size = ALIGN(size, s->align);
3527 if (forced_order >= 0)
3528 order = forced_order;
3530 order = calculate_order(size);
3537 s->allocflags |= __GFP_COMP;
3539 if (s->flags & SLAB_CACHE_DMA)
3540 s->allocflags |= GFP_DMA;
3542 if (s->flags & SLAB_CACHE_DMA32)
3543 s->allocflags |= GFP_DMA32;
3545 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3546 s->allocflags |= __GFP_RECLAIMABLE;
3549 * Determine the number of objects per slab
3551 s->oo = oo_make(order, size);
3552 s->min = oo_make(get_order(size), size);
3553 if (oo_objects(s->oo) > oo_objects(s->max))
3556 return !!oo_objects(s->oo);
3559 static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
3561 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3562 #ifdef CONFIG_SLAB_FREELIST_HARDENED
3563 s->random = get_random_long();
3566 if (!calculate_sizes(s, -1))
3568 if (disable_higher_order_debug) {
3570 * Disable debugging flags that store metadata if the min slab
3573 if (get_order(s->size) > get_order(s->object_size)) {
3574 s->flags &= ~DEBUG_METADATA_FLAGS;
3576 if (!calculate_sizes(s, -1))
3581 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3582 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3583 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
3584 /* Enable fast mode */
3585 s->flags |= __CMPXCHG_DOUBLE;
3589 * The larger the object size is, the more pages we want on the partial
3590 * list to avoid pounding the page allocator excessively.
3592 set_min_partial(s, ilog2(s->size) / 2);
3597 s->remote_node_defrag_ratio = 1000;
3600 /* Initialize the pre-computed randomized freelist if slab is up */
3601 if (slab_state >= UP) {
3602 if (init_cache_random_seq(s))
3606 if (!init_kmem_cache_nodes(s))
3609 if (alloc_kmem_cache_cpus(s))
3612 free_kmem_cache_nodes(s);
3614 if (flags & SLAB_PANIC)
3615 panic("Cannot create slab %s size=%u realsize=%u order=%u offset=%u flags=%lx\n",
3616 s->name, s->size, s->size,
3617 oo_order(s->oo), s->offset, (unsigned long)flags);
3621 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3624 #ifdef CONFIG_SLUB_DEBUG
3625 void *addr = page_address(page);
3627 unsigned long *map = kcalloc(BITS_TO_LONGS(page->objects),
3632 slab_err(s, page, text, s->name);
3635 get_map(s, page, map);
3636 for_each_object(p, s, addr, page->objects) {
3638 if (!test_bit(slab_index(p, s, addr), map)) {
3639 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3640 print_tracking(s, p);
3649 * Attempt to free all partial slabs on a node.
3650 * This is called from __kmem_cache_shutdown(). We must take list_lock
3651 * because sysfs file might still access partial list after the shutdowning.
3653 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3656 struct page *page, *h;
3658 BUG_ON(irqs_disabled());
3659 spin_lock_irq(&n->list_lock);
3660 list_for_each_entry_safe(page, h, &n->partial, lru) {
3662 remove_partial(n, page);
3663 list_add(&page->lru, &discard);
3665 list_slab_objects(s, page,
3666 "Objects remaining in %s on __kmem_cache_shutdown()");
3669 spin_unlock_irq(&n->list_lock);
3671 list_for_each_entry_safe(page, h, &discard, lru)
3672 discard_slab(s, page);
3675 bool __kmem_cache_empty(struct kmem_cache *s)
3678 struct kmem_cache_node *n;
3680 for_each_kmem_cache_node(s, node, n)
3681 if (n->nr_partial || slabs_node(s, node))
3687 * Release all resources used by a slab cache.
3689 int __kmem_cache_shutdown(struct kmem_cache *s)
3692 struct kmem_cache_node *n;
3695 /* Attempt to free all objects */
3696 for_each_kmem_cache_node(s, node, n) {
3698 if (n->nr_partial || slabs_node(s, node))
3701 sysfs_slab_remove(s);
3705 /********************************************************************
3707 *******************************************************************/
3709 static int __init setup_slub_min_order(char *str)
3711 get_option(&str, (int *)&slub_min_order);
3716 __setup("slub_min_order=", setup_slub_min_order);
3718 static int __init setup_slub_max_order(char *str)
3720 get_option(&str, (int *)&slub_max_order);
3721 slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1);
3726 __setup("slub_max_order=", setup_slub_max_order);
3728 static int __init setup_slub_min_objects(char *str)
3730 get_option(&str, (int *)&slub_min_objects);
3735 __setup("slub_min_objects=", setup_slub_min_objects);
3737 void *__kmalloc(size_t size, gfp_t flags)
3739 struct kmem_cache *s;
3742 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3743 return kmalloc_large(size, flags);
3745 s = kmalloc_slab(size, flags);
3747 if (unlikely(ZERO_OR_NULL_PTR(s)))
3750 ret = slab_alloc(s, flags, _RET_IP_);
3752 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3754 kasan_kmalloc(s, ret, size, flags);
3758 EXPORT_SYMBOL(__kmalloc);
3761 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3766 flags |= __GFP_COMP;
3767 page = alloc_pages_node(node, flags, get_order(size));
3769 ptr = page_address(page);
3771 kmalloc_large_node_hook(ptr, size, flags);
3775 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3777 struct kmem_cache *s;
3780 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3781 ret = kmalloc_large_node(size, flags, node);
3783 trace_kmalloc_node(_RET_IP_, ret,
3784 size, PAGE_SIZE << get_order(size),
3790 s = kmalloc_slab(size, flags);
3792 if (unlikely(ZERO_OR_NULL_PTR(s)))
3795 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3797 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3799 kasan_kmalloc(s, ret, size, flags);
3803 EXPORT_SYMBOL(__kmalloc_node);
3806 #ifdef CONFIG_HARDENED_USERCOPY
3808 * Rejects incorrectly sized objects and objects that are to be copied
3809 * to/from userspace but do not fall entirely within the containing slab
3810 * cache's usercopy region.
3812 * Returns NULL if check passes, otherwise const char * to name of cache
3813 * to indicate an error.
3815 void __check_heap_object(const void *ptr, unsigned long n, struct page *page,
3818 struct kmem_cache *s;
3819 unsigned int offset;
3822 /* Find object and usable object size. */
3823 s = page->slab_cache;
3825 /* Reject impossible pointers. */
3826 if (ptr < page_address(page))
3827 usercopy_abort("SLUB object not in SLUB page?!", NULL,
3830 /* Find offset within object. */
3831 offset = (ptr - page_address(page)) % s->size;
3833 /* Adjust for redzone and reject if within the redzone. */
3834 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE) {
3835 if (offset < s->red_left_pad)
3836 usercopy_abort("SLUB object in left red zone",
3837 s->name, to_user, offset, n);
3838 offset -= s->red_left_pad;
3841 /* Allow address range falling entirely within usercopy region. */
3842 if (offset >= s->useroffset &&
3843 offset - s->useroffset <= s->usersize &&
3844 n <= s->useroffset - offset + s->usersize)
3848 * If the copy is still within the allocated object, produce
3849 * a warning instead of rejecting the copy. This is intended
3850 * to be a temporary method to find any missing usercopy
3853 object_size = slab_ksize(s);
3854 if (usercopy_fallback &&
3855 offset <= object_size && n <= object_size - offset) {
3856 usercopy_warn("SLUB object", s->name, to_user, offset, n);
3860 usercopy_abort("SLUB object", s->name, to_user, offset, n);
3862 #endif /* CONFIG_HARDENED_USERCOPY */
3864 static size_t __ksize(const void *object)
3868 if (unlikely(object == ZERO_SIZE_PTR))
3871 page = virt_to_head_page(object);
3873 if (unlikely(!PageSlab(page))) {
3874 WARN_ON(!PageCompound(page));
3875 return PAGE_SIZE << compound_order(page);
3878 return slab_ksize(page->slab_cache);
3881 size_t ksize(const void *object)
3883 size_t size = __ksize(object);
3884 /* We assume that ksize callers could use whole allocated area,
3885 * so we need to unpoison this area.
3887 kasan_unpoison_shadow(object, size);
3890 EXPORT_SYMBOL(ksize);
3892 void kfree(const void *x)
3895 void *object = (void *)x;
3897 trace_kfree(_RET_IP_, x);
3899 if (unlikely(ZERO_OR_NULL_PTR(x)))
3902 page = virt_to_head_page(x);
3903 if (unlikely(!PageSlab(page))) {
3904 BUG_ON(!PageCompound(page));
3906 __free_pages(page, compound_order(page));
3909 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
3911 EXPORT_SYMBOL(kfree);
3913 #define SHRINK_PROMOTE_MAX 32
3916 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3917 * up most to the head of the partial lists. New allocations will then
3918 * fill those up and thus they can be removed from the partial lists.
3920 * The slabs with the least items are placed last. This results in them
3921 * being allocated from last increasing the chance that the last objects
3922 * are freed in them.
3924 int __kmem_cache_shrink(struct kmem_cache *s)
3928 struct kmem_cache_node *n;
3931 struct list_head discard;
3932 struct list_head promote[SHRINK_PROMOTE_MAX];
3933 unsigned long flags;
3937 for_each_kmem_cache_node(s, node, n) {
3938 INIT_LIST_HEAD(&discard);
3939 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3940 INIT_LIST_HEAD(promote + i);
3942 spin_lock_irqsave(&n->list_lock, flags);
3945 * Build lists of slabs to discard or promote.
3947 * Note that concurrent frees may occur while we hold the
3948 * list_lock. page->inuse here is the upper limit.
3950 list_for_each_entry_safe(page, t, &n->partial, lru) {
3951 int free = page->objects - page->inuse;
3953 /* Do not reread page->inuse */
3956 /* We do not keep full slabs on the list */
3959 if (free == page->objects) {
3960 list_move(&page->lru, &discard);
3962 } else if (free <= SHRINK_PROMOTE_MAX)
3963 list_move(&page->lru, promote + free - 1);
3967 * Promote the slabs filled up most to the head of the
3970 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
3971 list_splice(promote + i, &n->partial);
3973 spin_unlock_irqrestore(&n->list_lock, flags);
3975 /* Release empty slabs */
3976 list_for_each_entry_safe(page, t, &discard, lru)
3977 discard_slab(s, page);
3979 if (slabs_node(s, node))
3987 static void kmemcg_cache_deact_after_rcu(struct kmem_cache *s)
3990 * Called with all the locks held after a sched RCU grace period.
3991 * Even if @s becomes empty after shrinking, we can't know that @s
3992 * doesn't have allocations already in-flight and thus can't
3993 * destroy @s until the associated memcg is released.
3995 * However, let's remove the sysfs files for empty caches here.
3996 * Each cache has a lot of interface files which aren't
3997 * particularly useful for empty draining caches; otherwise, we can
3998 * easily end up with millions of unnecessary sysfs files on
3999 * systems which have a lot of memory and transient cgroups.
4001 if (!__kmem_cache_shrink(s))
4002 sysfs_slab_remove(s);
4005 void __kmemcg_cache_deactivate(struct kmem_cache *s)
4008 * Disable empty slabs caching. Used to avoid pinning offline
4009 * memory cgroups by kmem pages that can be freed.
4011 slub_set_cpu_partial(s, 0);
4015 * s->cpu_partial is checked locklessly (see put_cpu_partial), so
4016 * we have to make sure the change is visible before shrinking.
4018 slab_deactivate_memcg_cache_rcu_sched(s, kmemcg_cache_deact_after_rcu);
4022 static int slab_mem_going_offline_callback(void *arg)
4024 struct kmem_cache *s;
4026 mutex_lock(&slab_mutex);
4027 list_for_each_entry(s, &slab_caches, list)
4028 __kmem_cache_shrink(s);
4029 mutex_unlock(&slab_mutex);
4034 static void slab_mem_offline_callback(void *arg)
4036 struct kmem_cache_node *n;
4037 struct kmem_cache *s;
4038 struct memory_notify *marg = arg;
4041 offline_node = marg->status_change_nid_normal;
4044 * If the node still has available memory. we need kmem_cache_node
4047 if (offline_node < 0)
4050 mutex_lock(&slab_mutex);
4051 list_for_each_entry(s, &slab_caches, list) {
4052 n = get_node(s, offline_node);
4055 * if n->nr_slabs > 0, slabs still exist on the node
4056 * that is going down. We were unable to free them,
4057 * and offline_pages() function shouldn't call this
4058 * callback. So, we must fail.
4060 BUG_ON(slabs_node(s, offline_node));
4062 s->node[offline_node] = NULL;
4063 kmem_cache_free(kmem_cache_node, n);
4066 mutex_unlock(&slab_mutex);
4069 static int slab_mem_going_online_callback(void *arg)
4071 struct kmem_cache_node *n;
4072 struct kmem_cache *s;
4073 struct memory_notify *marg = arg;
4074 int nid = marg->status_change_nid_normal;
4078 * If the node's memory is already available, then kmem_cache_node is
4079 * already created. Nothing to do.
4085 * We are bringing a node online. No memory is available yet. We must
4086 * allocate a kmem_cache_node structure in order to bring the node
4089 mutex_lock(&slab_mutex);
4090 list_for_each_entry(s, &slab_caches, list) {
4092 * XXX: kmem_cache_alloc_node will fallback to other nodes
4093 * since memory is not yet available from the node that
4096 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4101 init_kmem_cache_node(n);
4105 mutex_unlock(&slab_mutex);
4109 static int slab_memory_callback(struct notifier_block *self,
4110 unsigned long action, void *arg)
4115 case MEM_GOING_ONLINE:
4116 ret = slab_mem_going_online_callback(arg);
4118 case MEM_GOING_OFFLINE:
4119 ret = slab_mem_going_offline_callback(arg);
4122 case MEM_CANCEL_ONLINE:
4123 slab_mem_offline_callback(arg);
4126 case MEM_CANCEL_OFFLINE:
4130 ret = notifier_from_errno(ret);
4136 static struct notifier_block slab_memory_callback_nb = {
4137 .notifier_call = slab_memory_callback,
4138 .priority = SLAB_CALLBACK_PRI,
4141 /********************************************************************
4142 * Basic setup of slabs
4143 *******************************************************************/
4146 * Used for early kmem_cache structures that were allocated using
4147 * the page allocator. Allocate them properly then fix up the pointers
4148 * that may be pointing to the wrong kmem_cache structure.
4151 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4154 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4155 struct kmem_cache_node *n;
4157 memcpy(s, static_cache, kmem_cache->object_size);
4160 * This runs very early, and only the boot processor is supposed to be
4161 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4164 __flush_cpu_slab(s, smp_processor_id());
4165 for_each_kmem_cache_node(s, node, n) {
4168 list_for_each_entry(p, &n->partial, lru)
4171 #ifdef CONFIG_SLUB_DEBUG
4172 list_for_each_entry(p, &n->full, lru)
4176 slab_init_memcg_params(s);
4177 list_add(&s->list, &slab_caches);
4178 memcg_link_cache(s);
4182 void __init kmem_cache_init(void)
4184 static __initdata struct kmem_cache boot_kmem_cache,
4185 boot_kmem_cache_node;
4187 if (debug_guardpage_minorder())
4190 kmem_cache_node = &boot_kmem_cache_node;
4191 kmem_cache = &boot_kmem_cache;
4193 create_boot_cache(kmem_cache_node, "kmem_cache_node",
4194 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
4196 register_hotmemory_notifier(&slab_memory_callback_nb);
4198 /* Able to allocate the per node structures */
4199 slab_state = PARTIAL;
4201 create_boot_cache(kmem_cache, "kmem_cache",
4202 offsetof(struct kmem_cache, node) +
4203 nr_node_ids * sizeof(struct kmem_cache_node *),
4204 SLAB_HWCACHE_ALIGN, 0, 0);
4206 kmem_cache = bootstrap(&boot_kmem_cache);
4207 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
4209 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4210 setup_kmalloc_cache_index_table();
4211 create_kmalloc_caches(0);
4213 /* Setup random freelists for each cache */
4214 init_freelist_randomization();
4216 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
4219 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%d\n",
4221 slub_min_order, slub_max_order, slub_min_objects,
4222 nr_cpu_ids, nr_node_ids);
4225 void __init kmem_cache_init_late(void)
4230 __kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
4231 slab_flags_t flags, void (*ctor)(void *))
4233 struct kmem_cache *s, *c;
4235 s = find_mergeable(size, align, flags, name, ctor);
4240 * Adjust the object sizes so that we clear
4241 * the complete object on kzalloc.
4243 s->object_size = max(s->object_size, size);
4244 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
4246 for_each_memcg_cache(c, s) {
4247 c->object_size = s->object_size;
4248 c->inuse = max(c->inuse, ALIGN(size, sizeof(void *)));
4251 if (sysfs_slab_alias(s, name)) {
4260 int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
4264 err = kmem_cache_open(s, flags);
4268 /* Mutex is not taken during early boot */
4269 if (slab_state <= UP)
4272 memcg_propagate_slab_attrs(s);
4273 err = sysfs_slab_add(s);
4275 __kmem_cache_release(s);
4280 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4282 struct kmem_cache *s;
4285 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4286 return kmalloc_large(size, gfpflags);
4288 s = kmalloc_slab(size, gfpflags);
4290 if (unlikely(ZERO_OR_NULL_PTR(s)))
4293 ret = slab_alloc(s, gfpflags, caller);
4295 /* Honor the call site pointer we received. */
4296 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4302 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4303 int node, unsigned long caller)
4305 struct kmem_cache *s;
4308 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4309 ret = kmalloc_large_node(size, gfpflags, node);
4311 trace_kmalloc_node(caller, ret,
4312 size, PAGE_SIZE << get_order(size),
4318 s = kmalloc_slab(size, gfpflags);
4320 if (unlikely(ZERO_OR_NULL_PTR(s)))
4323 ret = slab_alloc_node(s, gfpflags, node, caller);
4325 /* Honor the call site pointer we received. */
4326 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4333 static int count_inuse(struct page *page)
4338 static int count_total(struct page *page)
4340 return page->objects;
4344 #ifdef CONFIG_SLUB_DEBUG
4345 static int validate_slab(struct kmem_cache *s, struct page *page,
4349 void *addr = page_address(page);
4351 if (!check_slab(s, page) ||
4352 !on_freelist(s, page, NULL))
4355 /* Now we know that a valid freelist exists */
4356 bitmap_zero(map, page->objects);
4358 get_map(s, page, map);
4359 for_each_object(p, s, addr, page->objects) {
4360 if (test_bit(slab_index(p, s, addr), map))
4361 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4365 for_each_object(p, s, addr, page->objects)
4366 if (!test_bit(slab_index(p, s, addr), map))
4367 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4372 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4376 validate_slab(s, page, map);
4380 static int validate_slab_node(struct kmem_cache *s,
4381 struct kmem_cache_node *n, unsigned long *map)
4383 unsigned long count = 0;
4385 unsigned long flags;
4387 spin_lock_irqsave(&n->list_lock, flags);
4389 list_for_each_entry(page, &n->partial, lru) {
4390 validate_slab_slab(s, page, map);
4393 if (count != n->nr_partial)
4394 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4395 s->name, count, n->nr_partial);
4397 if (!(s->flags & SLAB_STORE_USER))
4400 list_for_each_entry(page, &n->full, lru) {
4401 validate_slab_slab(s, page, map);
4404 if (count != atomic_long_read(&n->nr_slabs))
4405 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4406 s->name, count, atomic_long_read(&n->nr_slabs));
4409 spin_unlock_irqrestore(&n->list_lock, flags);
4413 static long validate_slab_cache(struct kmem_cache *s)
4416 unsigned long count = 0;
4417 unsigned long *map = kmalloc_array(BITS_TO_LONGS(oo_objects(s->max)),
4418 sizeof(unsigned long),
4420 struct kmem_cache_node *n;
4426 for_each_kmem_cache_node(s, node, n)
4427 count += validate_slab_node(s, n, map);
4432 * Generate lists of code addresses where slabcache objects are allocated
4437 unsigned long count;
4444 DECLARE_BITMAP(cpus, NR_CPUS);
4450 unsigned long count;
4451 struct location *loc;
4454 static void free_loc_track(struct loc_track *t)
4457 free_pages((unsigned long)t->loc,
4458 get_order(sizeof(struct location) * t->max));
4461 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4466 order = get_order(sizeof(struct location) * max);
4468 l = (void *)__get_free_pages(flags, order);
4473 memcpy(l, t->loc, sizeof(struct location) * t->count);
4481 static int add_location(struct loc_track *t, struct kmem_cache *s,
4482 const struct track *track)
4484 long start, end, pos;
4486 unsigned long caddr;
4487 unsigned long age = jiffies - track->when;
4493 pos = start + (end - start + 1) / 2;
4496 * There is nothing at "end". If we end up there
4497 * we need to add something to before end.
4502 caddr = t->loc[pos].addr;
4503 if (track->addr == caddr) {
4509 if (age < l->min_time)
4511 if (age > l->max_time)
4514 if (track->pid < l->min_pid)
4515 l->min_pid = track->pid;
4516 if (track->pid > l->max_pid)
4517 l->max_pid = track->pid;
4519 cpumask_set_cpu(track->cpu,
4520 to_cpumask(l->cpus));
4522 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4526 if (track->addr < caddr)
4533 * Not found. Insert new tracking element.
4535 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4541 (t->count - pos) * sizeof(struct location));
4544 l->addr = track->addr;
4548 l->min_pid = track->pid;
4549 l->max_pid = track->pid;
4550 cpumask_clear(to_cpumask(l->cpus));
4551 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4552 nodes_clear(l->nodes);
4553 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4557 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4558 struct page *page, enum track_item alloc,
4561 void *addr = page_address(page);
4564 bitmap_zero(map, page->objects);
4565 get_map(s, page, map);
4567 for_each_object(p, s, addr, page->objects)
4568 if (!test_bit(slab_index(p, s, addr), map))
4569 add_location(t, s, get_track(s, p, alloc));
4572 static int list_locations(struct kmem_cache *s, char *buf,
4573 enum track_item alloc)
4577 struct loc_track t = { 0, 0, NULL };
4579 unsigned long *map = kmalloc_array(BITS_TO_LONGS(oo_objects(s->max)),
4580 sizeof(unsigned long),
4582 struct kmem_cache_node *n;
4584 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4587 return sprintf(buf, "Out of memory\n");
4589 /* Push back cpu slabs */
4592 for_each_kmem_cache_node(s, node, n) {
4593 unsigned long flags;
4596 if (!atomic_long_read(&n->nr_slabs))
4599 spin_lock_irqsave(&n->list_lock, flags);
4600 list_for_each_entry(page, &n->partial, lru)
4601 process_slab(&t, s, page, alloc, map);
4602 list_for_each_entry(page, &n->full, lru)
4603 process_slab(&t, s, page, alloc, map);
4604 spin_unlock_irqrestore(&n->list_lock, flags);
4607 for (i = 0; i < t.count; i++) {
4608 struct location *l = &t.loc[i];
4610 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4612 len += sprintf(buf + len, "%7ld ", l->count);
4615 len += sprintf(buf + len, "%pS", (void *)l->addr);
4617 len += sprintf(buf + len, "<not-available>");
4619 if (l->sum_time != l->min_time) {
4620 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4622 (long)div_u64(l->sum_time, l->count),
4625 len += sprintf(buf + len, " age=%ld",
4628 if (l->min_pid != l->max_pid)
4629 len += sprintf(buf + len, " pid=%ld-%ld",
4630 l->min_pid, l->max_pid);
4632 len += sprintf(buf + len, " pid=%ld",
4635 if (num_online_cpus() > 1 &&
4636 !cpumask_empty(to_cpumask(l->cpus)) &&
4637 len < PAGE_SIZE - 60)
4638 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4640 cpumask_pr_args(to_cpumask(l->cpus)));
4642 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4643 len < PAGE_SIZE - 60)
4644 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4646 nodemask_pr_args(&l->nodes));
4648 len += sprintf(buf + len, "\n");
4654 len += sprintf(buf, "No data\n");
4659 #ifdef SLUB_RESILIENCY_TEST
4660 static void __init resiliency_test(void)
4664 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4666 pr_err("SLUB resiliency testing\n");
4667 pr_err("-----------------------\n");
4668 pr_err("A. Corruption after allocation\n");
4670 p = kzalloc(16, GFP_KERNEL);
4672 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4675 validate_slab_cache(kmalloc_caches[4]);
4677 /* Hmmm... The next two are dangerous */
4678 p = kzalloc(32, GFP_KERNEL);
4679 p[32 + sizeof(void *)] = 0x34;
4680 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4682 pr_err("If allocated object is overwritten then not detectable\n\n");
4684 validate_slab_cache(kmalloc_caches[5]);
4685 p = kzalloc(64, GFP_KERNEL);
4686 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4688 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4690 pr_err("If allocated object is overwritten then not detectable\n\n");
4691 validate_slab_cache(kmalloc_caches[6]);
4693 pr_err("\nB. Corruption after free\n");
4694 p = kzalloc(128, GFP_KERNEL);
4697 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4698 validate_slab_cache(kmalloc_caches[7]);
4700 p = kzalloc(256, GFP_KERNEL);
4703 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4704 validate_slab_cache(kmalloc_caches[8]);
4706 p = kzalloc(512, GFP_KERNEL);
4709 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4710 validate_slab_cache(kmalloc_caches[9]);
4714 static void resiliency_test(void) {};
4719 enum slab_stat_type {
4720 SL_ALL, /* All slabs */
4721 SL_PARTIAL, /* Only partially allocated slabs */
4722 SL_CPU, /* Only slabs used for cpu caches */
4723 SL_OBJECTS, /* Determine allocated objects not slabs */
4724 SL_TOTAL /* Determine object capacity not slabs */
4727 #define SO_ALL (1 << SL_ALL)
4728 #define SO_PARTIAL (1 << SL_PARTIAL)
4729 #define SO_CPU (1 << SL_CPU)
4730 #define SO_OBJECTS (1 << SL_OBJECTS)
4731 #define SO_TOTAL (1 << SL_TOTAL)
4734 static bool memcg_sysfs_enabled = IS_ENABLED(CONFIG_SLUB_MEMCG_SYSFS_ON);
4736 static int __init setup_slub_memcg_sysfs(char *str)
4740 if (get_option(&str, &v) > 0)
4741 memcg_sysfs_enabled = v;
4746 __setup("slub_memcg_sysfs=", setup_slub_memcg_sysfs);
4749 static ssize_t show_slab_objects(struct kmem_cache *s,
4750 char *buf, unsigned long flags)
4752 unsigned long total = 0;
4755 unsigned long *nodes;
4757 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
4761 if (flags & SO_CPU) {
4764 for_each_possible_cpu(cpu) {
4765 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4770 page = READ_ONCE(c->page);
4774 node = page_to_nid(page);
4775 if (flags & SO_TOTAL)
4777 else if (flags & SO_OBJECTS)
4785 page = slub_percpu_partial_read_once(c);
4787 node = page_to_nid(page);
4788 if (flags & SO_TOTAL)
4790 else if (flags & SO_OBJECTS)
4801 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
4802 * already held which will conflict with an existing lock order:
4804 * mem_hotplug_lock->slab_mutex->kernfs_mutex
4806 * We don't really need mem_hotplug_lock (to hold off
4807 * slab_mem_going_offline_callback) here because slab's memory hot
4808 * unplug code doesn't destroy the kmem_cache->node[] data.
4811 #ifdef CONFIG_SLUB_DEBUG
4812 if (flags & SO_ALL) {
4813 struct kmem_cache_node *n;
4815 for_each_kmem_cache_node(s, node, n) {
4817 if (flags & SO_TOTAL)
4818 x = atomic_long_read(&n->total_objects);
4819 else if (flags & SO_OBJECTS)
4820 x = atomic_long_read(&n->total_objects) -
4821 count_partial(n, count_free);
4823 x = atomic_long_read(&n->nr_slabs);
4830 if (flags & SO_PARTIAL) {
4831 struct kmem_cache_node *n;
4833 for_each_kmem_cache_node(s, node, n) {
4834 if (flags & SO_TOTAL)
4835 x = count_partial(n, count_total);
4836 else if (flags & SO_OBJECTS)
4837 x = count_partial(n, count_inuse);
4844 x = sprintf(buf, "%lu", total);
4846 for (node = 0; node < nr_node_ids; node++)
4848 x += sprintf(buf + x, " N%d=%lu",
4852 return x + sprintf(buf + x, "\n");
4855 #ifdef CONFIG_SLUB_DEBUG
4856 static int any_slab_objects(struct kmem_cache *s)
4859 struct kmem_cache_node *n;
4861 for_each_kmem_cache_node(s, node, n)
4862 if (atomic_long_read(&n->total_objects))
4869 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4870 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4872 struct slab_attribute {
4873 struct attribute attr;
4874 ssize_t (*show)(struct kmem_cache *s, char *buf);
4875 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4878 #define SLAB_ATTR_RO(_name) \
4879 static struct slab_attribute _name##_attr = \
4880 __ATTR(_name, 0400, _name##_show, NULL)
4882 #define SLAB_ATTR(_name) \
4883 static struct slab_attribute _name##_attr = \
4884 __ATTR(_name, 0600, _name##_show, _name##_store)
4886 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4888 return sprintf(buf, "%u\n", s->size);
4890 SLAB_ATTR_RO(slab_size);
4892 static ssize_t align_show(struct kmem_cache *s, char *buf)
4894 return sprintf(buf, "%u\n", s->align);
4896 SLAB_ATTR_RO(align);
4898 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4900 return sprintf(buf, "%u\n", s->object_size);
4902 SLAB_ATTR_RO(object_size);
4904 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4906 return sprintf(buf, "%u\n", oo_objects(s->oo));
4908 SLAB_ATTR_RO(objs_per_slab);
4910 static ssize_t order_store(struct kmem_cache *s,
4911 const char *buf, size_t length)
4916 err = kstrtouint(buf, 10, &order);
4920 if (order > slub_max_order || order < slub_min_order)
4923 calculate_sizes(s, order);
4927 static ssize_t order_show(struct kmem_cache *s, char *buf)
4929 return sprintf(buf, "%u\n", oo_order(s->oo));
4933 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4935 return sprintf(buf, "%lu\n", s->min_partial);
4938 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4944 err = kstrtoul(buf, 10, &min);
4948 set_min_partial(s, min);
4951 SLAB_ATTR(min_partial);
4953 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4955 return sprintf(buf, "%u\n", slub_cpu_partial(s));
4958 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4961 unsigned int objects;
4964 err = kstrtouint(buf, 10, &objects);
4967 if (objects && !kmem_cache_has_cpu_partial(s))
4970 slub_set_cpu_partial(s, objects);
4974 SLAB_ATTR(cpu_partial);
4976 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4980 return sprintf(buf, "%pS\n", s->ctor);
4984 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4986 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
4988 SLAB_ATTR_RO(aliases);
4990 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4992 return show_slab_objects(s, buf, SO_PARTIAL);
4994 SLAB_ATTR_RO(partial);
4996 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4998 return show_slab_objects(s, buf, SO_CPU);
5000 SLAB_ATTR_RO(cpu_slabs);
5002 static ssize_t objects_show(struct kmem_cache *s, char *buf)
5004 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5006 SLAB_ATTR_RO(objects);
5008 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5010 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5012 SLAB_ATTR_RO(objects_partial);
5014 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5021 for_each_online_cpu(cpu) {
5024 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5027 pages += page->pages;
5028 objects += page->pobjects;
5032 len = sprintf(buf, "%d(%d)", objects, pages);
5035 for_each_online_cpu(cpu) {
5038 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5040 if (page && len < PAGE_SIZE - 20)
5041 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
5042 page->pobjects, page->pages);
5045 return len + sprintf(buf + len, "\n");
5047 SLAB_ATTR_RO(slabs_cpu_partial);
5049 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5051 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5054 static ssize_t reclaim_account_store(struct kmem_cache *s,
5055 const char *buf, size_t length)
5057 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
5059 s->flags |= SLAB_RECLAIM_ACCOUNT;
5062 SLAB_ATTR(reclaim_account);
5064 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5066 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5068 SLAB_ATTR_RO(hwcache_align);
5070 #ifdef CONFIG_ZONE_DMA
5071 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5073 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5075 SLAB_ATTR_RO(cache_dma);
5078 static ssize_t usersize_show(struct kmem_cache *s, char *buf)
5080 return sprintf(buf, "%u\n", s->usersize);
5082 SLAB_ATTR_RO(usersize);
5084 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5086 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
5088 SLAB_ATTR_RO(destroy_by_rcu);
5090 #ifdef CONFIG_SLUB_DEBUG
5091 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5093 return show_slab_objects(s, buf, SO_ALL);
5095 SLAB_ATTR_RO(slabs);
5097 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5099 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5101 SLAB_ATTR_RO(total_objects);
5103 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5105 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5108 static ssize_t sanity_checks_store(struct kmem_cache *s,
5109 const char *buf, size_t length)
5111 s->flags &= ~SLAB_CONSISTENCY_CHECKS;
5112 if (buf[0] == '1') {
5113 s->flags &= ~__CMPXCHG_DOUBLE;
5114 s->flags |= SLAB_CONSISTENCY_CHECKS;
5118 SLAB_ATTR(sanity_checks);
5120 static ssize_t trace_show(struct kmem_cache *s, char *buf)
5122 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5125 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
5129 * Tracing a merged cache is going to give confusing results
5130 * as well as cause other issues like converting a mergeable
5131 * cache into an umergeable one.
5133 if (s->refcount > 1)
5136 s->flags &= ~SLAB_TRACE;
5137 if (buf[0] == '1') {
5138 s->flags &= ~__CMPXCHG_DOUBLE;
5139 s->flags |= SLAB_TRACE;
5145 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5147 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5150 static ssize_t red_zone_store(struct kmem_cache *s,
5151 const char *buf, size_t length)
5153 if (any_slab_objects(s))
5156 s->flags &= ~SLAB_RED_ZONE;
5157 if (buf[0] == '1') {
5158 s->flags |= SLAB_RED_ZONE;
5160 calculate_sizes(s, -1);
5163 SLAB_ATTR(red_zone);
5165 static ssize_t poison_show(struct kmem_cache *s, char *buf)
5167 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
5170 static ssize_t poison_store(struct kmem_cache *s,
5171 const char *buf, size_t length)
5173 if (any_slab_objects(s))
5176 s->flags &= ~SLAB_POISON;
5177 if (buf[0] == '1') {
5178 s->flags |= SLAB_POISON;
5180 calculate_sizes(s, -1);
5185 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5187 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5190 static ssize_t store_user_store(struct kmem_cache *s,
5191 const char *buf, size_t length)
5193 if (any_slab_objects(s))
5196 s->flags &= ~SLAB_STORE_USER;
5197 if (buf[0] == '1') {
5198 s->flags &= ~__CMPXCHG_DOUBLE;
5199 s->flags |= SLAB_STORE_USER;
5201 calculate_sizes(s, -1);
5204 SLAB_ATTR(store_user);
5206 static ssize_t validate_show(struct kmem_cache *s, char *buf)
5211 static ssize_t validate_store(struct kmem_cache *s,
5212 const char *buf, size_t length)
5216 if (buf[0] == '1') {
5217 ret = validate_slab_cache(s);
5223 SLAB_ATTR(validate);
5225 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
5227 if (!(s->flags & SLAB_STORE_USER))
5229 return list_locations(s, buf, TRACK_ALLOC);
5231 SLAB_ATTR_RO(alloc_calls);
5233 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
5235 if (!(s->flags & SLAB_STORE_USER))
5237 return list_locations(s, buf, TRACK_FREE);
5239 SLAB_ATTR_RO(free_calls);
5240 #endif /* CONFIG_SLUB_DEBUG */
5242 #ifdef CONFIG_FAILSLAB
5243 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5245 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5248 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
5251 if (s->refcount > 1)
5254 s->flags &= ~SLAB_FAILSLAB;
5256 s->flags |= SLAB_FAILSLAB;
5259 SLAB_ATTR(failslab);
5262 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5267 static ssize_t shrink_store(struct kmem_cache *s,
5268 const char *buf, size_t length)
5271 kmem_cache_shrink(s);
5279 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5281 return sprintf(buf, "%u\n", s->remote_node_defrag_ratio / 10);
5284 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5285 const char *buf, size_t length)
5290 err = kstrtouint(buf, 10, &ratio);
5296 s->remote_node_defrag_ratio = ratio * 10;
5300 SLAB_ATTR(remote_node_defrag_ratio);
5303 #ifdef CONFIG_SLUB_STATS
5304 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5306 unsigned long sum = 0;
5309 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
5314 for_each_online_cpu(cpu) {
5315 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5321 len = sprintf(buf, "%lu", sum);
5324 for_each_online_cpu(cpu) {
5325 if (data[cpu] && len < PAGE_SIZE - 20)
5326 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5330 return len + sprintf(buf + len, "\n");
5333 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5337 for_each_online_cpu(cpu)
5338 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5341 #define STAT_ATTR(si, text) \
5342 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5344 return show_stat(s, buf, si); \
5346 static ssize_t text##_store(struct kmem_cache *s, \
5347 const char *buf, size_t length) \
5349 if (buf[0] != '0') \
5351 clear_stat(s, si); \
5356 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5357 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5358 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5359 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5360 STAT_ATTR(FREE_FROZEN, free_frozen);
5361 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5362 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5363 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5364 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5365 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5366 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5367 STAT_ATTR(FREE_SLAB, free_slab);
5368 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5369 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5370 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5371 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5372 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5373 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5374 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5375 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5376 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5377 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5378 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5379 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5380 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5381 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5384 static struct attribute *slab_attrs[] = {
5385 &slab_size_attr.attr,
5386 &object_size_attr.attr,
5387 &objs_per_slab_attr.attr,
5389 &min_partial_attr.attr,
5390 &cpu_partial_attr.attr,
5392 &objects_partial_attr.attr,
5394 &cpu_slabs_attr.attr,
5398 &hwcache_align_attr.attr,
5399 &reclaim_account_attr.attr,
5400 &destroy_by_rcu_attr.attr,
5402 &slabs_cpu_partial_attr.attr,
5403 #ifdef CONFIG_SLUB_DEBUG
5404 &total_objects_attr.attr,
5406 &sanity_checks_attr.attr,
5408 &red_zone_attr.attr,
5410 &store_user_attr.attr,
5411 &validate_attr.attr,
5412 &alloc_calls_attr.attr,
5413 &free_calls_attr.attr,
5415 #ifdef CONFIG_ZONE_DMA
5416 &cache_dma_attr.attr,
5419 &remote_node_defrag_ratio_attr.attr,
5421 #ifdef CONFIG_SLUB_STATS
5422 &alloc_fastpath_attr.attr,
5423 &alloc_slowpath_attr.attr,
5424 &free_fastpath_attr.attr,
5425 &free_slowpath_attr.attr,
5426 &free_frozen_attr.attr,
5427 &free_add_partial_attr.attr,
5428 &free_remove_partial_attr.attr,
5429 &alloc_from_partial_attr.attr,
5430 &alloc_slab_attr.attr,
5431 &alloc_refill_attr.attr,
5432 &alloc_node_mismatch_attr.attr,
5433 &free_slab_attr.attr,
5434 &cpuslab_flush_attr.attr,
5435 &deactivate_full_attr.attr,
5436 &deactivate_empty_attr.attr,
5437 &deactivate_to_head_attr.attr,
5438 &deactivate_to_tail_attr.attr,
5439 &deactivate_remote_frees_attr.attr,
5440 &deactivate_bypass_attr.attr,
5441 &order_fallback_attr.attr,
5442 &cmpxchg_double_fail_attr.attr,
5443 &cmpxchg_double_cpu_fail_attr.attr,
5444 &cpu_partial_alloc_attr.attr,
5445 &cpu_partial_free_attr.attr,
5446 &cpu_partial_node_attr.attr,
5447 &cpu_partial_drain_attr.attr,
5449 #ifdef CONFIG_FAILSLAB
5450 &failslab_attr.attr,
5452 &usersize_attr.attr,
5457 static const struct attribute_group slab_attr_group = {
5458 .attrs = slab_attrs,
5461 static ssize_t slab_attr_show(struct kobject *kobj,
5462 struct attribute *attr,
5465 struct slab_attribute *attribute;
5466 struct kmem_cache *s;
5469 attribute = to_slab_attr(attr);
5472 if (!attribute->show)
5475 err = attribute->show(s, buf);
5480 static ssize_t slab_attr_store(struct kobject *kobj,
5481 struct attribute *attr,
5482 const char *buf, size_t len)
5484 struct slab_attribute *attribute;
5485 struct kmem_cache *s;
5488 attribute = to_slab_attr(attr);
5491 if (!attribute->store)
5494 err = attribute->store(s, buf, len);
5496 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5497 struct kmem_cache *c;
5499 mutex_lock(&slab_mutex);
5500 if (s->max_attr_size < len)
5501 s->max_attr_size = len;
5504 * This is a best effort propagation, so this function's return
5505 * value will be determined by the parent cache only. This is
5506 * basically because not all attributes will have a well
5507 * defined semantics for rollbacks - most of the actions will
5508 * have permanent effects.
5510 * Returning the error value of any of the children that fail
5511 * is not 100 % defined, in the sense that users seeing the
5512 * error code won't be able to know anything about the state of
5515 * Only returning the error code for the parent cache at least
5516 * has well defined semantics. The cache being written to
5517 * directly either failed or succeeded, in which case we loop
5518 * through the descendants with best-effort propagation.
5520 for_each_memcg_cache(c, s)
5521 attribute->store(c, buf, len);
5522 mutex_unlock(&slab_mutex);
5528 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5532 char *buffer = NULL;
5533 struct kmem_cache *root_cache;
5535 if (is_root_cache(s))
5538 root_cache = s->memcg_params.root_cache;
5541 * This mean this cache had no attribute written. Therefore, no point
5542 * in copying default values around
5544 if (!root_cache->max_attr_size)
5547 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5550 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5553 if (!attr || !attr->store || !attr->show)
5557 * It is really bad that we have to allocate here, so we will
5558 * do it only as a fallback. If we actually allocate, though,
5559 * we can just use the allocated buffer until the end.
5561 * Most of the slub attributes will tend to be very small in
5562 * size, but sysfs allows buffers up to a page, so they can
5563 * theoretically happen.
5567 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5570 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5571 if (WARN_ON(!buffer))
5576 len = attr->show(root_cache, buf);
5578 attr->store(s, buf, len);
5582 free_page((unsigned long)buffer);
5586 static void kmem_cache_release(struct kobject *k)
5588 slab_kmem_cache_release(to_slab(k));
5591 static const struct sysfs_ops slab_sysfs_ops = {
5592 .show = slab_attr_show,
5593 .store = slab_attr_store,
5596 static struct kobj_type slab_ktype = {
5597 .sysfs_ops = &slab_sysfs_ops,
5598 .release = kmem_cache_release,
5601 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5603 struct kobj_type *ktype = get_ktype(kobj);
5605 if (ktype == &slab_ktype)
5610 static const struct kset_uevent_ops slab_uevent_ops = {
5611 .filter = uevent_filter,
5614 static struct kset *slab_kset;
5616 static inline struct kset *cache_kset(struct kmem_cache *s)
5619 if (!is_root_cache(s))
5620 return s->memcg_params.root_cache->memcg_kset;
5625 #define ID_STR_LENGTH 64
5627 /* Create a unique string id for a slab cache:
5629 * Format :[flags-]size
5631 static char *create_unique_id(struct kmem_cache *s)
5633 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5640 * First flags affecting slabcache operations. We will only
5641 * get here for aliasable slabs so we do not need to support
5642 * too many flags. The flags here must cover all flags that
5643 * are matched during merging to guarantee that the id is
5646 if (s->flags & SLAB_CACHE_DMA)
5648 if (s->flags & SLAB_CACHE_DMA32)
5650 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5652 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5654 if (s->flags & SLAB_ACCOUNT)
5658 p += sprintf(p, "%07u", s->size);
5660 BUG_ON(p > name + ID_STR_LENGTH - 1);
5664 static void sysfs_slab_remove_workfn(struct work_struct *work)
5666 struct kmem_cache *s =
5667 container_of(work, struct kmem_cache, kobj_remove_work);
5669 if (!s->kobj.state_in_sysfs)
5671 * For a memcg cache, this may be called during
5672 * deactivation and again on shutdown. Remove only once.
5673 * A cache is never shut down before deactivation is
5674 * complete, so no need to worry about synchronization.
5679 kset_unregister(s->memcg_kset);
5681 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5683 kobject_put(&s->kobj);
5686 static int sysfs_slab_add(struct kmem_cache *s)
5690 struct kset *kset = cache_kset(s);
5691 int unmergeable = slab_unmergeable(s);
5693 INIT_WORK(&s->kobj_remove_work, sysfs_slab_remove_workfn);
5696 kobject_init(&s->kobj, &slab_ktype);
5700 if (!unmergeable && disable_higher_order_debug &&
5701 (slub_debug & DEBUG_METADATA_FLAGS))
5706 * Slabcache can never be merged so we can use the name proper.
5707 * This is typically the case for debug situations. In that
5708 * case we can catch duplicate names easily.
5710 sysfs_remove_link(&slab_kset->kobj, s->name);
5714 * Create a unique name for the slab as a target
5717 name = create_unique_id(s);
5720 s->kobj.kset = kset;
5721 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5725 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5730 if (is_root_cache(s) && memcg_sysfs_enabled) {
5731 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5732 if (!s->memcg_kset) {
5739 kobject_uevent(&s->kobj, KOBJ_ADD);
5741 /* Setup first alias */
5742 sysfs_slab_alias(s, s->name);
5749 kobject_del(&s->kobj);
5753 static void sysfs_slab_remove(struct kmem_cache *s)
5755 if (slab_state < FULL)
5757 * Sysfs has not been setup yet so no need to remove the
5762 kobject_get(&s->kobj);
5763 schedule_work(&s->kobj_remove_work);
5766 void sysfs_slab_unlink(struct kmem_cache *s)
5768 if (slab_state >= FULL)
5769 kobject_del(&s->kobj);
5772 void sysfs_slab_release(struct kmem_cache *s)
5774 if (slab_state >= FULL)
5775 kobject_put(&s->kobj);
5779 * Need to buffer aliases during bootup until sysfs becomes
5780 * available lest we lose that information.
5782 struct saved_alias {
5783 struct kmem_cache *s;
5785 struct saved_alias *next;
5788 static struct saved_alias *alias_list;
5790 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5792 struct saved_alias *al;
5794 if (slab_state == FULL) {
5796 * If we have a leftover link then remove it.
5798 sysfs_remove_link(&slab_kset->kobj, name);
5799 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5802 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5808 al->next = alias_list;
5813 static int __init slab_sysfs_init(void)
5815 struct kmem_cache *s;
5818 mutex_lock(&slab_mutex);
5820 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5822 mutex_unlock(&slab_mutex);
5823 pr_err("Cannot register slab subsystem.\n");
5829 list_for_each_entry(s, &slab_caches, list) {
5830 err = sysfs_slab_add(s);
5832 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5836 while (alias_list) {
5837 struct saved_alias *al = alias_list;
5839 alias_list = alias_list->next;
5840 err = sysfs_slab_alias(al->s, al->name);
5842 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5847 mutex_unlock(&slab_mutex);
5852 __initcall(slab_sysfs_init);
5853 #endif /* CONFIG_SYSFS */
5856 * The /proc/slabinfo ABI
5858 #ifdef CONFIG_SLUB_DEBUG
5859 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5861 unsigned long nr_slabs = 0;
5862 unsigned long nr_objs = 0;
5863 unsigned long nr_free = 0;
5865 struct kmem_cache_node *n;
5867 for_each_kmem_cache_node(s, node, n) {
5868 nr_slabs += node_nr_slabs(n);
5869 nr_objs += node_nr_objs(n);
5870 nr_free += count_partial(n, count_free);
5873 sinfo->active_objs = nr_objs - nr_free;
5874 sinfo->num_objs = nr_objs;
5875 sinfo->active_slabs = nr_slabs;
5876 sinfo->num_slabs = nr_slabs;
5877 sinfo->objects_per_slab = oo_objects(s->oo);
5878 sinfo->cache_order = oo_order(s->oo);
5881 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5885 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5886 size_t count, loff_t *ppos)
5890 #endif /* CONFIG_SLUB_DEBUG */