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 ^ swab(ptr_addr));
258 /* Returns the freelist pointer recorded at location ptr_addr. */
259 static inline void *freelist_dereference(const struct kmem_cache *s,
262 return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr),
263 (unsigned long)ptr_addr);
266 static inline void *get_freepointer(struct kmem_cache *s, void *object)
268 return freelist_dereference(s, object + s->offset);
271 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
273 prefetch(object + s->offset);
276 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
278 unsigned long freepointer_addr;
281 if (!debug_pagealloc_enabled())
282 return get_freepointer(s, object);
284 freepointer_addr = (unsigned long)object + s->offset;
285 probe_kernel_read(&p, (void **)freepointer_addr, sizeof(p));
286 return freelist_ptr(s, p, freepointer_addr);
289 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
291 unsigned long freeptr_addr = (unsigned long)object + s->offset;
293 #ifdef CONFIG_SLAB_FREELIST_HARDENED
294 BUG_ON(object == fp); /* naive detection of double free or corruption */
297 *(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr);
300 /* Loop over all objects in a slab */
301 #define for_each_object(__p, __s, __addr, __objects) \
302 for (__p = fixup_red_left(__s, __addr); \
303 __p < (__addr) + (__objects) * (__s)->size; \
306 #define for_each_object_idx(__p, __idx, __s, __addr, __objects) \
307 for (__p = fixup_red_left(__s, __addr), __idx = 1; \
308 __idx <= __objects; \
309 __p += (__s)->size, __idx++)
311 /* Determine object index from a given position */
312 static inline 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 bool freelist_corrupted(struct kmem_cache *s, struct page *page,
649 void **freelist, void *nextfree)
651 if ((s->flags & SLAB_CONSISTENCY_CHECKS) &&
652 !check_valid_pointer(s, page, nextfree) && freelist) {
653 object_err(s, page, *freelist, "Freechain corrupt");
655 slab_fix(s, "Isolate corrupted freechain");
662 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
664 unsigned int off; /* Offset of last byte */
665 u8 *addr = page_address(page);
667 print_tracking(s, p);
669 print_page_info(page);
671 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
672 p, p - addr, get_freepointer(s, p));
674 if (s->flags & SLAB_RED_ZONE)
675 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
677 else if (p > addr + 16)
678 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
680 print_section(KERN_ERR, "Object ", p,
681 min_t(unsigned int, s->object_size, PAGE_SIZE));
682 if (s->flags & SLAB_RED_ZONE)
683 print_section(KERN_ERR, "Redzone ", p + s->object_size,
684 s->inuse - s->object_size);
687 off = s->offset + sizeof(void *);
691 if (s->flags & SLAB_STORE_USER)
692 off += 2 * sizeof(struct track);
694 off += kasan_metadata_size(s);
696 if (off != size_from_object(s))
697 /* Beginning of the filler is the free pointer */
698 print_section(KERN_ERR, "Padding ", p + off,
699 size_from_object(s) - off);
704 void object_err(struct kmem_cache *s, struct page *page,
705 u8 *object, char *reason)
707 slab_bug(s, "%s", reason);
708 print_trailer(s, page, object);
711 static __printf(3, 4) void slab_err(struct kmem_cache *s, struct page *page,
712 const char *fmt, ...)
718 vsnprintf(buf, sizeof(buf), fmt, args);
720 slab_bug(s, "%s", buf);
721 print_page_info(page);
725 static void init_object(struct kmem_cache *s, void *object, u8 val)
729 if (s->flags & SLAB_RED_ZONE)
730 memset(p - s->red_left_pad, val, s->red_left_pad);
732 if (s->flags & __OBJECT_POISON) {
733 memset(p, POISON_FREE, s->object_size - 1);
734 p[s->object_size - 1] = POISON_END;
737 if (s->flags & SLAB_RED_ZONE)
738 memset(p + s->object_size, val, s->inuse - s->object_size);
741 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
742 void *from, void *to)
744 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
745 memset(from, data, to - from);
748 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
749 u8 *object, char *what,
750 u8 *start, unsigned int value, unsigned int bytes)
755 metadata_access_enable();
756 fault = memchr_inv(start, value, bytes);
757 metadata_access_disable();
762 while (end > fault && end[-1] == value)
765 slab_bug(s, "%s overwritten", what);
766 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
767 fault, end - 1, fault[0], value);
768 print_trailer(s, page, object);
770 restore_bytes(s, what, value, fault, end);
778 * Bytes of the object to be managed.
779 * If the freepointer may overlay the object then the free
780 * pointer is the first word of the object.
782 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
785 * object + s->object_size
786 * Padding to reach word boundary. This is also used for Redzoning.
787 * Padding is extended by another word if Redzoning is enabled and
788 * object_size == inuse.
790 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
791 * 0xcc (RED_ACTIVE) for objects in use.
794 * Meta data starts here.
796 * A. Free pointer (if we cannot overwrite object on free)
797 * B. Tracking data for SLAB_STORE_USER
798 * C. Padding to reach required alignment boundary or at mininum
799 * one word if debugging is on to be able to detect writes
800 * before the word boundary.
802 * Padding is done using 0x5a (POISON_INUSE)
805 * Nothing is used beyond s->size.
807 * If slabcaches are merged then the object_size and inuse boundaries are mostly
808 * ignored. And therefore no slab options that rely on these boundaries
809 * may be used with merged slabcaches.
812 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
814 unsigned long off = s->inuse; /* The end of info */
817 /* Freepointer is placed after the object. */
818 off += sizeof(void *);
820 if (s->flags & SLAB_STORE_USER)
821 /* We also have user information there */
822 off += 2 * sizeof(struct track);
824 off += kasan_metadata_size(s);
826 if (size_from_object(s) == off)
829 return check_bytes_and_report(s, page, p, "Object padding",
830 p + off, POISON_INUSE, size_from_object(s) - off);
833 /* Check the pad bytes at the end of a slab page */
834 static int slab_pad_check(struct kmem_cache *s, struct page *page)
843 if (!(s->flags & SLAB_POISON))
846 start = page_address(page);
847 length = PAGE_SIZE << compound_order(page);
848 end = start + length;
849 remainder = length % s->size;
853 pad = end - remainder;
854 metadata_access_enable();
855 fault = memchr_inv(pad, POISON_INUSE, remainder);
856 metadata_access_disable();
859 while (end > fault && end[-1] == POISON_INUSE)
862 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
863 print_section(KERN_ERR, "Padding ", pad, remainder);
865 restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
869 static int check_object(struct kmem_cache *s, struct page *page,
870 void *object, u8 val)
873 u8 *endobject = object + s->object_size;
875 if (s->flags & SLAB_RED_ZONE) {
876 if (!check_bytes_and_report(s, page, object, "Redzone",
877 object - s->red_left_pad, val, s->red_left_pad))
880 if (!check_bytes_and_report(s, page, object, "Redzone",
881 endobject, val, s->inuse - s->object_size))
884 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
885 check_bytes_and_report(s, page, p, "Alignment padding",
886 endobject, POISON_INUSE,
887 s->inuse - s->object_size);
891 if (s->flags & SLAB_POISON) {
892 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
893 (!check_bytes_and_report(s, page, p, "Poison", p,
894 POISON_FREE, s->object_size - 1) ||
895 !check_bytes_and_report(s, page, p, "Poison",
896 p + s->object_size - 1, POISON_END, 1)))
899 * check_pad_bytes cleans up on its own.
901 check_pad_bytes(s, page, p);
904 if (!s->offset && val == SLUB_RED_ACTIVE)
906 * Object and freepointer overlap. Cannot check
907 * freepointer while object is allocated.
911 /* Check free pointer validity */
912 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
913 object_err(s, page, p, "Freepointer corrupt");
915 * No choice but to zap it and thus lose the remainder
916 * of the free objects in this slab. May cause
917 * another error because the object count is now wrong.
919 set_freepointer(s, p, NULL);
925 static int check_slab(struct kmem_cache *s, struct page *page)
929 VM_BUG_ON(!irqs_disabled());
931 if (!PageSlab(page)) {
932 slab_err(s, page, "Not a valid slab page");
936 maxobj = order_objects(compound_order(page), s->size);
937 if (page->objects > maxobj) {
938 slab_err(s, page, "objects %u > max %u",
939 page->objects, maxobj);
942 if (page->inuse > page->objects) {
943 slab_err(s, page, "inuse %u > max %u",
944 page->inuse, page->objects);
947 /* Slab_pad_check fixes things up after itself */
948 slab_pad_check(s, page);
953 * Determine if a certain object on a page is on the freelist. Must hold the
954 * slab lock to guarantee that the chains are in a consistent state.
956 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
964 while (fp && nr <= page->objects) {
967 if (!check_valid_pointer(s, page, fp)) {
969 object_err(s, page, object,
970 "Freechain corrupt");
971 set_freepointer(s, object, NULL);
973 slab_err(s, page, "Freepointer corrupt");
974 page->freelist = NULL;
975 page->inuse = page->objects;
976 slab_fix(s, "Freelist cleared");
982 fp = get_freepointer(s, object);
986 max_objects = order_objects(compound_order(page), s->size);
987 if (max_objects > MAX_OBJS_PER_PAGE)
988 max_objects = MAX_OBJS_PER_PAGE;
990 if (page->objects != max_objects) {
991 slab_err(s, page, "Wrong number of objects. Found %d but should be %d",
992 page->objects, max_objects);
993 page->objects = max_objects;
994 slab_fix(s, "Number of objects adjusted.");
996 if (page->inuse != page->objects - nr) {
997 slab_err(s, page, "Wrong object count. Counter is %d but counted were %d",
998 page->inuse, page->objects - nr);
999 page->inuse = page->objects - nr;
1000 slab_fix(s, "Object count adjusted.");
1002 return search == NULL;
1005 static void trace(struct kmem_cache *s, struct page *page, void *object,
1008 if (s->flags & SLAB_TRACE) {
1009 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1011 alloc ? "alloc" : "free",
1012 object, page->inuse,
1016 print_section(KERN_INFO, "Object ", (void *)object,
1024 * Tracking of fully allocated slabs for debugging purposes.
1026 static void add_full(struct kmem_cache *s,
1027 struct kmem_cache_node *n, struct page *page)
1029 if (!(s->flags & SLAB_STORE_USER))
1032 lockdep_assert_held(&n->list_lock);
1033 list_add(&page->lru, &n->full);
1036 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
1038 if (!(s->flags & SLAB_STORE_USER))
1041 lockdep_assert_held(&n->list_lock);
1042 list_del(&page->lru);
1045 /* Tracking of the number of slabs for debugging purposes */
1046 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1048 struct kmem_cache_node *n = get_node(s, node);
1050 return atomic_long_read(&n->nr_slabs);
1053 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1055 return atomic_long_read(&n->nr_slabs);
1058 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1060 struct kmem_cache_node *n = get_node(s, node);
1063 * May be called early in order to allocate a slab for the
1064 * kmem_cache_node structure. Solve the chicken-egg
1065 * dilemma by deferring the increment of the count during
1066 * bootstrap (see early_kmem_cache_node_alloc).
1069 atomic_long_inc(&n->nr_slabs);
1070 atomic_long_add(objects, &n->total_objects);
1073 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1075 struct kmem_cache_node *n = get_node(s, node);
1077 atomic_long_dec(&n->nr_slabs);
1078 atomic_long_sub(objects, &n->total_objects);
1081 /* Object debug checks for alloc/free paths */
1082 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1085 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1088 init_object(s, object, SLUB_RED_INACTIVE);
1089 init_tracking(s, object);
1092 static inline int alloc_consistency_checks(struct kmem_cache *s,
1094 void *object, unsigned long addr)
1096 if (!check_slab(s, page))
1099 if (!check_valid_pointer(s, page, object)) {
1100 object_err(s, page, object, "Freelist Pointer check fails");
1104 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1110 static noinline int alloc_debug_processing(struct kmem_cache *s,
1112 void *object, unsigned long addr)
1114 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1115 if (!alloc_consistency_checks(s, page, object, addr))
1119 /* Success perform special debug activities for allocs */
1120 if (s->flags & SLAB_STORE_USER)
1121 set_track(s, object, TRACK_ALLOC, addr);
1122 trace(s, page, object, 1);
1123 init_object(s, object, SLUB_RED_ACTIVE);
1127 if (PageSlab(page)) {
1129 * If this is a slab page then lets do the best we can
1130 * to avoid issues in the future. Marking all objects
1131 * as used avoids touching the remaining objects.
1133 slab_fix(s, "Marking all objects used");
1134 page->inuse = page->objects;
1135 page->freelist = NULL;
1140 static inline int free_consistency_checks(struct kmem_cache *s,
1141 struct page *page, void *object, unsigned long addr)
1143 if (!check_valid_pointer(s, page, object)) {
1144 slab_err(s, page, "Invalid object pointer 0x%p", object);
1148 if (on_freelist(s, page, object)) {
1149 object_err(s, page, object, "Object already free");
1153 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1156 if (unlikely(s != page->slab_cache)) {
1157 if (!PageSlab(page)) {
1158 slab_err(s, page, "Attempt to free object(0x%p) outside of slab",
1160 } else if (!page->slab_cache) {
1161 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1165 object_err(s, page, object,
1166 "page slab pointer corrupt.");
1172 /* Supports checking bulk free of a constructed freelist */
1173 static noinline int free_debug_processing(
1174 struct kmem_cache *s, struct page *page,
1175 void *head, void *tail, int bulk_cnt,
1178 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1179 void *object = head;
1181 unsigned long uninitialized_var(flags);
1184 spin_lock_irqsave(&n->list_lock, flags);
1187 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1188 if (!check_slab(s, page))
1195 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1196 if (!free_consistency_checks(s, page, object, addr))
1200 if (s->flags & SLAB_STORE_USER)
1201 set_track(s, object, TRACK_FREE, addr);
1202 trace(s, page, object, 0);
1203 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1204 init_object(s, object, SLUB_RED_INACTIVE);
1206 /* Reached end of constructed freelist yet? */
1207 if (object != tail) {
1208 object = get_freepointer(s, object);
1214 if (cnt != bulk_cnt)
1215 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1219 spin_unlock_irqrestore(&n->list_lock, flags);
1221 slab_fix(s, "Object at 0x%p not freed", object);
1225 static int __init setup_slub_debug(char *str)
1227 slub_debug = DEBUG_DEFAULT_FLAGS;
1228 if (*str++ != '=' || !*str)
1230 * No options specified. Switch on full debugging.
1236 * No options but restriction on slabs. This means full
1237 * debugging for slabs matching a pattern.
1244 * Switch off all debugging measures.
1249 * Determine which debug features should be switched on
1251 for (; *str && *str != ','; str++) {
1252 switch (tolower(*str)) {
1254 slub_debug |= SLAB_CONSISTENCY_CHECKS;
1257 slub_debug |= SLAB_RED_ZONE;
1260 slub_debug |= SLAB_POISON;
1263 slub_debug |= SLAB_STORE_USER;
1266 slub_debug |= SLAB_TRACE;
1269 slub_debug |= SLAB_FAILSLAB;
1273 * Avoid enabling debugging on caches if its minimum
1274 * order would increase as a result.
1276 disable_higher_order_debug = 1;
1279 pr_err("slub_debug option '%c' unknown. skipped\n",
1286 slub_debug_slabs = str + 1;
1291 __setup("slub_debug", setup_slub_debug);
1293 slab_flags_t kmem_cache_flags(unsigned int object_size,
1294 slab_flags_t flags, const char *name,
1295 void (*ctor)(void *))
1298 * Enable debugging if selected on the kernel commandline.
1300 if (slub_debug && (!slub_debug_slabs || (name &&
1301 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1302 flags |= slub_debug;
1306 #else /* !CONFIG_SLUB_DEBUG */
1307 static inline void setup_object_debug(struct kmem_cache *s,
1308 struct page *page, void *object) {}
1310 static inline int alloc_debug_processing(struct kmem_cache *s,
1311 struct page *page, void *object, unsigned long addr) { return 0; }
1313 static inline int free_debug_processing(
1314 struct kmem_cache *s, struct page *page,
1315 void *head, void *tail, int bulk_cnt,
1316 unsigned long addr) { return 0; }
1318 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1320 static inline int check_object(struct kmem_cache *s, struct page *page,
1321 void *object, u8 val) { return 1; }
1322 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1323 struct page *page) {}
1324 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1325 struct page *page) {}
1326 slab_flags_t kmem_cache_flags(unsigned int object_size,
1327 slab_flags_t flags, const char *name,
1328 void (*ctor)(void *))
1332 #define slub_debug 0
1334 #define disable_higher_order_debug 0
1336 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1338 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1340 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1342 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1345 static bool freelist_corrupted(struct kmem_cache *s, struct page *page,
1346 void **freelist, void *nextfree)
1350 #endif /* CONFIG_SLUB_DEBUG */
1353 * Hooks for other subsystems that check memory allocations. In a typical
1354 * production configuration these hooks all should produce no code at all.
1356 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1358 kmemleak_alloc(ptr, size, 1, flags);
1359 kasan_kmalloc_large(ptr, size, flags);
1362 static __always_inline void kfree_hook(void *x)
1365 kasan_kfree_large(x, _RET_IP_);
1368 static __always_inline bool slab_free_hook(struct kmem_cache *s, void *x)
1370 kmemleak_free_recursive(x, s->flags);
1373 * Trouble is that we may no longer disable interrupts in the fast path
1374 * So in order to make the debug calls that expect irqs to be
1375 * disabled we need to disable interrupts temporarily.
1377 #ifdef CONFIG_LOCKDEP
1379 unsigned long flags;
1381 local_irq_save(flags);
1382 debug_check_no_locks_freed(x, s->object_size);
1383 local_irq_restore(flags);
1386 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1387 debug_check_no_obj_freed(x, s->object_size);
1389 /* KASAN might put x into memory quarantine, delaying its reuse */
1390 return kasan_slab_free(s, x, _RET_IP_);
1393 static inline bool slab_free_freelist_hook(struct kmem_cache *s,
1394 void **head, void **tail)
1397 * Compiler cannot detect this function can be removed if slab_free_hook()
1398 * evaluates to nothing. Thus, catch all relevant config debug options here.
1400 #if defined(CONFIG_LOCKDEP) || \
1401 defined(CONFIG_DEBUG_KMEMLEAK) || \
1402 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1403 defined(CONFIG_KASAN)
1407 void *old_tail = *tail ? *tail : *head;
1409 /* Head and tail of the reconstructed freelist */
1415 next = get_freepointer(s, object);
1416 /* If object's reuse doesn't have to be delayed */
1417 if (!slab_free_hook(s, object)) {
1418 /* Move object to the new freelist */
1419 set_freepointer(s, object, *head);
1424 } while (object != old_tail);
1429 return *head != NULL;
1435 static void setup_object(struct kmem_cache *s, struct page *page,
1438 setup_object_debug(s, page, object);
1439 kasan_init_slab_obj(s, object);
1440 if (unlikely(s->ctor)) {
1441 kasan_unpoison_object_data(s, object);
1443 kasan_poison_object_data(s, object);
1448 * Slab allocation and freeing
1450 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1451 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1454 unsigned int order = oo_order(oo);
1456 if (node == NUMA_NO_NODE)
1457 page = alloc_pages(flags, order);
1459 page = __alloc_pages_node(node, flags, order);
1461 if (page && memcg_charge_slab(page, flags, order, s)) {
1462 __free_pages(page, order);
1469 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1470 /* Pre-initialize the random sequence cache */
1471 static int init_cache_random_seq(struct kmem_cache *s)
1473 unsigned int count = oo_objects(s->oo);
1476 /* Bailout if already initialised */
1480 err = cache_random_seq_create(s, count, GFP_KERNEL);
1482 pr_err("SLUB: Unable to initialize free list for %s\n",
1487 /* Transform to an offset on the set of pages */
1488 if (s->random_seq) {
1491 for (i = 0; i < count; i++)
1492 s->random_seq[i] *= s->size;
1497 /* Initialize each random sequence freelist per cache */
1498 static void __init init_freelist_randomization(void)
1500 struct kmem_cache *s;
1502 mutex_lock(&slab_mutex);
1504 list_for_each_entry(s, &slab_caches, list)
1505 init_cache_random_seq(s);
1507 mutex_unlock(&slab_mutex);
1510 /* Get the next entry on the pre-computed freelist randomized */
1511 static void *next_freelist_entry(struct kmem_cache *s, struct page *page,
1512 unsigned long *pos, void *start,
1513 unsigned long page_limit,
1514 unsigned long freelist_count)
1519 * If the target page allocation failed, the number of objects on the
1520 * page might be smaller than the usual size defined by the cache.
1523 idx = s->random_seq[*pos];
1525 if (*pos >= freelist_count)
1527 } while (unlikely(idx >= page_limit));
1529 return (char *)start + idx;
1532 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1533 static bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1538 unsigned long idx, pos, page_limit, freelist_count;
1540 if (page->objects < 2 || !s->random_seq)
1543 freelist_count = oo_objects(s->oo);
1544 pos = get_random_int() % freelist_count;
1546 page_limit = page->objects * s->size;
1547 start = fixup_red_left(s, page_address(page));
1549 /* First entry is used as the base of the freelist */
1550 cur = next_freelist_entry(s, page, &pos, start, page_limit,
1552 page->freelist = cur;
1554 for (idx = 1; idx < page->objects; idx++) {
1555 setup_object(s, page, cur);
1556 next = next_freelist_entry(s, page, &pos, start, page_limit,
1558 set_freepointer(s, cur, next);
1561 setup_object(s, page, cur);
1562 set_freepointer(s, cur, NULL);
1567 static inline int init_cache_random_seq(struct kmem_cache *s)
1571 static inline void init_freelist_randomization(void) { }
1572 static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1576 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1578 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1581 struct kmem_cache_order_objects oo = s->oo;
1587 flags &= gfp_allowed_mask;
1589 if (gfpflags_allow_blocking(flags))
1592 flags |= s->allocflags;
1595 * Let the initial higher-order allocation fail under memory pressure
1596 * so we fall-back to the minimum order allocation.
1598 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1599 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1600 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
1602 page = alloc_slab_page(s, alloc_gfp, node, oo);
1603 if (unlikely(!page)) {
1607 * Allocation may have failed due to fragmentation.
1608 * Try a lower order alloc if possible
1610 page = alloc_slab_page(s, alloc_gfp, node, oo);
1611 if (unlikely(!page))
1613 stat(s, ORDER_FALLBACK);
1616 page->objects = oo_objects(oo);
1618 order = compound_order(page);
1619 page->slab_cache = s;
1620 __SetPageSlab(page);
1621 if (page_is_pfmemalloc(page))
1622 SetPageSlabPfmemalloc(page);
1624 start = page_address(page);
1626 if (unlikely(s->flags & SLAB_POISON))
1627 memset(start, POISON_INUSE, PAGE_SIZE << order);
1629 kasan_poison_slab(page);
1631 shuffle = shuffle_freelist(s, page);
1634 for_each_object_idx(p, idx, s, start, page->objects) {
1635 setup_object(s, page, p);
1636 if (likely(idx < page->objects))
1637 set_freepointer(s, p, p + s->size);
1639 set_freepointer(s, p, NULL);
1641 page->freelist = fixup_red_left(s, start);
1644 page->inuse = page->objects;
1648 if (gfpflags_allow_blocking(flags))
1649 local_irq_disable();
1653 mod_lruvec_page_state(page,
1654 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1655 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1658 inc_slabs_node(s, page_to_nid(page), page->objects);
1663 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1665 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1666 gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK;
1667 flags &= ~GFP_SLAB_BUG_MASK;
1668 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
1669 invalid_mask, &invalid_mask, flags, &flags);
1673 return allocate_slab(s,
1674 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1677 static void __free_slab(struct kmem_cache *s, struct page *page)
1679 int order = compound_order(page);
1680 int pages = 1 << order;
1682 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1685 slab_pad_check(s, page);
1686 for_each_object(p, s, page_address(page),
1688 check_object(s, page, p, SLUB_RED_INACTIVE);
1691 mod_lruvec_page_state(page,
1692 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1693 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1696 __ClearPageSlabPfmemalloc(page);
1697 __ClearPageSlab(page);
1699 page->mapping = NULL;
1700 if (current->reclaim_state)
1701 current->reclaim_state->reclaimed_slab += pages;
1702 memcg_uncharge_slab(page, order, s);
1703 __free_pages(page, order);
1706 static void rcu_free_slab(struct rcu_head *h)
1708 struct page *page = container_of(h, struct page, rcu_head);
1710 __free_slab(page->slab_cache, page);
1713 static void free_slab(struct kmem_cache *s, struct page *page)
1715 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) {
1716 call_rcu(&page->rcu_head, rcu_free_slab);
1718 __free_slab(s, page);
1721 static void discard_slab(struct kmem_cache *s, struct page *page)
1723 dec_slabs_node(s, page_to_nid(page), page->objects);
1728 * Management of partially allocated slabs.
1731 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1734 if (tail == DEACTIVATE_TO_TAIL)
1735 list_add_tail(&page->lru, &n->partial);
1737 list_add(&page->lru, &n->partial);
1740 static inline void add_partial(struct kmem_cache_node *n,
1741 struct page *page, int tail)
1743 lockdep_assert_held(&n->list_lock);
1744 __add_partial(n, page, tail);
1747 static inline void remove_partial(struct kmem_cache_node *n,
1750 lockdep_assert_held(&n->list_lock);
1751 list_del(&page->lru);
1756 * Remove slab from the partial list, freeze it and
1757 * return the pointer to the freelist.
1759 * Returns a list of objects or NULL if it fails.
1761 static inline void *acquire_slab(struct kmem_cache *s,
1762 struct kmem_cache_node *n, struct page *page,
1763 int mode, int *objects)
1766 unsigned long counters;
1769 lockdep_assert_held(&n->list_lock);
1772 * Zap the freelist and set the frozen bit.
1773 * The old freelist is the list of objects for the
1774 * per cpu allocation list.
1776 freelist = page->freelist;
1777 counters = page->counters;
1778 new.counters = counters;
1779 *objects = new.objects - new.inuse;
1781 new.inuse = page->objects;
1782 new.freelist = NULL;
1784 new.freelist = freelist;
1787 VM_BUG_ON(new.frozen);
1790 if (!__cmpxchg_double_slab(s, page,
1792 new.freelist, new.counters,
1796 remove_partial(n, page);
1801 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1802 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1805 * Try to allocate a partial slab from a specific node.
1807 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1808 struct kmem_cache_cpu *c, gfp_t flags)
1810 struct page *page, *page2;
1811 void *object = NULL;
1812 unsigned int available = 0;
1816 * Racy check. If we mistakenly see no partial slabs then we
1817 * just allocate an empty slab. If we mistakenly try to get a
1818 * partial slab and there is none available then get_partials()
1821 if (!n || !n->nr_partial)
1824 spin_lock(&n->list_lock);
1825 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1828 if (!pfmemalloc_match(page, flags))
1831 t = acquire_slab(s, n, page, object == NULL, &objects);
1835 available += objects;
1838 stat(s, ALLOC_FROM_PARTIAL);
1841 put_cpu_partial(s, page, 0);
1842 stat(s, CPU_PARTIAL_NODE);
1844 if (!kmem_cache_has_cpu_partial(s)
1845 || available > slub_cpu_partial(s) / 2)
1849 spin_unlock(&n->list_lock);
1854 * Get a page from somewhere. Search in increasing NUMA distances.
1856 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1857 struct kmem_cache_cpu *c)
1860 struct zonelist *zonelist;
1863 enum zone_type high_zoneidx = gfp_zone(flags);
1865 unsigned int cpuset_mems_cookie;
1868 * The defrag ratio allows a configuration of the tradeoffs between
1869 * inter node defragmentation and node local allocations. A lower
1870 * defrag_ratio increases the tendency to do local allocations
1871 * instead of attempting to obtain partial slabs from other nodes.
1873 * If the defrag_ratio is set to 0 then kmalloc() always
1874 * returns node local objects. If the ratio is higher then kmalloc()
1875 * may return off node objects because partial slabs are obtained
1876 * from other nodes and filled up.
1878 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
1879 * (which makes defrag_ratio = 1000) then every (well almost)
1880 * allocation will first attempt to defrag slab caches on other nodes.
1881 * This means scanning over all nodes to look for partial slabs which
1882 * may be expensive if we do it every time we are trying to find a slab
1883 * with available objects.
1885 if (!s->remote_node_defrag_ratio ||
1886 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1890 cpuset_mems_cookie = read_mems_allowed_begin();
1891 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1892 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1893 struct kmem_cache_node *n;
1895 n = get_node(s, zone_to_nid(zone));
1897 if (n && cpuset_zone_allowed(zone, flags) &&
1898 n->nr_partial > s->min_partial) {
1899 object = get_partial_node(s, n, c, flags);
1902 * Don't check read_mems_allowed_retry()
1903 * here - if mems_allowed was updated in
1904 * parallel, that was a harmless race
1905 * between allocation and the cpuset
1912 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1918 * Get a partial page, lock it and return it.
1920 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1921 struct kmem_cache_cpu *c)
1924 int searchnode = node;
1926 if (node == NUMA_NO_NODE)
1927 searchnode = numa_mem_id();
1929 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1930 if (object || node != NUMA_NO_NODE)
1933 return get_any_partial(s, flags, c);
1936 #ifdef CONFIG_PREEMPT
1938 * Calculate the next globally unique transaction for disambiguiation
1939 * during cmpxchg. The transactions start with the cpu number and are then
1940 * incremented by CONFIG_NR_CPUS.
1942 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1945 * No preemption supported therefore also no need to check for
1951 static inline unsigned long next_tid(unsigned long tid)
1953 return tid + TID_STEP;
1956 static inline unsigned int tid_to_cpu(unsigned long tid)
1958 return tid % TID_STEP;
1961 static inline unsigned long tid_to_event(unsigned long tid)
1963 return tid / TID_STEP;
1966 static inline unsigned int init_tid(int cpu)
1971 static inline void note_cmpxchg_failure(const char *n,
1972 const struct kmem_cache *s, unsigned long tid)
1974 #ifdef SLUB_DEBUG_CMPXCHG
1975 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1977 pr_info("%s %s: cmpxchg redo ", n, s->name);
1979 #ifdef CONFIG_PREEMPT
1980 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1981 pr_warn("due to cpu change %d -> %d\n",
1982 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1985 if (tid_to_event(tid) != tid_to_event(actual_tid))
1986 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1987 tid_to_event(tid), tid_to_event(actual_tid));
1989 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1990 actual_tid, tid, next_tid(tid));
1992 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1995 static void init_kmem_cache_cpus(struct kmem_cache *s)
1999 for_each_possible_cpu(cpu)
2000 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
2004 * Remove the cpu slab
2006 static void deactivate_slab(struct kmem_cache *s, struct page *page,
2007 void *freelist, struct kmem_cache_cpu *c)
2009 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
2010 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
2012 enum slab_modes l = M_NONE, m = M_NONE;
2014 int tail = DEACTIVATE_TO_HEAD;
2018 if (page->freelist) {
2019 stat(s, DEACTIVATE_REMOTE_FREES);
2020 tail = DEACTIVATE_TO_TAIL;
2024 * Stage one: Free all available per cpu objects back
2025 * to the page freelist while it is still frozen. Leave the
2028 * There is no need to take the list->lock because the page
2031 while (freelist && (nextfree = get_freepointer(s, freelist))) {
2033 unsigned long counters;
2036 * If 'nextfree' is invalid, it is possible that the object at
2037 * 'freelist' is already corrupted. So isolate all objects
2038 * starting at 'freelist'.
2040 if (freelist_corrupted(s, page, &freelist, nextfree))
2044 prior = page->freelist;
2045 counters = page->counters;
2046 set_freepointer(s, freelist, prior);
2047 new.counters = counters;
2049 VM_BUG_ON(!new.frozen);
2051 } while (!__cmpxchg_double_slab(s, page,
2053 freelist, new.counters,
2054 "drain percpu freelist"));
2056 freelist = nextfree;
2060 * Stage two: Ensure that the page is unfrozen while the
2061 * list presence reflects the actual number of objects
2064 * We setup the list membership and then perform a cmpxchg
2065 * with the count. If there is a mismatch then the page
2066 * is not unfrozen but the page is on the wrong list.
2068 * Then we restart the process which may have to remove
2069 * the page from the list that we just put it on again
2070 * because the number of objects in the slab may have
2075 old.freelist = page->freelist;
2076 old.counters = page->counters;
2077 VM_BUG_ON(!old.frozen);
2079 /* Determine target state of the slab */
2080 new.counters = old.counters;
2083 set_freepointer(s, freelist, old.freelist);
2084 new.freelist = freelist;
2086 new.freelist = old.freelist;
2090 if (!new.inuse && n->nr_partial >= s->min_partial)
2092 else if (new.freelist) {
2097 * Taking the spinlock removes the possiblity
2098 * that acquire_slab() will see a slab page that
2101 spin_lock(&n->list_lock);
2105 if (kmem_cache_debug(s) && !lock) {
2108 * This also ensures that the scanning of full
2109 * slabs from diagnostic functions will not see
2112 spin_lock(&n->list_lock);
2120 remove_partial(n, page);
2122 else if (l == M_FULL)
2124 remove_full(s, n, page);
2126 if (m == M_PARTIAL) {
2128 add_partial(n, page, tail);
2131 } else if (m == M_FULL) {
2133 stat(s, DEACTIVATE_FULL);
2134 add_full(s, n, page);
2140 if (!__cmpxchg_double_slab(s, page,
2141 old.freelist, old.counters,
2142 new.freelist, new.counters,
2147 spin_unlock(&n->list_lock);
2150 stat(s, DEACTIVATE_EMPTY);
2151 discard_slab(s, page);
2160 * Unfreeze all the cpu partial slabs.
2162 * This function must be called with interrupts disabled
2163 * for the cpu using c (or some other guarantee must be there
2164 * to guarantee no concurrent accesses).
2166 static void unfreeze_partials(struct kmem_cache *s,
2167 struct kmem_cache_cpu *c)
2169 #ifdef CONFIG_SLUB_CPU_PARTIAL
2170 struct kmem_cache_node *n = NULL, *n2 = NULL;
2171 struct page *page, *discard_page = NULL;
2173 while ((page = c->partial)) {
2177 c->partial = page->next;
2179 n2 = get_node(s, page_to_nid(page));
2182 spin_unlock(&n->list_lock);
2185 spin_lock(&n->list_lock);
2190 old.freelist = page->freelist;
2191 old.counters = page->counters;
2192 VM_BUG_ON(!old.frozen);
2194 new.counters = old.counters;
2195 new.freelist = old.freelist;
2199 } while (!__cmpxchg_double_slab(s, page,
2200 old.freelist, old.counters,
2201 new.freelist, new.counters,
2202 "unfreezing slab"));
2204 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2205 page->next = discard_page;
2206 discard_page = page;
2208 add_partial(n, page, DEACTIVATE_TO_TAIL);
2209 stat(s, FREE_ADD_PARTIAL);
2214 spin_unlock(&n->list_lock);
2216 while (discard_page) {
2217 page = discard_page;
2218 discard_page = discard_page->next;
2220 stat(s, DEACTIVATE_EMPTY);
2221 discard_slab(s, page);
2228 * Put a page that was just frozen (in __slab_free) into a partial page
2229 * slot if available.
2231 * If we did not find a slot then simply move all the partials to the
2232 * per node partial list.
2234 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2236 #ifdef CONFIG_SLUB_CPU_PARTIAL
2237 struct page *oldpage;
2245 oldpage = this_cpu_read(s->cpu_slab->partial);
2248 pobjects = oldpage->pobjects;
2249 pages = oldpage->pages;
2250 if (drain && pobjects > s->cpu_partial) {
2251 unsigned long flags;
2253 * partial array is full. Move the existing
2254 * set to the per node partial list.
2256 local_irq_save(flags);
2257 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2258 local_irq_restore(flags);
2262 stat(s, CPU_PARTIAL_DRAIN);
2267 pobjects += page->objects - page->inuse;
2269 page->pages = pages;
2270 page->pobjects = pobjects;
2271 page->next = oldpage;
2273 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2275 if (unlikely(!s->cpu_partial)) {
2276 unsigned long flags;
2278 local_irq_save(flags);
2279 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2280 local_irq_restore(flags);
2286 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2288 stat(s, CPUSLAB_FLUSH);
2289 deactivate_slab(s, c->page, c->freelist, c);
2291 c->tid = next_tid(c->tid);
2297 * Called from IPI handler with interrupts disabled.
2299 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2301 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2307 unfreeze_partials(s, c);
2311 static void flush_cpu_slab(void *d)
2313 struct kmem_cache *s = d;
2315 __flush_cpu_slab(s, smp_processor_id());
2318 static bool has_cpu_slab(int cpu, void *info)
2320 struct kmem_cache *s = info;
2321 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2323 return c->page || slub_percpu_partial(c);
2326 static void flush_all(struct kmem_cache *s)
2328 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2332 * Use the cpu notifier to insure that the cpu slabs are flushed when
2335 static int slub_cpu_dead(unsigned int cpu)
2337 struct kmem_cache *s;
2338 unsigned long flags;
2340 mutex_lock(&slab_mutex);
2341 list_for_each_entry(s, &slab_caches, list) {
2342 local_irq_save(flags);
2343 __flush_cpu_slab(s, cpu);
2344 local_irq_restore(flags);
2346 mutex_unlock(&slab_mutex);
2351 * Check if the objects in a per cpu structure fit numa
2352 * locality expectations.
2354 static inline int node_match(struct page *page, int node)
2357 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2363 #ifdef CONFIG_SLUB_DEBUG
2364 static int count_free(struct page *page)
2366 return page->objects - page->inuse;
2369 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2371 return atomic_long_read(&n->total_objects);
2373 #endif /* CONFIG_SLUB_DEBUG */
2375 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2376 static unsigned long count_partial(struct kmem_cache_node *n,
2377 int (*get_count)(struct page *))
2379 unsigned long flags;
2380 unsigned long x = 0;
2383 spin_lock_irqsave(&n->list_lock, flags);
2384 list_for_each_entry(page, &n->partial, lru)
2385 x += get_count(page);
2386 spin_unlock_irqrestore(&n->list_lock, flags);
2389 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2391 static noinline void
2392 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2394 #ifdef CONFIG_SLUB_DEBUG
2395 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2396 DEFAULT_RATELIMIT_BURST);
2398 struct kmem_cache_node *n;
2400 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2403 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2404 nid, gfpflags, &gfpflags);
2405 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2406 s->name, s->object_size, s->size, oo_order(s->oo),
2409 if (oo_order(s->min) > get_order(s->object_size))
2410 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2413 for_each_kmem_cache_node(s, node, n) {
2414 unsigned long nr_slabs;
2415 unsigned long nr_objs;
2416 unsigned long nr_free;
2418 nr_free = count_partial(n, count_free);
2419 nr_slabs = node_nr_slabs(n);
2420 nr_objs = node_nr_objs(n);
2422 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2423 node, nr_slabs, nr_objs, nr_free);
2428 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2429 int node, struct kmem_cache_cpu **pc)
2432 struct kmem_cache_cpu *c = *pc;
2435 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
2437 freelist = get_partial(s, flags, node, c);
2442 page = new_slab(s, flags, node);
2444 c = raw_cpu_ptr(s->cpu_slab);
2449 * No other reference to the page yet so we can
2450 * muck around with it freely without cmpxchg
2452 freelist = page->freelist;
2453 page->freelist = NULL;
2455 stat(s, ALLOC_SLAB);
2464 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2466 if (unlikely(PageSlabPfmemalloc(page)))
2467 return gfp_pfmemalloc_allowed(gfpflags);
2473 * Check the page->freelist of a page and either transfer the freelist to the
2474 * per cpu freelist or deactivate the page.
2476 * The page is still frozen if the return value is not NULL.
2478 * If this function returns NULL then the page has been unfrozen.
2480 * This function must be called with interrupt disabled.
2482 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2485 unsigned long counters;
2489 freelist = page->freelist;
2490 counters = page->counters;
2492 new.counters = counters;
2493 VM_BUG_ON(!new.frozen);
2495 new.inuse = page->objects;
2496 new.frozen = freelist != NULL;
2498 } while (!__cmpxchg_double_slab(s, page,
2507 * Slow path. The lockless freelist is empty or we need to perform
2510 * Processing is still very fast if new objects have been freed to the
2511 * regular freelist. In that case we simply take over the regular freelist
2512 * as the lockless freelist and zap the regular freelist.
2514 * If that is not working then we fall back to the partial lists. We take the
2515 * first element of the freelist as the object to allocate now and move the
2516 * rest of the freelist to the lockless freelist.
2518 * And if we were unable to get a new slab from the partial slab lists then
2519 * we need to allocate a new slab. This is the slowest path since it involves
2520 * a call to the page allocator and the setup of a new slab.
2522 * Version of __slab_alloc to use when we know that interrupts are
2523 * already disabled (which is the case for bulk allocation).
2525 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2526 unsigned long addr, struct kmem_cache_cpu *c)
2534 * if the node is not online or has no normal memory, just
2535 * ignore the node constraint
2537 if (unlikely(node != NUMA_NO_NODE &&
2538 !node_state(node, N_NORMAL_MEMORY)))
2539 node = NUMA_NO_NODE;
2544 if (unlikely(!node_match(page, node))) {
2546 * same as above but node_match() being false already
2547 * implies node != NUMA_NO_NODE
2549 if (!node_state(node, N_NORMAL_MEMORY)) {
2550 node = NUMA_NO_NODE;
2553 stat(s, ALLOC_NODE_MISMATCH);
2554 deactivate_slab(s, page, c->freelist, c);
2560 * By rights, we should be searching for a slab page that was
2561 * PFMEMALLOC but right now, we are losing the pfmemalloc
2562 * information when the page leaves the per-cpu allocator
2564 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2565 deactivate_slab(s, page, c->freelist, c);
2569 /* must check again c->freelist in case of cpu migration or IRQ */
2570 freelist = c->freelist;
2574 freelist = get_freelist(s, page);
2578 stat(s, DEACTIVATE_BYPASS);
2582 stat(s, ALLOC_REFILL);
2586 * freelist is pointing to the list of objects to be used.
2587 * page is pointing to the page from which the objects are obtained.
2588 * That page must be frozen for per cpu allocations to work.
2590 VM_BUG_ON(!c->page->frozen);
2591 c->freelist = get_freepointer(s, freelist);
2592 c->tid = next_tid(c->tid);
2597 if (slub_percpu_partial(c)) {
2598 page = c->page = slub_percpu_partial(c);
2599 slub_set_percpu_partial(c, page);
2600 stat(s, CPU_PARTIAL_ALLOC);
2604 freelist = new_slab_objects(s, gfpflags, node, &c);
2606 if (unlikely(!freelist)) {
2607 slab_out_of_memory(s, gfpflags, node);
2612 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2615 /* Only entered in the debug case */
2616 if (kmem_cache_debug(s) &&
2617 !alloc_debug_processing(s, page, freelist, addr))
2618 goto new_slab; /* Slab failed checks. Next slab needed */
2620 deactivate_slab(s, page, get_freepointer(s, freelist), c);
2625 * Another one that disabled interrupt and compensates for possible
2626 * cpu changes by refetching the per cpu area pointer.
2628 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2629 unsigned long addr, struct kmem_cache_cpu *c)
2632 unsigned long flags;
2634 local_irq_save(flags);
2635 #ifdef CONFIG_PREEMPT
2637 * We may have been preempted and rescheduled on a different
2638 * cpu before disabling interrupts. Need to reload cpu area
2641 c = this_cpu_ptr(s->cpu_slab);
2644 p = ___slab_alloc(s, gfpflags, node, addr, c);
2645 local_irq_restore(flags);
2650 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2651 * have the fastpath folded into their functions. So no function call
2652 * overhead for requests that can be satisfied on the fastpath.
2654 * The fastpath works by first checking if the lockless freelist can be used.
2655 * If not then __slab_alloc is called for slow processing.
2657 * Otherwise we can simply pick the next object from the lockless free list.
2659 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2660 gfp_t gfpflags, int node, unsigned long addr)
2663 struct kmem_cache_cpu *c;
2667 s = slab_pre_alloc_hook(s, gfpflags);
2672 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2673 * enabled. We may switch back and forth between cpus while
2674 * reading from one cpu area. That does not matter as long
2675 * as we end up on the original cpu again when doing the cmpxchg.
2677 * We should guarantee that tid and kmem_cache are retrieved on
2678 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2679 * to check if it is matched or not.
2682 tid = this_cpu_read(s->cpu_slab->tid);
2683 c = raw_cpu_ptr(s->cpu_slab);
2684 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2685 unlikely(tid != READ_ONCE(c->tid)));
2688 * Irqless object alloc/free algorithm used here depends on sequence
2689 * of fetching cpu_slab's data. tid should be fetched before anything
2690 * on c to guarantee that object and page associated with previous tid
2691 * won't be used with current tid. If we fetch tid first, object and
2692 * page could be one associated with next tid and our alloc/free
2693 * request will be failed. In this case, we will retry. So, no problem.
2698 * The transaction ids are globally unique per cpu and per operation on
2699 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2700 * occurs on the right processor and that there was no operation on the
2701 * linked list in between.
2704 object = c->freelist;
2706 if (unlikely(!object || !node_match(page, node))) {
2707 object = __slab_alloc(s, gfpflags, node, addr, c);
2708 stat(s, ALLOC_SLOWPATH);
2710 void *next_object = get_freepointer_safe(s, object);
2713 * The cmpxchg will only match if there was no additional
2714 * operation and if we are on the right processor.
2716 * The cmpxchg does the following atomically (without lock
2718 * 1. Relocate first pointer to the current per cpu area.
2719 * 2. Verify that tid and freelist have not been changed
2720 * 3. If they were not changed replace tid and freelist
2722 * Since this is without lock semantics the protection is only
2723 * against code executing on this cpu *not* from access by
2726 if (unlikely(!this_cpu_cmpxchg_double(
2727 s->cpu_slab->freelist, s->cpu_slab->tid,
2729 next_object, next_tid(tid)))) {
2731 note_cmpxchg_failure("slab_alloc", s, tid);
2734 prefetch_freepointer(s, next_object);
2735 stat(s, ALLOC_FASTPATH);
2738 if (unlikely(gfpflags & __GFP_ZERO) && object)
2739 memset(object, 0, s->object_size);
2741 slab_post_alloc_hook(s, gfpflags, 1, &object);
2746 static __always_inline void *slab_alloc(struct kmem_cache *s,
2747 gfp_t gfpflags, unsigned long addr)
2749 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2752 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2754 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2756 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2761 EXPORT_SYMBOL(kmem_cache_alloc);
2763 #ifdef CONFIG_TRACING
2764 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2766 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2767 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2768 kasan_kmalloc(s, ret, size, gfpflags);
2771 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2775 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2777 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2779 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2780 s->object_size, s->size, gfpflags, node);
2784 EXPORT_SYMBOL(kmem_cache_alloc_node);
2786 #ifdef CONFIG_TRACING
2787 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2789 int node, size_t size)
2791 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2793 trace_kmalloc_node(_RET_IP_, ret,
2794 size, s->size, gfpflags, node);
2796 kasan_kmalloc(s, ret, size, gfpflags);
2799 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2804 * Slow path handling. This may still be called frequently since objects
2805 * have a longer lifetime than the cpu slabs in most processing loads.
2807 * So we still attempt to reduce cache line usage. Just take the slab
2808 * lock and free the item. If there is no additional partial page
2809 * handling required then we can return immediately.
2811 static void __slab_free(struct kmem_cache *s, struct page *page,
2812 void *head, void *tail, int cnt,
2819 unsigned long counters;
2820 struct kmem_cache_node *n = NULL;
2821 unsigned long uninitialized_var(flags);
2823 stat(s, FREE_SLOWPATH);
2825 if (kmem_cache_debug(s) &&
2826 !free_debug_processing(s, page, head, tail, cnt, addr))
2831 spin_unlock_irqrestore(&n->list_lock, flags);
2834 prior = page->freelist;
2835 counters = page->counters;
2836 set_freepointer(s, tail, prior);
2837 new.counters = counters;
2838 was_frozen = new.frozen;
2840 if ((!new.inuse || !prior) && !was_frozen) {
2842 if (kmem_cache_has_cpu_partial(s) && !prior) {
2845 * Slab was on no list before and will be
2847 * We can defer the list move and instead
2852 } else { /* Needs to be taken off a list */
2854 n = get_node(s, page_to_nid(page));
2856 * Speculatively acquire the list_lock.
2857 * If the cmpxchg does not succeed then we may
2858 * drop the list_lock without any processing.
2860 * Otherwise the list_lock will synchronize with
2861 * other processors updating the list of slabs.
2863 spin_lock_irqsave(&n->list_lock, flags);
2868 } while (!cmpxchg_double_slab(s, page,
2876 * If we just froze the page then put it onto the
2877 * per cpu partial list.
2879 if (new.frozen && !was_frozen) {
2880 put_cpu_partial(s, page, 1);
2881 stat(s, CPU_PARTIAL_FREE);
2884 * The list lock was not taken therefore no list
2885 * activity can be necessary.
2888 stat(s, FREE_FROZEN);
2892 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2896 * Objects left in the slab. If it was not on the partial list before
2899 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2900 if (kmem_cache_debug(s))
2901 remove_full(s, n, page);
2902 add_partial(n, page, DEACTIVATE_TO_TAIL);
2903 stat(s, FREE_ADD_PARTIAL);
2905 spin_unlock_irqrestore(&n->list_lock, flags);
2911 * Slab on the partial list.
2913 remove_partial(n, page);
2914 stat(s, FREE_REMOVE_PARTIAL);
2916 /* Slab must be on the full list */
2917 remove_full(s, n, page);
2920 spin_unlock_irqrestore(&n->list_lock, flags);
2922 discard_slab(s, page);
2926 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2927 * can perform fastpath freeing without additional function calls.
2929 * The fastpath is only possible if we are freeing to the current cpu slab
2930 * of this processor. This typically the case if we have just allocated
2933 * If fastpath is not possible then fall back to __slab_free where we deal
2934 * with all sorts of special processing.
2936 * Bulk free of a freelist with several objects (all pointing to the
2937 * same page) possible by specifying head and tail ptr, plus objects
2938 * count (cnt). Bulk free indicated by tail pointer being set.
2940 static __always_inline void do_slab_free(struct kmem_cache *s,
2941 struct page *page, void *head, void *tail,
2942 int cnt, unsigned long addr)
2944 void *tail_obj = tail ? : head;
2945 struct kmem_cache_cpu *c;
2949 * Determine the currently cpus per cpu slab.
2950 * The cpu may change afterward. However that does not matter since
2951 * data is retrieved via this pointer. If we are on the same cpu
2952 * during the cmpxchg then the free will succeed.
2955 tid = this_cpu_read(s->cpu_slab->tid);
2956 c = raw_cpu_ptr(s->cpu_slab);
2957 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2958 unlikely(tid != READ_ONCE(c->tid)));
2960 /* Same with comment on barrier() in slab_alloc_node() */
2963 if (likely(page == c->page)) {
2964 void **freelist = READ_ONCE(c->freelist);
2966 set_freepointer(s, tail_obj, freelist);
2968 if (unlikely(!this_cpu_cmpxchg_double(
2969 s->cpu_slab->freelist, s->cpu_slab->tid,
2971 head, next_tid(tid)))) {
2973 note_cmpxchg_failure("slab_free", s, tid);
2976 stat(s, FREE_FASTPATH);
2978 __slab_free(s, page, head, tail_obj, cnt, addr);
2982 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
2983 void *head, void *tail, int cnt,
2987 * With KASAN enabled slab_free_freelist_hook modifies the freelist
2988 * to remove objects, whose reuse must be delayed.
2990 if (slab_free_freelist_hook(s, &head, &tail))
2991 do_slab_free(s, page, head, tail, cnt, addr);
2995 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
2997 do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr);
3001 void kmem_cache_free(struct kmem_cache *s, void *x)
3003 s = cache_from_obj(s, x);
3006 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
3007 trace_kmem_cache_free(_RET_IP_, x);
3009 EXPORT_SYMBOL(kmem_cache_free);
3011 struct detached_freelist {
3016 struct kmem_cache *s;
3020 * This function progressively scans the array with free objects (with
3021 * a limited look ahead) and extract objects belonging to the same
3022 * page. It builds a detached freelist directly within the given
3023 * page/objects. This can happen without any need for
3024 * synchronization, because the objects are owned by running process.
3025 * The freelist is build up as a single linked list in the objects.
3026 * The idea is, that this detached freelist can then be bulk
3027 * transferred to the real freelist(s), but only requiring a single
3028 * synchronization primitive. Look ahead in the array is limited due
3029 * to performance reasons.
3032 int build_detached_freelist(struct kmem_cache *s, size_t size,
3033 void **p, struct detached_freelist *df)
3035 size_t first_skipped_index = 0;
3040 /* Always re-init detached_freelist */
3045 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3046 } while (!object && size);
3051 page = virt_to_head_page(object);
3053 /* Handle kalloc'ed objects */
3054 if (unlikely(!PageSlab(page))) {
3055 BUG_ON(!PageCompound(page));
3057 __free_pages(page, compound_order(page));
3058 p[size] = NULL; /* mark object processed */
3061 /* Derive kmem_cache from object */
3062 df->s = page->slab_cache;
3064 df->s = cache_from_obj(s, object); /* Support for memcg */
3067 /* Start new detached freelist */
3069 set_freepointer(df->s, object, NULL);
3071 df->freelist = object;
3072 p[size] = NULL; /* mark object processed */
3078 continue; /* Skip processed objects */
3080 /* df->page is always set at this point */
3081 if (df->page == virt_to_head_page(object)) {
3082 /* Opportunity build freelist */
3083 set_freepointer(df->s, object, df->freelist);
3084 df->freelist = object;
3086 p[size] = NULL; /* mark object processed */
3091 /* Limit look ahead search */
3095 if (!first_skipped_index)
3096 first_skipped_index = size + 1;
3099 return first_skipped_index;
3102 /* Note that interrupts must be enabled when calling this function. */
3103 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3109 struct detached_freelist df;
3111 size = build_detached_freelist(s, size, p, &df);
3115 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
3116 } while (likely(size));
3118 EXPORT_SYMBOL(kmem_cache_free_bulk);
3120 /* Note that interrupts must be enabled when calling this function. */
3121 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3124 struct kmem_cache_cpu *c;
3127 /* memcg and kmem_cache debug support */
3128 s = slab_pre_alloc_hook(s, flags);
3132 * Drain objects in the per cpu slab, while disabling local
3133 * IRQs, which protects against PREEMPT and interrupts
3134 * handlers invoking normal fastpath.
3136 local_irq_disable();
3137 c = this_cpu_ptr(s->cpu_slab);
3139 for (i = 0; i < size; i++) {
3140 void *object = c->freelist;
3142 if (unlikely(!object)) {
3144 * We may have removed an object from c->freelist using
3145 * the fastpath in the previous iteration; in that case,
3146 * c->tid has not been bumped yet.
3147 * Since ___slab_alloc() may reenable interrupts while
3148 * allocating memory, we should bump c->tid now.
3150 c->tid = next_tid(c->tid);
3153 * Invoking slow path likely have side-effect
3154 * of re-populating per CPU c->freelist
3156 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3158 if (unlikely(!p[i]))
3161 c = this_cpu_ptr(s->cpu_slab);
3162 continue; /* goto for-loop */
3164 c->freelist = get_freepointer(s, object);
3167 c->tid = next_tid(c->tid);
3170 /* Clear memory outside IRQ disabled fastpath loop */
3171 if (unlikely(flags & __GFP_ZERO)) {
3174 for (j = 0; j < i; j++)
3175 memset(p[j], 0, s->object_size);
3178 /* memcg and kmem_cache debug support */
3179 slab_post_alloc_hook(s, flags, size, p);
3183 slab_post_alloc_hook(s, flags, i, p);
3184 __kmem_cache_free_bulk(s, i, p);
3187 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3191 * Object placement in a slab is made very easy because we always start at
3192 * offset 0. If we tune the size of the object to the alignment then we can
3193 * get the required alignment by putting one properly sized object after
3196 * Notice that the allocation order determines the sizes of the per cpu
3197 * caches. Each processor has always one slab available for allocations.
3198 * Increasing the allocation order reduces the number of times that slabs
3199 * must be moved on and off the partial lists and is therefore a factor in
3204 * Mininum / Maximum order of slab pages. This influences locking overhead
3205 * and slab fragmentation. A higher order reduces the number of partial slabs
3206 * and increases the number of allocations possible without having to
3207 * take the list_lock.
3209 static unsigned int slub_min_order;
3210 static unsigned int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3211 static unsigned int slub_min_objects;
3214 * Calculate the order of allocation given an slab object size.
3216 * The order of allocation has significant impact on performance and other
3217 * system components. Generally order 0 allocations should be preferred since
3218 * order 0 does not cause fragmentation in the page allocator. Larger objects
3219 * be problematic to put into order 0 slabs because there may be too much
3220 * unused space left. We go to a higher order if more than 1/16th of the slab
3223 * In order to reach satisfactory performance we must ensure that a minimum
3224 * number of objects is in one slab. Otherwise we may generate too much
3225 * activity on the partial lists which requires taking the list_lock. This is
3226 * less a concern for large slabs though which are rarely used.
3228 * slub_max_order specifies the order where we begin to stop considering the
3229 * number of objects in a slab as critical. If we reach slub_max_order then
3230 * we try to keep the page order as low as possible. So we accept more waste
3231 * of space in favor of a small page order.
3233 * Higher order allocations also allow the placement of more objects in a
3234 * slab and thereby reduce object handling overhead. If the user has
3235 * requested a higher mininum order then we start with that one instead of
3236 * the smallest order which will fit the object.
3238 static inline unsigned int slab_order(unsigned int size,
3239 unsigned int min_objects, unsigned int max_order,
3240 unsigned int fract_leftover)
3242 unsigned int min_order = slub_min_order;
3245 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
3246 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3248 for (order = max(min_order, (unsigned int)get_order(min_objects * size));
3249 order <= max_order; order++) {
3251 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
3254 rem = slab_size % size;
3256 if (rem <= slab_size / fract_leftover)
3263 static inline int calculate_order(unsigned int size)
3266 unsigned int min_objects;
3267 unsigned int max_objects;
3270 * Attempt to find best configuration for a slab. This
3271 * works by first attempting to generate a layout with
3272 * the best configuration and backing off gradually.
3274 * First we increase the acceptable waste in a slab. Then
3275 * we reduce the minimum objects required in a slab.
3277 min_objects = slub_min_objects;
3279 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3280 max_objects = order_objects(slub_max_order, size);
3281 min_objects = min(min_objects, max_objects);
3283 while (min_objects > 1) {
3284 unsigned int fraction;
3287 while (fraction >= 4) {
3288 order = slab_order(size, min_objects,
3289 slub_max_order, fraction);
3290 if (order <= slub_max_order)
3298 * We were unable to place multiple objects in a slab. Now
3299 * lets see if we can place a single object there.
3301 order = slab_order(size, 1, slub_max_order, 1);
3302 if (order <= slub_max_order)
3306 * Doh this slab cannot be placed using slub_max_order.
3308 order = slab_order(size, 1, MAX_ORDER, 1);
3309 if (order < MAX_ORDER)
3315 init_kmem_cache_node(struct kmem_cache_node *n)
3318 spin_lock_init(&n->list_lock);
3319 INIT_LIST_HEAD(&n->partial);
3320 #ifdef CONFIG_SLUB_DEBUG
3321 atomic_long_set(&n->nr_slabs, 0);
3322 atomic_long_set(&n->total_objects, 0);
3323 INIT_LIST_HEAD(&n->full);
3327 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3329 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3330 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3333 * Must align to double word boundary for the double cmpxchg
3334 * instructions to work; see __pcpu_double_call_return_bool().
3336 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3337 2 * sizeof(void *));
3342 init_kmem_cache_cpus(s);
3347 static struct kmem_cache *kmem_cache_node;
3350 * No kmalloc_node yet so do it by hand. We know that this is the first
3351 * slab on the node for this slabcache. There are no concurrent accesses
3354 * Note that this function only works on the kmem_cache_node
3355 * when allocating for the kmem_cache_node. This is used for bootstrapping
3356 * memory on a fresh node that has no slab structures yet.
3358 static void early_kmem_cache_node_alloc(int node)
3361 struct kmem_cache_node *n;
3363 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3365 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3368 if (page_to_nid(page) != node) {
3369 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3370 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3375 page->freelist = get_freepointer(kmem_cache_node, n);
3378 kmem_cache_node->node[node] = n;
3379 #ifdef CONFIG_SLUB_DEBUG
3380 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3381 init_tracking(kmem_cache_node, n);
3383 kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node),
3385 init_kmem_cache_node(n);
3386 inc_slabs_node(kmem_cache_node, node, page->objects);
3389 * No locks need to be taken here as it has just been
3390 * initialized and there is no concurrent access.
3392 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3395 static void free_kmem_cache_nodes(struct kmem_cache *s)
3398 struct kmem_cache_node *n;
3400 for_each_kmem_cache_node(s, node, n) {
3401 s->node[node] = NULL;
3402 kmem_cache_free(kmem_cache_node, n);
3406 void __kmem_cache_release(struct kmem_cache *s)
3408 cache_random_seq_destroy(s);
3409 free_percpu(s->cpu_slab);
3410 free_kmem_cache_nodes(s);
3413 static int init_kmem_cache_nodes(struct kmem_cache *s)
3417 for_each_node_state(node, N_NORMAL_MEMORY) {
3418 struct kmem_cache_node *n;
3420 if (slab_state == DOWN) {
3421 early_kmem_cache_node_alloc(node);
3424 n = kmem_cache_alloc_node(kmem_cache_node,
3428 free_kmem_cache_nodes(s);
3432 init_kmem_cache_node(n);
3438 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3440 if (min < MIN_PARTIAL)
3442 else if (min > MAX_PARTIAL)
3444 s->min_partial = min;
3447 static void set_cpu_partial(struct kmem_cache *s)
3449 #ifdef CONFIG_SLUB_CPU_PARTIAL
3451 * cpu_partial determined the maximum number of objects kept in the
3452 * per cpu partial lists of a processor.
3454 * Per cpu partial lists mainly contain slabs that just have one
3455 * object freed. If they are used for allocation then they can be
3456 * filled up again with minimal effort. The slab will never hit the
3457 * per node partial lists and therefore no locking will be required.
3459 * This setting also determines
3461 * A) The number of objects from per cpu partial slabs dumped to the
3462 * per node list when we reach the limit.
3463 * B) The number of objects in cpu partial slabs to extract from the
3464 * per node list when we run out of per cpu objects. We only fetch
3465 * 50% to keep some capacity around for frees.
3467 if (!kmem_cache_has_cpu_partial(s))
3469 else if (s->size >= PAGE_SIZE)
3471 else if (s->size >= 1024)
3473 else if (s->size >= 256)
3474 s->cpu_partial = 13;
3476 s->cpu_partial = 30;
3481 * calculate_sizes() determines the order and the distribution of data within
3484 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3486 slab_flags_t flags = s->flags;
3487 unsigned int size = s->object_size;
3491 * Round up object size to the next word boundary. We can only
3492 * place the free pointer at word boundaries and this determines
3493 * the possible location of the free pointer.
3495 size = ALIGN(size, sizeof(void *));
3497 #ifdef CONFIG_SLUB_DEBUG
3499 * Determine if we can poison the object itself. If the user of
3500 * the slab may touch the object after free or before allocation
3501 * then we should never poison the object itself.
3503 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
3505 s->flags |= __OBJECT_POISON;
3507 s->flags &= ~__OBJECT_POISON;
3511 * If we are Redzoning then check if there is some space between the
3512 * end of the object and the free pointer. If not then add an
3513 * additional word to have some bytes to store Redzone information.
3515 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3516 size += sizeof(void *);
3520 * With that we have determined the number of bytes in actual use
3521 * by the object. This is the potential offset to the free pointer.
3525 if (((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
3528 * Relocate free pointer after the object if it is not
3529 * permitted to overwrite the first word of the object on
3532 * This is the case if we do RCU, have a constructor or
3533 * destructor or are poisoning the objects.
3536 size += sizeof(void *);
3539 #ifdef CONFIG_SLUB_DEBUG
3540 if (flags & SLAB_STORE_USER)
3542 * Need to store information about allocs and frees after
3545 size += 2 * sizeof(struct track);
3548 kasan_cache_create(s, &size, &s->flags);
3549 #ifdef CONFIG_SLUB_DEBUG
3550 if (flags & SLAB_RED_ZONE) {
3552 * Add some empty padding so that we can catch
3553 * overwrites from earlier objects rather than let
3554 * tracking information or the free pointer be
3555 * corrupted if a user writes before the start
3558 size += sizeof(void *);
3560 s->red_left_pad = sizeof(void *);
3561 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3562 size += s->red_left_pad;
3567 * SLUB stores one object immediately after another beginning from
3568 * offset 0. In order to align the objects we have to simply size
3569 * each object to conform to the alignment.
3571 size = ALIGN(size, s->align);
3573 if (forced_order >= 0)
3574 order = forced_order;
3576 order = calculate_order(size);
3583 s->allocflags |= __GFP_COMP;
3585 if (s->flags & SLAB_CACHE_DMA)
3586 s->allocflags |= GFP_DMA;
3588 if (s->flags & SLAB_CACHE_DMA32)
3589 s->allocflags |= GFP_DMA32;
3591 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3592 s->allocflags |= __GFP_RECLAIMABLE;
3595 * Determine the number of objects per slab
3597 s->oo = oo_make(order, size);
3598 s->min = oo_make(get_order(size), size);
3599 if (oo_objects(s->oo) > oo_objects(s->max))
3602 return !!oo_objects(s->oo);
3605 static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
3607 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3608 #ifdef CONFIG_SLAB_FREELIST_HARDENED
3609 s->random = get_random_long();
3612 if (!calculate_sizes(s, -1))
3614 if (disable_higher_order_debug) {
3616 * Disable debugging flags that store metadata if the min slab
3619 if (get_order(s->size) > get_order(s->object_size)) {
3620 s->flags &= ~DEBUG_METADATA_FLAGS;
3622 if (!calculate_sizes(s, -1))
3627 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3628 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3629 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
3630 /* Enable fast mode */
3631 s->flags |= __CMPXCHG_DOUBLE;
3635 * The larger the object size is, the more pages we want on the partial
3636 * list to avoid pounding the page allocator excessively.
3638 set_min_partial(s, ilog2(s->size) / 2);
3643 s->remote_node_defrag_ratio = 1000;
3646 /* Initialize the pre-computed randomized freelist if slab is up */
3647 if (slab_state >= UP) {
3648 if (init_cache_random_seq(s))
3652 if (!init_kmem_cache_nodes(s))
3655 if (alloc_kmem_cache_cpus(s))
3658 free_kmem_cache_nodes(s);
3660 if (flags & SLAB_PANIC)
3661 panic("Cannot create slab %s size=%u realsize=%u order=%u offset=%u flags=%lx\n",
3662 s->name, s->size, s->size,
3663 oo_order(s->oo), s->offset, (unsigned long)flags);
3667 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3670 #ifdef CONFIG_SLUB_DEBUG
3671 void *addr = page_address(page);
3673 unsigned long *map = kcalloc(BITS_TO_LONGS(page->objects),
3678 slab_err(s, page, text, s->name);
3681 get_map(s, page, map);
3682 for_each_object(p, s, addr, page->objects) {
3684 if (!test_bit(slab_index(p, s, addr), map)) {
3685 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3686 print_tracking(s, p);
3695 * Attempt to free all partial slabs on a node.
3696 * This is called from __kmem_cache_shutdown(). We must take list_lock
3697 * because sysfs file might still access partial list after the shutdowning.
3699 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3702 struct page *page, *h;
3704 BUG_ON(irqs_disabled());
3705 spin_lock_irq(&n->list_lock);
3706 list_for_each_entry_safe(page, h, &n->partial, lru) {
3708 remove_partial(n, page);
3709 list_add(&page->lru, &discard);
3711 list_slab_objects(s, page,
3712 "Objects remaining in %s on __kmem_cache_shutdown()");
3715 spin_unlock_irq(&n->list_lock);
3717 list_for_each_entry_safe(page, h, &discard, lru)
3718 discard_slab(s, page);
3721 bool __kmem_cache_empty(struct kmem_cache *s)
3724 struct kmem_cache_node *n;
3726 for_each_kmem_cache_node(s, node, n)
3727 if (n->nr_partial || slabs_node(s, node))
3733 * Release all resources used by a slab cache.
3735 int __kmem_cache_shutdown(struct kmem_cache *s)
3738 struct kmem_cache_node *n;
3741 /* Attempt to free all objects */
3742 for_each_kmem_cache_node(s, node, n) {
3744 if (n->nr_partial || slabs_node(s, node))
3747 sysfs_slab_remove(s);
3751 /********************************************************************
3753 *******************************************************************/
3755 static int __init setup_slub_min_order(char *str)
3757 get_option(&str, (int *)&slub_min_order);
3762 __setup("slub_min_order=", setup_slub_min_order);
3764 static int __init setup_slub_max_order(char *str)
3766 get_option(&str, (int *)&slub_max_order);
3767 slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1);
3772 __setup("slub_max_order=", setup_slub_max_order);
3774 static int __init setup_slub_min_objects(char *str)
3776 get_option(&str, (int *)&slub_min_objects);
3781 __setup("slub_min_objects=", setup_slub_min_objects);
3783 void *__kmalloc(size_t size, gfp_t flags)
3785 struct kmem_cache *s;
3788 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3789 return kmalloc_large(size, flags);
3791 s = kmalloc_slab(size, flags);
3793 if (unlikely(ZERO_OR_NULL_PTR(s)))
3796 ret = slab_alloc(s, flags, _RET_IP_);
3798 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3800 kasan_kmalloc(s, ret, size, flags);
3804 EXPORT_SYMBOL(__kmalloc);
3807 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3812 flags |= __GFP_COMP;
3813 page = alloc_pages_node(node, flags, get_order(size));
3815 ptr = page_address(page);
3817 kmalloc_large_node_hook(ptr, size, flags);
3821 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3823 struct kmem_cache *s;
3826 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3827 ret = kmalloc_large_node(size, flags, node);
3829 trace_kmalloc_node(_RET_IP_, ret,
3830 size, PAGE_SIZE << get_order(size),
3836 s = kmalloc_slab(size, flags);
3838 if (unlikely(ZERO_OR_NULL_PTR(s)))
3841 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3843 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3845 kasan_kmalloc(s, ret, size, flags);
3849 EXPORT_SYMBOL(__kmalloc_node);
3852 #ifdef CONFIG_HARDENED_USERCOPY
3854 * Rejects incorrectly sized objects and objects that are to be copied
3855 * to/from userspace but do not fall entirely within the containing slab
3856 * cache's usercopy region.
3858 * Returns NULL if check passes, otherwise const char * to name of cache
3859 * to indicate an error.
3861 void __check_heap_object(const void *ptr, unsigned long n, struct page *page,
3864 struct kmem_cache *s;
3865 unsigned int offset;
3868 /* Find object and usable object size. */
3869 s = page->slab_cache;
3871 /* Reject impossible pointers. */
3872 if (ptr < page_address(page))
3873 usercopy_abort("SLUB object not in SLUB page?!", NULL,
3876 /* Find offset within object. */
3877 offset = (ptr - page_address(page)) % s->size;
3879 /* Adjust for redzone and reject if within the redzone. */
3880 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE) {
3881 if (offset < s->red_left_pad)
3882 usercopy_abort("SLUB object in left red zone",
3883 s->name, to_user, offset, n);
3884 offset -= s->red_left_pad;
3887 /* Allow address range falling entirely within usercopy region. */
3888 if (offset >= s->useroffset &&
3889 offset - s->useroffset <= s->usersize &&
3890 n <= s->useroffset - offset + s->usersize)
3894 * If the copy is still within the allocated object, produce
3895 * a warning instead of rejecting the copy. This is intended
3896 * to be a temporary method to find any missing usercopy
3899 object_size = slab_ksize(s);
3900 if (usercopy_fallback &&
3901 offset <= object_size && n <= object_size - offset) {
3902 usercopy_warn("SLUB object", s->name, to_user, offset, n);
3906 usercopy_abort("SLUB object", s->name, to_user, offset, n);
3908 #endif /* CONFIG_HARDENED_USERCOPY */
3910 static size_t __ksize(const void *object)
3914 if (unlikely(object == ZERO_SIZE_PTR))
3917 page = virt_to_head_page(object);
3919 if (unlikely(!PageSlab(page))) {
3920 WARN_ON(!PageCompound(page));
3921 return PAGE_SIZE << compound_order(page);
3924 return slab_ksize(page->slab_cache);
3927 size_t ksize(const void *object)
3929 size_t size = __ksize(object);
3930 /* We assume that ksize callers could use whole allocated area,
3931 * so we need to unpoison this area.
3933 kasan_unpoison_shadow(object, size);
3936 EXPORT_SYMBOL(ksize);
3938 void kfree(const void *x)
3941 void *object = (void *)x;
3943 trace_kfree(_RET_IP_, x);
3945 if (unlikely(ZERO_OR_NULL_PTR(x)))
3948 page = virt_to_head_page(x);
3949 if (unlikely(!PageSlab(page))) {
3950 BUG_ON(!PageCompound(page));
3952 __free_pages(page, compound_order(page));
3955 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
3957 EXPORT_SYMBOL(kfree);
3959 #define SHRINK_PROMOTE_MAX 32
3962 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3963 * up most to the head of the partial lists. New allocations will then
3964 * fill those up and thus they can be removed from the partial lists.
3966 * The slabs with the least items are placed last. This results in them
3967 * being allocated from last increasing the chance that the last objects
3968 * are freed in them.
3970 int __kmem_cache_shrink(struct kmem_cache *s)
3974 struct kmem_cache_node *n;
3977 struct list_head discard;
3978 struct list_head promote[SHRINK_PROMOTE_MAX];
3979 unsigned long flags;
3983 for_each_kmem_cache_node(s, node, n) {
3984 INIT_LIST_HEAD(&discard);
3985 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3986 INIT_LIST_HEAD(promote + i);
3988 spin_lock_irqsave(&n->list_lock, flags);
3991 * Build lists of slabs to discard or promote.
3993 * Note that concurrent frees may occur while we hold the
3994 * list_lock. page->inuse here is the upper limit.
3996 list_for_each_entry_safe(page, t, &n->partial, lru) {
3997 int free = page->objects - page->inuse;
3999 /* Do not reread page->inuse */
4002 /* We do not keep full slabs on the list */
4005 if (free == page->objects) {
4006 list_move(&page->lru, &discard);
4008 } else if (free <= SHRINK_PROMOTE_MAX)
4009 list_move(&page->lru, promote + free - 1);
4013 * Promote the slabs filled up most to the head of the
4016 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
4017 list_splice(promote + i, &n->partial);
4019 spin_unlock_irqrestore(&n->list_lock, flags);
4021 /* Release empty slabs */
4022 list_for_each_entry_safe(page, t, &discard, lru)
4023 discard_slab(s, page);
4025 if (slabs_node(s, node))
4033 static void kmemcg_cache_deact_after_rcu(struct kmem_cache *s)
4036 * Called with all the locks held after a sched RCU grace period.
4037 * Even if @s becomes empty after shrinking, we can't know that @s
4038 * doesn't have allocations already in-flight and thus can't
4039 * destroy @s until the associated memcg is released.
4041 * However, let's remove the sysfs files for empty caches here.
4042 * Each cache has a lot of interface files which aren't
4043 * particularly useful for empty draining caches; otherwise, we can
4044 * easily end up with millions of unnecessary sysfs files on
4045 * systems which have a lot of memory and transient cgroups.
4047 if (!__kmem_cache_shrink(s))
4048 sysfs_slab_remove(s);
4051 void __kmemcg_cache_deactivate(struct kmem_cache *s)
4054 * Disable empty slabs caching. Used to avoid pinning offline
4055 * memory cgroups by kmem pages that can be freed.
4057 slub_set_cpu_partial(s, 0);
4061 * s->cpu_partial is checked locklessly (see put_cpu_partial), so
4062 * we have to make sure the change is visible before shrinking.
4064 slab_deactivate_memcg_cache_rcu_sched(s, kmemcg_cache_deact_after_rcu);
4068 static int slab_mem_going_offline_callback(void *arg)
4070 struct kmem_cache *s;
4072 mutex_lock(&slab_mutex);
4073 list_for_each_entry(s, &slab_caches, list)
4074 __kmem_cache_shrink(s);
4075 mutex_unlock(&slab_mutex);
4080 static void slab_mem_offline_callback(void *arg)
4082 struct kmem_cache_node *n;
4083 struct kmem_cache *s;
4084 struct memory_notify *marg = arg;
4087 offline_node = marg->status_change_nid_normal;
4090 * If the node still has available memory. we need kmem_cache_node
4093 if (offline_node < 0)
4096 mutex_lock(&slab_mutex);
4097 list_for_each_entry(s, &slab_caches, list) {
4098 n = get_node(s, offline_node);
4101 * if n->nr_slabs > 0, slabs still exist on the node
4102 * that is going down. We were unable to free them,
4103 * and offline_pages() function shouldn't call this
4104 * callback. So, we must fail.
4106 BUG_ON(slabs_node(s, offline_node));
4108 s->node[offline_node] = NULL;
4109 kmem_cache_free(kmem_cache_node, n);
4112 mutex_unlock(&slab_mutex);
4115 static int slab_mem_going_online_callback(void *arg)
4117 struct kmem_cache_node *n;
4118 struct kmem_cache *s;
4119 struct memory_notify *marg = arg;
4120 int nid = marg->status_change_nid_normal;
4124 * If the node's memory is already available, then kmem_cache_node is
4125 * already created. Nothing to do.
4131 * We are bringing a node online. No memory is available yet. We must
4132 * allocate a kmem_cache_node structure in order to bring the node
4135 mutex_lock(&slab_mutex);
4136 list_for_each_entry(s, &slab_caches, list) {
4138 * XXX: kmem_cache_alloc_node will fallback to other nodes
4139 * since memory is not yet available from the node that
4142 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4147 init_kmem_cache_node(n);
4151 mutex_unlock(&slab_mutex);
4155 static int slab_memory_callback(struct notifier_block *self,
4156 unsigned long action, void *arg)
4161 case MEM_GOING_ONLINE:
4162 ret = slab_mem_going_online_callback(arg);
4164 case MEM_GOING_OFFLINE:
4165 ret = slab_mem_going_offline_callback(arg);
4168 case MEM_CANCEL_ONLINE:
4169 slab_mem_offline_callback(arg);
4172 case MEM_CANCEL_OFFLINE:
4176 ret = notifier_from_errno(ret);
4182 static struct notifier_block slab_memory_callback_nb = {
4183 .notifier_call = slab_memory_callback,
4184 .priority = SLAB_CALLBACK_PRI,
4187 /********************************************************************
4188 * Basic setup of slabs
4189 *******************************************************************/
4192 * Used for early kmem_cache structures that were allocated using
4193 * the page allocator. Allocate them properly then fix up the pointers
4194 * that may be pointing to the wrong kmem_cache structure.
4197 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4200 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4201 struct kmem_cache_node *n;
4203 memcpy(s, static_cache, kmem_cache->object_size);
4206 * This runs very early, and only the boot processor is supposed to be
4207 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4210 __flush_cpu_slab(s, smp_processor_id());
4211 for_each_kmem_cache_node(s, node, n) {
4214 list_for_each_entry(p, &n->partial, lru)
4217 #ifdef CONFIG_SLUB_DEBUG
4218 list_for_each_entry(p, &n->full, lru)
4222 slab_init_memcg_params(s);
4223 list_add(&s->list, &slab_caches);
4224 memcg_link_cache(s);
4228 void __init kmem_cache_init(void)
4230 static __initdata struct kmem_cache boot_kmem_cache,
4231 boot_kmem_cache_node;
4233 if (debug_guardpage_minorder())
4236 kmem_cache_node = &boot_kmem_cache_node;
4237 kmem_cache = &boot_kmem_cache;
4239 create_boot_cache(kmem_cache_node, "kmem_cache_node",
4240 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
4242 register_hotmemory_notifier(&slab_memory_callback_nb);
4244 /* Able to allocate the per node structures */
4245 slab_state = PARTIAL;
4247 create_boot_cache(kmem_cache, "kmem_cache",
4248 offsetof(struct kmem_cache, node) +
4249 nr_node_ids * sizeof(struct kmem_cache_node *),
4250 SLAB_HWCACHE_ALIGN, 0, 0);
4252 kmem_cache = bootstrap(&boot_kmem_cache);
4253 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
4255 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4256 setup_kmalloc_cache_index_table();
4257 create_kmalloc_caches(0);
4259 /* Setup random freelists for each cache */
4260 init_freelist_randomization();
4262 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
4265 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%d\n",
4267 slub_min_order, slub_max_order, slub_min_objects,
4268 nr_cpu_ids, nr_node_ids);
4271 void __init kmem_cache_init_late(void)
4276 __kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
4277 slab_flags_t flags, void (*ctor)(void *))
4279 struct kmem_cache *s, *c;
4281 s = find_mergeable(size, align, flags, name, ctor);
4286 * Adjust the object sizes so that we clear
4287 * the complete object on kzalloc.
4289 s->object_size = max(s->object_size, size);
4290 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
4292 for_each_memcg_cache(c, s) {
4293 c->object_size = s->object_size;
4294 c->inuse = max(c->inuse, ALIGN(size, sizeof(void *)));
4297 if (sysfs_slab_alias(s, name)) {
4306 int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
4310 err = kmem_cache_open(s, flags);
4314 /* Mutex is not taken during early boot */
4315 if (slab_state <= UP)
4318 memcg_propagate_slab_attrs(s);
4319 err = sysfs_slab_add(s);
4321 __kmem_cache_release(s);
4326 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4328 struct kmem_cache *s;
4331 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4332 return kmalloc_large(size, gfpflags);
4334 s = kmalloc_slab(size, gfpflags);
4336 if (unlikely(ZERO_OR_NULL_PTR(s)))
4339 ret = slab_alloc(s, gfpflags, caller);
4341 /* Honor the call site pointer we received. */
4342 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4348 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4349 int node, unsigned long caller)
4351 struct kmem_cache *s;
4354 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4355 ret = kmalloc_large_node(size, gfpflags, node);
4357 trace_kmalloc_node(caller, ret,
4358 size, PAGE_SIZE << get_order(size),
4364 s = kmalloc_slab(size, gfpflags);
4366 if (unlikely(ZERO_OR_NULL_PTR(s)))
4369 ret = slab_alloc_node(s, gfpflags, node, caller);
4371 /* Honor the call site pointer we received. */
4372 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4379 static int count_inuse(struct page *page)
4384 static int count_total(struct page *page)
4386 return page->objects;
4390 #ifdef CONFIG_SLUB_DEBUG
4391 static int validate_slab(struct kmem_cache *s, struct page *page,
4395 void *addr = page_address(page);
4397 if (!check_slab(s, page) ||
4398 !on_freelist(s, page, NULL))
4401 /* Now we know that a valid freelist exists */
4402 bitmap_zero(map, page->objects);
4404 get_map(s, page, map);
4405 for_each_object(p, s, addr, page->objects) {
4406 if (test_bit(slab_index(p, s, addr), map))
4407 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4411 for_each_object(p, s, addr, page->objects)
4412 if (!test_bit(slab_index(p, s, addr), map))
4413 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4418 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4422 validate_slab(s, page, map);
4426 static int validate_slab_node(struct kmem_cache *s,
4427 struct kmem_cache_node *n, unsigned long *map)
4429 unsigned long count = 0;
4431 unsigned long flags;
4433 spin_lock_irqsave(&n->list_lock, flags);
4435 list_for_each_entry(page, &n->partial, lru) {
4436 validate_slab_slab(s, page, map);
4439 if (count != n->nr_partial)
4440 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4441 s->name, count, n->nr_partial);
4443 if (!(s->flags & SLAB_STORE_USER))
4446 list_for_each_entry(page, &n->full, lru) {
4447 validate_slab_slab(s, page, map);
4450 if (count != atomic_long_read(&n->nr_slabs))
4451 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4452 s->name, count, atomic_long_read(&n->nr_slabs));
4455 spin_unlock_irqrestore(&n->list_lock, flags);
4459 static long validate_slab_cache(struct kmem_cache *s)
4462 unsigned long count = 0;
4463 unsigned long *map = kmalloc_array(BITS_TO_LONGS(oo_objects(s->max)),
4464 sizeof(unsigned long),
4466 struct kmem_cache_node *n;
4472 for_each_kmem_cache_node(s, node, n)
4473 count += validate_slab_node(s, n, map);
4478 * Generate lists of code addresses where slabcache objects are allocated
4483 unsigned long count;
4490 DECLARE_BITMAP(cpus, NR_CPUS);
4496 unsigned long count;
4497 struct location *loc;
4500 static void free_loc_track(struct loc_track *t)
4503 free_pages((unsigned long)t->loc,
4504 get_order(sizeof(struct location) * t->max));
4507 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4512 order = get_order(sizeof(struct location) * max);
4514 l = (void *)__get_free_pages(flags, order);
4519 memcpy(l, t->loc, sizeof(struct location) * t->count);
4527 static int add_location(struct loc_track *t, struct kmem_cache *s,
4528 const struct track *track)
4530 long start, end, pos;
4532 unsigned long caddr;
4533 unsigned long age = jiffies - track->when;
4539 pos = start + (end - start + 1) / 2;
4542 * There is nothing at "end". If we end up there
4543 * we need to add something to before end.
4548 caddr = t->loc[pos].addr;
4549 if (track->addr == caddr) {
4555 if (age < l->min_time)
4557 if (age > l->max_time)
4560 if (track->pid < l->min_pid)
4561 l->min_pid = track->pid;
4562 if (track->pid > l->max_pid)
4563 l->max_pid = track->pid;
4565 cpumask_set_cpu(track->cpu,
4566 to_cpumask(l->cpus));
4568 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4572 if (track->addr < caddr)
4579 * Not found. Insert new tracking element.
4581 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4587 (t->count - pos) * sizeof(struct location));
4590 l->addr = track->addr;
4594 l->min_pid = track->pid;
4595 l->max_pid = track->pid;
4596 cpumask_clear(to_cpumask(l->cpus));
4597 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4598 nodes_clear(l->nodes);
4599 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4603 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4604 struct page *page, enum track_item alloc,
4607 void *addr = page_address(page);
4610 bitmap_zero(map, page->objects);
4611 get_map(s, page, map);
4613 for_each_object(p, s, addr, page->objects)
4614 if (!test_bit(slab_index(p, s, addr), map))
4615 add_location(t, s, get_track(s, p, alloc));
4618 static int list_locations(struct kmem_cache *s, char *buf,
4619 enum track_item alloc)
4623 struct loc_track t = { 0, 0, NULL };
4625 unsigned long *map = kmalloc_array(BITS_TO_LONGS(oo_objects(s->max)),
4626 sizeof(unsigned long),
4628 struct kmem_cache_node *n;
4630 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4633 return sprintf(buf, "Out of memory\n");
4635 /* Push back cpu slabs */
4638 for_each_kmem_cache_node(s, node, n) {
4639 unsigned long flags;
4642 if (!atomic_long_read(&n->nr_slabs))
4645 spin_lock_irqsave(&n->list_lock, flags);
4646 list_for_each_entry(page, &n->partial, lru)
4647 process_slab(&t, s, page, alloc, map);
4648 list_for_each_entry(page, &n->full, lru)
4649 process_slab(&t, s, page, alloc, map);
4650 spin_unlock_irqrestore(&n->list_lock, flags);
4653 for (i = 0; i < t.count; i++) {
4654 struct location *l = &t.loc[i];
4656 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4658 len += sprintf(buf + len, "%7ld ", l->count);
4661 len += sprintf(buf + len, "%pS", (void *)l->addr);
4663 len += sprintf(buf + len, "<not-available>");
4665 if (l->sum_time != l->min_time) {
4666 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4668 (long)div_u64(l->sum_time, l->count),
4671 len += sprintf(buf + len, " age=%ld",
4674 if (l->min_pid != l->max_pid)
4675 len += sprintf(buf + len, " pid=%ld-%ld",
4676 l->min_pid, l->max_pid);
4678 len += sprintf(buf + len, " pid=%ld",
4681 if (num_online_cpus() > 1 &&
4682 !cpumask_empty(to_cpumask(l->cpus)) &&
4683 len < PAGE_SIZE - 60)
4684 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4686 cpumask_pr_args(to_cpumask(l->cpus)));
4688 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4689 len < PAGE_SIZE - 60)
4690 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4692 nodemask_pr_args(&l->nodes));
4694 len += sprintf(buf + len, "\n");
4700 len += sprintf(buf, "No data\n");
4705 #ifdef SLUB_RESILIENCY_TEST
4706 static void __init resiliency_test(void)
4710 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4712 pr_err("SLUB resiliency testing\n");
4713 pr_err("-----------------------\n");
4714 pr_err("A. Corruption after allocation\n");
4716 p = kzalloc(16, GFP_KERNEL);
4718 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4721 validate_slab_cache(kmalloc_caches[4]);
4723 /* Hmmm... The next two are dangerous */
4724 p = kzalloc(32, GFP_KERNEL);
4725 p[32 + sizeof(void *)] = 0x34;
4726 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4728 pr_err("If allocated object is overwritten then not detectable\n\n");
4730 validate_slab_cache(kmalloc_caches[5]);
4731 p = kzalloc(64, GFP_KERNEL);
4732 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4734 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4736 pr_err("If allocated object is overwritten then not detectable\n\n");
4737 validate_slab_cache(kmalloc_caches[6]);
4739 pr_err("\nB. Corruption after free\n");
4740 p = kzalloc(128, GFP_KERNEL);
4743 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4744 validate_slab_cache(kmalloc_caches[7]);
4746 p = kzalloc(256, GFP_KERNEL);
4749 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4750 validate_slab_cache(kmalloc_caches[8]);
4752 p = kzalloc(512, GFP_KERNEL);
4755 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4756 validate_slab_cache(kmalloc_caches[9]);
4760 static void resiliency_test(void) {};
4765 enum slab_stat_type {
4766 SL_ALL, /* All slabs */
4767 SL_PARTIAL, /* Only partially allocated slabs */
4768 SL_CPU, /* Only slabs used for cpu caches */
4769 SL_OBJECTS, /* Determine allocated objects not slabs */
4770 SL_TOTAL /* Determine object capacity not slabs */
4773 #define SO_ALL (1 << SL_ALL)
4774 #define SO_PARTIAL (1 << SL_PARTIAL)
4775 #define SO_CPU (1 << SL_CPU)
4776 #define SO_OBJECTS (1 << SL_OBJECTS)
4777 #define SO_TOTAL (1 << SL_TOTAL)
4780 static bool memcg_sysfs_enabled = IS_ENABLED(CONFIG_SLUB_MEMCG_SYSFS_ON);
4782 static int __init setup_slub_memcg_sysfs(char *str)
4786 if (get_option(&str, &v) > 0)
4787 memcg_sysfs_enabled = v;
4792 __setup("slub_memcg_sysfs=", setup_slub_memcg_sysfs);
4795 static ssize_t show_slab_objects(struct kmem_cache *s,
4796 char *buf, unsigned long flags)
4798 unsigned long total = 0;
4801 unsigned long *nodes;
4803 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
4807 if (flags & SO_CPU) {
4810 for_each_possible_cpu(cpu) {
4811 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4816 page = READ_ONCE(c->page);
4820 node = page_to_nid(page);
4821 if (flags & SO_TOTAL)
4823 else if (flags & SO_OBJECTS)
4831 page = slub_percpu_partial_read_once(c);
4833 node = page_to_nid(page);
4834 if (flags & SO_TOTAL)
4836 else if (flags & SO_OBJECTS)
4847 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
4848 * already held which will conflict with an existing lock order:
4850 * mem_hotplug_lock->slab_mutex->kernfs_mutex
4852 * We don't really need mem_hotplug_lock (to hold off
4853 * slab_mem_going_offline_callback) here because slab's memory hot
4854 * unplug code doesn't destroy the kmem_cache->node[] data.
4857 #ifdef CONFIG_SLUB_DEBUG
4858 if (flags & SO_ALL) {
4859 struct kmem_cache_node *n;
4861 for_each_kmem_cache_node(s, node, n) {
4863 if (flags & SO_TOTAL)
4864 x = atomic_long_read(&n->total_objects);
4865 else if (flags & SO_OBJECTS)
4866 x = atomic_long_read(&n->total_objects) -
4867 count_partial(n, count_free);
4869 x = atomic_long_read(&n->nr_slabs);
4876 if (flags & SO_PARTIAL) {
4877 struct kmem_cache_node *n;
4879 for_each_kmem_cache_node(s, node, n) {
4880 if (flags & SO_TOTAL)
4881 x = count_partial(n, count_total);
4882 else if (flags & SO_OBJECTS)
4883 x = count_partial(n, count_inuse);
4890 x = sprintf(buf, "%lu", total);
4892 for (node = 0; node < nr_node_ids; node++)
4894 x += sprintf(buf + x, " N%d=%lu",
4898 return x + sprintf(buf + x, "\n");
4901 #ifdef CONFIG_SLUB_DEBUG
4902 static int any_slab_objects(struct kmem_cache *s)
4905 struct kmem_cache_node *n;
4907 for_each_kmem_cache_node(s, node, n)
4908 if (atomic_long_read(&n->total_objects))
4915 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4916 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4918 struct slab_attribute {
4919 struct attribute attr;
4920 ssize_t (*show)(struct kmem_cache *s, char *buf);
4921 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4924 #define SLAB_ATTR_RO(_name) \
4925 static struct slab_attribute _name##_attr = \
4926 __ATTR(_name, 0400, _name##_show, NULL)
4928 #define SLAB_ATTR(_name) \
4929 static struct slab_attribute _name##_attr = \
4930 __ATTR(_name, 0600, _name##_show, _name##_store)
4932 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4934 return sprintf(buf, "%u\n", s->size);
4936 SLAB_ATTR_RO(slab_size);
4938 static ssize_t align_show(struct kmem_cache *s, char *buf)
4940 return sprintf(buf, "%u\n", s->align);
4942 SLAB_ATTR_RO(align);
4944 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4946 return sprintf(buf, "%u\n", s->object_size);
4948 SLAB_ATTR_RO(object_size);
4950 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4952 return sprintf(buf, "%u\n", oo_objects(s->oo));
4954 SLAB_ATTR_RO(objs_per_slab);
4956 static ssize_t order_store(struct kmem_cache *s,
4957 const char *buf, size_t length)
4962 err = kstrtouint(buf, 10, &order);
4966 if (order > slub_max_order || order < slub_min_order)
4969 calculate_sizes(s, order);
4973 static ssize_t order_show(struct kmem_cache *s, char *buf)
4975 return sprintf(buf, "%u\n", oo_order(s->oo));
4979 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4981 return sprintf(buf, "%lu\n", s->min_partial);
4984 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4990 err = kstrtoul(buf, 10, &min);
4994 set_min_partial(s, min);
4997 SLAB_ATTR(min_partial);
4999 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
5001 return sprintf(buf, "%u\n", slub_cpu_partial(s));
5004 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
5007 unsigned int objects;
5010 err = kstrtouint(buf, 10, &objects);
5013 if (objects && !kmem_cache_has_cpu_partial(s))
5016 slub_set_cpu_partial(s, objects);
5020 SLAB_ATTR(cpu_partial);
5022 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
5026 return sprintf(buf, "%pS\n", s->ctor);
5030 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
5032 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
5034 SLAB_ATTR_RO(aliases);
5036 static ssize_t partial_show(struct kmem_cache *s, char *buf)
5038 return show_slab_objects(s, buf, SO_PARTIAL);
5040 SLAB_ATTR_RO(partial);
5042 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
5044 return show_slab_objects(s, buf, SO_CPU);
5046 SLAB_ATTR_RO(cpu_slabs);
5048 static ssize_t objects_show(struct kmem_cache *s, char *buf)
5050 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5052 SLAB_ATTR_RO(objects);
5054 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5056 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5058 SLAB_ATTR_RO(objects_partial);
5060 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5067 for_each_online_cpu(cpu) {
5070 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5073 pages += page->pages;
5074 objects += page->pobjects;
5078 len = sprintf(buf, "%d(%d)", objects, pages);
5081 for_each_online_cpu(cpu) {
5084 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5086 if (page && len < PAGE_SIZE - 20)
5087 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
5088 page->pobjects, page->pages);
5091 return len + sprintf(buf + len, "\n");
5093 SLAB_ATTR_RO(slabs_cpu_partial);
5095 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5097 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5100 static ssize_t reclaim_account_store(struct kmem_cache *s,
5101 const char *buf, size_t length)
5103 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
5105 s->flags |= SLAB_RECLAIM_ACCOUNT;
5108 SLAB_ATTR(reclaim_account);
5110 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5112 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5114 SLAB_ATTR_RO(hwcache_align);
5116 #ifdef CONFIG_ZONE_DMA
5117 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5119 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5121 SLAB_ATTR_RO(cache_dma);
5124 static ssize_t usersize_show(struct kmem_cache *s, char *buf)
5126 return sprintf(buf, "%u\n", s->usersize);
5128 SLAB_ATTR_RO(usersize);
5130 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5132 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
5134 SLAB_ATTR_RO(destroy_by_rcu);
5136 #ifdef CONFIG_SLUB_DEBUG
5137 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5139 return show_slab_objects(s, buf, SO_ALL);
5141 SLAB_ATTR_RO(slabs);
5143 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5145 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5147 SLAB_ATTR_RO(total_objects);
5149 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5151 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5154 static ssize_t sanity_checks_store(struct kmem_cache *s,
5155 const char *buf, size_t length)
5157 s->flags &= ~SLAB_CONSISTENCY_CHECKS;
5158 if (buf[0] == '1') {
5159 s->flags &= ~__CMPXCHG_DOUBLE;
5160 s->flags |= SLAB_CONSISTENCY_CHECKS;
5164 SLAB_ATTR(sanity_checks);
5166 static ssize_t trace_show(struct kmem_cache *s, char *buf)
5168 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5171 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
5175 * Tracing a merged cache is going to give confusing results
5176 * as well as cause other issues like converting a mergeable
5177 * cache into an umergeable one.
5179 if (s->refcount > 1)
5182 s->flags &= ~SLAB_TRACE;
5183 if (buf[0] == '1') {
5184 s->flags &= ~__CMPXCHG_DOUBLE;
5185 s->flags |= SLAB_TRACE;
5191 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5193 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5196 static ssize_t red_zone_store(struct kmem_cache *s,
5197 const char *buf, size_t length)
5199 if (any_slab_objects(s))
5202 s->flags &= ~SLAB_RED_ZONE;
5203 if (buf[0] == '1') {
5204 s->flags |= SLAB_RED_ZONE;
5206 calculate_sizes(s, -1);
5209 SLAB_ATTR(red_zone);
5211 static ssize_t poison_show(struct kmem_cache *s, char *buf)
5213 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
5216 static ssize_t poison_store(struct kmem_cache *s,
5217 const char *buf, size_t length)
5219 if (any_slab_objects(s))
5222 s->flags &= ~SLAB_POISON;
5223 if (buf[0] == '1') {
5224 s->flags |= SLAB_POISON;
5226 calculate_sizes(s, -1);
5231 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5233 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5236 static ssize_t store_user_store(struct kmem_cache *s,
5237 const char *buf, size_t length)
5239 if (any_slab_objects(s))
5242 s->flags &= ~SLAB_STORE_USER;
5243 if (buf[0] == '1') {
5244 s->flags &= ~__CMPXCHG_DOUBLE;
5245 s->flags |= SLAB_STORE_USER;
5247 calculate_sizes(s, -1);
5250 SLAB_ATTR(store_user);
5252 static ssize_t validate_show(struct kmem_cache *s, char *buf)
5257 static ssize_t validate_store(struct kmem_cache *s,
5258 const char *buf, size_t length)
5262 if (buf[0] == '1') {
5263 ret = validate_slab_cache(s);
5269 SLAB_ATTR(validate);
5271 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
5273 if (!(s->flags & SLAB_STORE_USER))
5275 return list_locations(s, buf, TRACK_ALLOC);
5277 SLAB_ATTR_RO(alloc_calls);
5279 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
5281 if (!(s->flags & SLAB_STORE_USER))
5283 return list_locations(s, buf, TRACK_FREE);
5285 SLAB_ATTR_RO(free_calls);
5286 #endif /* CONFIG_SLUB_DEBUG */
5288 #ifdef CONFIG_FAILSLAB
5289 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5291 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5294 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
5297 if (s->refcount > 1)
5300 s->flags &= ~SLAB_FAILSLAB;
5302 s->flags |= SLAB_FAILSLAB;
5305 SLAB_ATTR(failslab);
5308 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5313 static ssize_t shrink_store(struct kmem_cache *s,
5314 const char *buf, size_t length)
5317 kmem_cache_shrink(s);
5325 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5327 return sprintf(buf, "%u\n", s->remote_node_defrag_ratio / 10);
5330 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5331 const char *buf, size_t length)
5336 err = kstrtouint(buf, 10, &ratio);
5342 s->remote_node_defrag_ratio = ratio * 10;
5346 SLAB_ATTR(remote_node_defrag_ratio);
5349 #ifdef CONFIG_SLUB_STATS
5350 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5352 unsigned long sum = 0;
5355 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
5360 for_each_online_cpu(cpu) {
5361 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5367 len = sprintf(buf, "%lu", sum);
5370 for_each_online_cpu(cpu) {
5371 if (data[cpu] && len < PAGE_SIZE - 20)
5372 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5376 return len + sprintf(buf + len, "\n");
5379 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5383 for_each_online_cpu(cpu)
5384 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5387 #define STAT_ATTR(si, text) \
5388 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5390 return show_stat(s, buf, si); \
5392 static ssize_t text##_store(struct kmem_cache *s, \
5393 const char *buf, size_t length) \
5395 if (buf[0] != '0') \
5397 clear_stat(s, si); \
5402 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5403 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5404 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5405 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5406 STAT_ATTR(FREE_FROZEN, free_frozen);
5407 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5408 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5409 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5410 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5411 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5412 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5413 STAT_ATTR(FREE_SLAB, free_slab);
5414 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5415 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5416 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5417 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5418 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5419 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5420 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5421 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5422 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5423 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5424 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5425 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5426 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5427 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5430 static struct attribute *slab_attrs[] = {
5431 &slab_size_attr.attr,
5432 &object_size_attr.attr,
5433 &objs_per_slab_attr.attr,
5435 &min_partial_attr.attr,
5436 &cpu_partial_attr.attr,
5438 &objects_partial_attr.attr,
5440 &cpu_slabs_attr.attr,
5444 &hwcache_align_attr.attr,
5445 &reclaim_account_attr.attr,
5446 &destroy_by_rcu_attr.attr,
5448 &slabs_cpu_partial_attr.attr,
5449 #ifdef CONFIG_SLUB_DEBUG
5450 &total_objects_attr.attr,
5452 &sanity_checks_attr.attr,
5454 &red_zone_attr.attr,
5456 &store_user_attr.attr,
5457 &validate_attr.attr,
5458 &alloc_calls_attr.attr,
5459 &free_calls_attr.attr,
5461 #ifdef CONFIG_ZONE_DMA
5462 &cache_dma_attr.attr,
5465 &remote_node_defrag_ratio_attr.attr,
5467 #ifdef CONFIG_SLUB_STATS
5468 &alloc_fastpath_attr.attr,
5469 &alloc_slowpath_attr.attr,
5470 &free_fastpath_attr.attr,
5471 &free_slowpath_attr.attr,
5472 &free_frozen_attr.attr,
5473 &free_add_partial_attr.attr,
5474 &free_remove_partial_attr.attr,
5475 &alloc_from_partial_attr.attr,
5476 &alloc_slab_attr.attr,
5477 &alloc_refill_attr.attr,
5478 &alloc_node_mismatch_attr.attr,
5479 &free_slab_attr.attr,
5480 &cpuslab_flush_attr.attr,
5481 &deactivate_full_attr.attr,
5482 &deactivate_empty_attr.attr,
5483 &deactivate_to_head_attr.attr,
5484 &deactivate_to_tail_attr.attr,
5485 &deactivate_remote_frees_attr.attr,
5486 &deactivate_bypass_attr.attr,
5487 &order_fallback_attr.attr,
5488 &cmpxchg_double_fail_attr.attr,
5489 &cmpxchg_double_cpu_fail_attr.attr,
5490 &cpu_partial_alloc_attr.attr,
5491 &cpu_partial_free_attr.attr,
5492 &cpu_partial_node_attr.attr,
5493 &cpu_partial_drain_attr.attr,
5495 #ifdef CONFIG_FAILSLAB
5496 &failslab_attr.attr,
5498 &usersize_attr.attr,
5503 static const struct attribute_group slab_attr_group = {
5504 .attrs = slab_attrs,
5507 static ssize_t slab_attr_show(struct kobject *kobj,
5508 struct attribute *attr,
5511 struct slab_attribute *attribute;
5512 struct kmem_cache *s;
5515 attribute = to_slab_attr(attr);
5518 if (!attribute->show)
5521 err = attribute->show(s, buf);
5526 static ssize_t slab_attr_store(struct kobject *kobj,
5527 struct attribute *attr,
5528 const char *buf, size_t len)
5530 struct slab_attribute *attribute;
5531 struct kmem_cache *s;
5534 attribute = to_slab_attr(attr);
5537 if (!attribute->store)
5540 err = attribute->store(s, buf, len);
5542 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5543 struct kmem_cache *c;
5545 mutex_lock(&slab_mutex);
5546 if (s->max_attr_size < len)
5547 s->max_attr_size = len;
5550 * This is a best effort propagation, so this function's return
5551 * value will be determined by the parent cache only. This is
5552 * basically because not all attributes will have a well
5553 * defined semantics for rollbacks - most of the actions will
5554 * have permanent effects.
5556 * Returning the error value of any of the children that fail
5557 * is not 100 % defined, in the sense that users seeing the
5558 * error code won't be able to know anything about the state of
5561 * Only returning the error code for the parent cache at least
5562 * has well defined semantics. The cache being written to
5563 * directly either failed or succeeded, in which case we loop
5564 * through the descendants with best-effort propagation.
5566 for_each_memcg_cache(c, s)
5567 attribute->store(c, buf, len);
5568 mutex_unlock(&slab_mutex);
5574 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5578 char *buffer = NULL;
5579 struct kmem_cache *root_cache;
5581 if (is_root_cache(s))
5584 root_cache = s->memcg_params.root_cache;
5587 * This mean this cache had no attribute written. Therefore, no point
5588 * in copying default values around
5590 if (!root_cache->max_attr_size)
5593 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5596 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5599 if (!attr || !attr->store || !attr->show)
5603 * It is really bad that we have to allocate here, so we will
5604 * do it only as a fallback. If we actually allocate, though,
5605 * we can just use the allocated buffer until the end.
5607 * Most of the slub attributes will tend to be very small in
5608 * size, but sysfs allows buffers up to a page, so they can
5609 * theoretically happen.
5613 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf) &&
5614 !IS_ENABLED(CONFIG_SLUB_STATS))
5617 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5618 if (WARN_ON(!buffer))
5623 len = attr->show(root_cache, buf);
5625 attr->store(s, buf, len);
5629 free_page((unsigned long)buffer);
5633 static void kmem_cache_release(struct kobject *k)
5635 slab_kmem_cache_release(to_slab(k));
5638 static const struct sysfs_ops slab_sysfs_ops = {
5639 .show = slab_attr_show,
5640 .store = slab_attr_store,
5643 static struct kobj_type slab_ktype = {
5644 .sysfs_ops = &slab_sysfs_ops,
5645 .release = kmem_cache_release,
5648 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5650 struct kobj_type *ktype = get_ktype(kobj);
5652 if (ktype == &slab_ktype)
5657 static const struct kset_uevent_ops slab_uevent_ops = {
5658 .filter = uevent_filter,
5661 static struct kset *slab_kset;
5663 static inline struct kset *cache_kset(struct kmem_cache *s)
5666 if (!is_root_cache(s))
5667 return s->memcg_params.root_cache->memcg_kset;
5672 #define ID_STR_LENGTH 64
5674 /* Create a unique string id for a slab cache:
5676 * Format :[flags-]size
5678 static char *create_unique_id(struct kmem_cache *s)
5680 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5687 * First flags affecting slabcache operations. We will only
5688 * get here for aliasable slabs so we do not need to support
5689 * too many flags. The flags here must cover all flags that
5690 * are matched during merging to guarantee that the id is
5693 if (s->flags & SLAB_CACHE_DMA)
5695 if (s->flags & SLAB_CACHE_DMA32)
5697 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5699 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5701 if (s->flags & SLAB_ACCOUNT)
5705 p += sprintf(p, "%07u", s->size);
5707 BUG_ON(p > name + ID_STR_LENGTH - 1);
5711 static void sysfs_slab_remove_workfn(struct work_struct *work)
5713 struct kmem_cache *s =
5714 container_of(work, struct kmem_cache, kobj_remove_work);
5716 if (!s->kobj.state_in_sysfs)
5718 * For a memcg cache, this may be called during
5719 * deactivation and again on shutdown. Remove only once.
5720 * A cache is never shut down before deactivation is
5721 * complete, so no need to worry about synchronization.
5726 kset_unregister(s->memcg_kset);
5728 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5730 kobject_put(&s->kobj);
5733 static int sysfs_slab_add(struct kmem_cache *s)
5737 struct kset *kset = cache_kset(s);
5738 int unmergeable = slab_unmergeable(s);
5740 INIT_WORK(&s->kobj_remove_work, sysfs_slab_remove_workfn);
5743 kobject_init(&s->kobj, &slab_ktype);
5747 if (!unmergeable && disable_higher_order_debug &&
5748 (slub_debug & DEBUG_METADATA_FLAGS))
5753 * Slabcache can never be merged so we can use the name proper.
5754 * This is typically the case for debug situations. In that
5755 * case we can catch duplicate names easily.
5757 sysfs_remove_link(&slab_kset->kobj, s->name);
5761 * Create a unique name for the slab as a target
5764 name = create_unique_id(s);
5767 s->kobj.kset = kset;
5768 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5772 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5777 if (is_root_cache(s) && memcg_sysfs_enabled) {
5778 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5779 if (!s->memcg_kset) {
5786 kobject_uevent(&s->kobj, KOBJ_ADD);
5788 /* Setup first alias */
5789 sysfs_slab_alias(s, s->name);
5796 kobject_del(&s->kobj);
5800 static void sysfs_slab_remove(struct kmem_cache *s)
5802 if (slab_state < FULL)
5804 * Sysfs has not been setup yet so no need to remove the
5809 kobject_get(&s->kobj);
5810 schedule_work(&s->kobj_remove_work);
5813 void sysfs_slab_unlink(struct kmem_cache *s)
5815 if (slab_state >= FULL)
5816 kobject_del(&s->kobj);
5819 void sysfs_slab_release(struct kmem_cache *s)
5821 if (slab_state >= FULL)
5822 kobject_put(&s->kobj);
5826 * Need to buffer aliases during bootup until sysfs becomes
5827 * available lest we lose that information.
5829 struct saved_alias {
5830 struct kmem_cache *s;
5832 struct saved_alias *next;
5835 static struct saved_alias *alias_list;
5837 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5839 struct saved_alias *al;
5841 if (slab_state == FULL) {
5843 * If we have a leftover link then remove it.
5845 sysfs_remove_link(&slab_kset->kobj, name);
5846 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5849 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5855 al->next = alias_list;
5860 static int __init slab_sysfs_init(void)
5862 struct kmem_cache *s;
5865 mutex_lock(&slab_mutex);
5867 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5869 mutex_unlock(&slab_mutex);
5870 pr_err("Cannot register slab subsystem.\n");
5876 list_for_each_entry(s, &slab_caches, list) {
5877 err = sysfs_slab_add(s);
5879 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5883 while (alias_list) {
5884 struct saved_alias *al = alias_list;
5886 alias_list = alias_list->next;
5887 err = sysfs_slab_alias(al->s, al->name);
5889 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5894 mutex_unlock(&slab_mutex);
5899 __initcall(slab_sysfs_init);
5900 #endif /* CONFIG_SYSFS */
5903 * The /proc/slabinfo ABI
5905 #ifdef CONFIG_SLUB_DEBUG
5906 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5908 unsigned long nr_slabs = 0;
5909 unsigned long nr_objs = 0;
5910 unsigned long nr_free = 0;
5912 struct kmem_cache_node *n;
5914 for_each_kmem_cache_node(s, node, n) {
5915 nr_slabs += node_nr_slabs(n);
5916 nr_objs += node_nr_objs(n);
5917 nr_free += count_partial(n, count_free);
5920 sinfo->active_objs = nr_objs - nr_free;
5921 sinfo->num_objs = nr_objs;
5922 sinfo->active_slabs = nr_slabs;
5923 sinfo->num_slabs = nr_slabs;
5924 sinfo->objects_per_slab = oo_objects(s->oo);
5925 sinfo->cache_order = oo_order(s->oo);
5928 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5932 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5933 size_t count, loff_t *ppos)
5937 #endif /* CONFIG_SLUB_DEBUG */