2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks or atomic operatios
6 * and only uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
9 * (C) 2011 Linux Foundation, Christoph Lameter
13 #include <linux/swap.h> /* struct reclaim_state */
14 #include <linux/module.h>
15 #include <linux/bit_spinlock.h>
16 #include <linux/interrupt.h>
17 #include <linux/bitops.h>
18 #include <linux/slab.h>
20 #include <linux/proc_fs.h>
21 #include <linux/notifier.h>
22 #include <linux/seq_file.h>
23 #include <linux/kmemcheck.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>
37 #include <trace/events/kmem.h>
43 * 1. slab_mutex (Global Mutex)
45 * 3. slab_lock(page) (Only on some arches and for debugging)
49 * The role of the slab_mutex is to protect the list of all the slabs
50 * and to synchronize major metadata changes to slab cache structures.
52 * The slab_lock is only used for debugging and on arches that do not
53 * have the ability to do a cmpxchg_double. It only protects the second
54 * double word in the page struct. Meaning
55 * A. page->freelist -> List of object free in a page
56 * B. page->counters -> Counters of objects
57 * C. page->frozen -> frozen state
59 * If a slab is frozen then it is exempt from list management. It is not
60 * on any list. The processor that froze the slab is the one who can
61 * perform list operations on the page. Other processors may put objects
62 * onto the freelist but the processor that froze the slab is the only
63 * one that can retrieve the objects from the page's freelist.
65 * The list_lock protects the partial and full list on each node and
66 * the partial slab counter. If taken then no new slabs may be added or
67 * removed from the lists nor make the number of partial slabs be modified.
68 * (Note that the total number of slabs is an atomic value that may be
69 * modified without taking the list lock).
71 * The list_lock is a centralized lock and thus we avoid taking it as
72 * much as possible. As long as SLUB does not have to handle partial
73 * slabs, operations can continue without any centralized lock. F.e.
74 * allocating a long series of objects that fill up slabs does not require
76 * Interrupts are disabled during allocation and deallocation in order to
77 * make the slab allocator safe to use in the context of an irq. In addition
78 * interrupts are disabled to ensure that the processor does not change
79 * while handling per_cpu slabs, due to kernel preemption.
81 * SLUB assigns one slab for allocation to each processor.
82 * Allocations only occur from these slabs called cpu slabs.
84 * Slabs with free elements are kept on a partial list and during regular
85 * operations no list for full slabs is used. If an object in a full slab is
86 * freed then the slab will show up again on the partial lists.
87 * We track full slabs for debugging purposes though because otherwise we
88 * cannot scan all objects.
90 * Slabs are freed when they become empty. Teardown and setup is
91 * minimal so we rely on the page allocators per cpu caches for
92 * fast frees and allocs.
94 * Overloading of page flags that are otherwise used for LRU management.
96 * PageActive The slab is frozen and exempt from list processing.
97 * This means that the slab is dedicated to a purpose
98 * such as satisfying allocations for a specific
99 * processor. Objects may be freed in the slab while
100 * it is frozen but slab_free will then skip the usual
101 * list operations. It is up to the processor holding
102 * the slab to integrate the slab into the slab lists
103 * when the slab is no longer needed.
105 * One use of this flag is to mark slabs that are
106 * used for allocations. Then such a slab becomes a cpu
107 * slab. The cpu slab may be equipped with an additional
108 * freelist that allows lockless access to
109 * free objects in addition to the regular freelist
110 * that requires the slab lock.
112 * PageError Slab requires special handling due to debug
113 * options set. This moves slab handling out of
114 * the fast path and disables lockless freelists.
117 static inline int kmem_cache_debug(struct kmem_cache *s)
119 #ifdef CONFIG_SLUB_DEBUG
120 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
126 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
128 #ifdef CONFIG_SLUB_CPU_PARTIAL
129 return !kmem_cache_debug(s);
136 * Issues still to be resolved:
138 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
140 * - Variable sizing of the per node arrays
143 /* Enable to test recovery from slab corruption on boot */
144 #undef SLUB_RESILIENCY_TEST
146 /* Enable to log cmpxchg failures */
147 #undef SLUB_DEBUG_CMPXCHG
150 * Mininum number of partial slabs. These will be left on the partial
151 * lists even if they are empty. kmem_cache_shrink may reclaim them.
153 #define MIN_PARTIAL 5
156 * Maximum number of desirable partial slabs.
157 * The existence of more partial slabs makes kmem_cache_shrink
158 * sort the partial list by the number of objects in use.
160 #define MAX_PARTIAL 10
162 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
163 SLAB_POISON | SLAB_STORE_USER)
166 * Debugging flags that require metadata to be stored in the slab. These get
167 * disabled when slub_debug=O is used and a cache's min order increases with
170 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
173 #define OO_MASK ((1 << OO_SHIFT) - 1)
174 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
176 /* Internal SLUB flags */
177 #define __OBJECT_POISON 0x80000000UL /* Poison object */
178 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
181 static struct notifier_block slab_notifier;
185 * Tracking user of a slab.
187 #define TRACK_ADDRS_COUNT 16
189 unsigned long addr; /* Called from address */
190 #ifdef CONFIG_STACKTRACE
191 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
193 int cpu; /* Was running on cpu */
194 int pid; /* Pid context */
195 unsigned long when; /* When did the operation occur */
198 enum track_item { TRACK_ALLOC, TRACK_FREE };
201 static int sysfs_slab_add(struct kmem_cache *);
202 static int sysfs_slab_alias(struct kmem_cache *, const char *);
203 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
205 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
206 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
208 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
211 static inline void stat(const struct kmem_cache *s, enum stat_item si)
213 #ifdef CONFIG_SLUB_STATS
215 * The rmw is racy on a preemptible kernel but this is acceptable, so
216 * avoid this_cpu_add()'s irq-disable overhead.
218 raw_cpu_inc(s->cpu_slab->stat[si]);
222 /********************************************************************
223 * Core slab cache functions
224 *******************************************************************/
226 /* Verify that a pointer has an address that is valid within a slab page */
227 static inline int check_valid_pointer(struct kmem_cache *s,
228 struct page *page, const void *object)
235 base = page_address(page);
236 if (object < base || object >= base + page->objects * s->size ||
237 (object - base) % s->size) {
244 static inline void *get_freepointer(struct kmem_cache *s, void *object)
246 return *(void **)(object + s->offset);
249 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
251 prefetch(object + s->offset);
254 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
258 #ifdef CONFIG_DEBUG_PAGEALLOC
259 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
261 p = get_freepointer(s, object);
266 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
268 *(void **)(object + s->offset) = fp;
271 /* Loop over all objects in a slab */
272 #define for_each_object(__p, __s, __addr, __objects) \
273 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
276 #define for_each_object_idx(__p, __idx, __s, __addr, __objects) \
277 for (__p = (__addr), __idx = 1; __idx <= __objects;\
278 __p += (__s)->size, __idx++)
280 /* Determine object index from a given position */
281 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
283 return (p - addr) / s->size;
286 static inline size_t slab_ksize(const struct kmem_cache *s)
288 #ifdef CONFIG_SLUB_DEBUG
290 * Debugging requires use of the padding between object
291 * and whatever may come after it.
293 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
294 return s->object_size;
298 * If we have the need to store the freelist pointer
299 * back there or track user information then we can
300 * only use the space before that information.
302 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
305 * Else we can use all the padding etc for the allocation
310 static inline int order_objects(int order, unsigned long size, int reserved)
312 return ((PAGE_SIZE << order) - reserved) / size;
315 static inline struct kmem_cache_order_objects oo_make(int order,
316 unsigned long size, int reserved)
318 struct kmem_cache_order_objects x = {
319 (order << OO_SHIFT) + order_objects(order, size, reserved)
325 static inline int oo_order(struct kmem_cache_order_objects x)
327 return x.x >> OO_SHIFT;
330 static inline int oo_objects(struct kmem_cache_order_objects x)
332 return x.x & OO_MASK;
336 * Per slab locking using the pagelock
338 static __always_inline void slab_lock(struct page *page)
340 bit_spin_lock(PG_locked, &page->flags);
343 static __always_inline void slab_unlock(struct page *page)
345 __bit_spin_unlock(PG_locked, &page->flags);
348 static inline void set_page_slub_counters(struct page *page, unsigned long counters_new)
351 tmp.counters = counters_new;
353 * page->counters can cover frozen/inuse/objects as well
354 * as page->_count. If we assign to ->counters directly
355 * we run the risk of losing updates to page->_count, so
356 * be careful and only assign to the fields we need.
358 page->frozen = tmp.frozen;
359 page->inuse = tmp.inuse;
360 page->objects = tmp.objects;
363 /* Interrupts must be disabled (for the fallback code to work right) */
364 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
365 void *freelist_old, unsigned long counters_old,
366 void *freelist_new, unsigned long counters_new,
369 VM_BUG_ON(!irqs_disabled());
370 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
371 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
372 if (s->flags & __CMPXCHG_DOUBLE) {
373 if (cmpxchg_double(&page->freelist, &page->counters,
374 freelist_old, counters_old,
375 freelist_new, counters_new))
381 if (page->freelist == freelist_old &&
382 page->counters == counters_old) {
383 page->freelist = freelist_new;
384 set_page_slub_counters(page, counters_new);
392 stat(s, CMPXCHG_DOUBLE_FAIL);
394 #ifdef SLUB_DEBUG_CMPXCHG
395 pr_info("%s %s: cmpxchg double redo ", n, s->name);
401 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
402 void *freelist_old, unsigned long counters_old,
403 void *freelist_new, unsigned long counters_new,
406 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
407 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
408 if (s->flags & __CMPXCHG_DOUBLE) {
409 if (cmpxchg_double(&page->freelist, &page->counters,
410 freelist_old, counters_old,
411 freelist_new, counters_new))
418 local_irq_save(flags);
420 if (page->freelist == freelist_old &&
421 page->counters == counters_old) {
422 page->freelist = freelist_new;
423 set_page_slub_counters(page, counters_new);
425 local_irq_restore(flags);
429 local_irq_restore(flags);
433 stat(s, CMPXCHG_DOUBLE_FAIL);
435 #ifdef SLUB_DEBUG_CMPXCHG
436 pr_info("%s %s: cmpxchg double redo ", n, s->name);
442 #ifdef CONFIG_SLUB_DEBUG
444 * Determine a map of object in use on a page.
446 * Node listlock must be held to guarantee that the page does
447 * not vanish from under us.
449 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
452 void *addr = page_address(page);
454 for (p = page->freelist; p; p = get_freepointer(s, p))
455 set_bit(slab_index(p, s, addr), map);
461 #ifdef CONFIG_SLUB_DEBUG_ON
462 static int slub_debug = DEBUG_DEFAULT_FLAGS;
464 static int slub_debug;
467 static char *slub_debug_slabs;
468 static int disable_higher_order_debug;
473 static void print_section(char *text, u8 *addr, unsigned int length)
475 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
479 static struct track *get_track(struct kmem_cache *s, void *object,
480 enum track_item alloc)
485 p = object + s->offset + sizeof(void *);
487 p = object + s->inuse;
492 static void set_track(struct kmem_cache *s, void *object,
493 enum track_item alloc, unsigned long addr)
495 struct track *p = get_track(s, object, alloc);
498 #ifdef CONFIG_STACKTRACE
499 struct stack_trace trace;
502 trace.nr_entries = 0;
503 trace.max_entries = TRACK_ADDRS_COUNT;
504 trace.entries = p->addrs;
506 save_stack_trace(&trace);
508 /* See rant in lockdep.c */
509 if (trace.nr_entries != 0 &&
510 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
513 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
517 p->cpu = smp_processor_id();
518 p->pid = current->pid;
521 memset(p, 0, sizeof(struct track));
524 static void init_tracking(struct kmem_cache *s, void *object)
526 if (!(s->flags & SLAB_STORE_USER))
529 set_track(s, object, TRACK_FREE, 0UL);
530 set_track(s, object, TRACK_ALLOC, 0UL);
533 static void print_track(const char *s, struct track *t)
538 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
539 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
540 #ifdef CONFIG_STACKTRACE
543 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
545 pr_err("\t%pS\n", (void *)t->addrs[i]);
552 static void print_tracking(struct kmem_cache *s, void *object)
554 if (!(s->flags & SLAB_STORE_USER))
557 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
558 print_track("Freed", get_track(s, object, TRACK_FREE));
561 static void print_page_info(struct page *page)
563 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
564 page, page->objects, page->inuse, page->freelist, page->flags);
568 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
570 struct va_format vaf;
576 pr_err("=============================================================================\n");
577 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
578 pr_err("-----------------------------------------------------------------------------\n\n");
580 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
584 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
586 struct va_format vaf;
592 pr_err("FIX %s: %pV\n", s->name, &vaf);
596 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
598 unsigned int off; /* Offset of last byte */
599 u8 *addr = page_address(page);
601 print_tracking(s, p);
603 print_page_info(page);
605 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
606 p, p - addr, get_freepointer(s, p));
609 print_section("Bytes b4 ", p - 16, 16);
611 print_section("Object ", p, min_t(unsigned long, s->object_size,
613 if (s->flags & SLAB_RED_ZONE)
614 print_section("Redzone ", p + s->object_size,
615 s->inuse - s->object_size);
618 off = s->offset + sizeof(void *);
622 if (s->flags & SLAB_STORE_USER)
623 off += 2 * sizeof(struct track);
626 /* Beginning of the filler is the free pointer */
627 print_section("Padding ", p + off, s->size - off);
632 static void object_err(struct kmem_cache *s, struct page *page,
633 u8 *object, char *reason)
635 slab_bug(s, "%s", reason);
636 print_trailer(s, page, object);
639 static void slab_err(struct kmem_cache *s, struct page *page,
640 const char *fmt, ...)
646 vsnprintf(buf, sizeof(buf), fmt, args);
648 slab_bug(s, "%s", buf);
649 print_page_info(page);
653 static void init_object(struct kmem_cache *s, void *object, u8 val)
657 if (s->flags & __OBJECT_POISON) {
658 memset(p, POISON_FREE, s->object_size - 1);
659 p[s->object_size - 1] = POISON_END;
662 if (s->flags & SLAB_RED_ZONE)
663 memset(p + s->object_size, val, s->inuse - s->object_size);
666 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
667 void *from, void *to)
669 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
670 memset(from, data, to - from);
673 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
674 u8 *object, char *what,
675 u8 *start, unsigned int value, unsigned int bytes)
680 fault = memchr_inv(start, value, bytes);
685 while (end > fault && end[-1] == value)
688 slab_bug(s, "%s overwritten", what);
689 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
690 fault, end - 1, fault[0], value);
691 print_trailer(s, page, object);
693 restore_bytes(s, what, value, fault, end);
701 * Bytes of the object to be managed.
702 * If the freepointer may overlay the object then the free
703 * pointer is the first word of the object.
705 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
708 * object + s->object_size
709 * Padding to reach word boundary. This is also used for Redzoning.
710 * Padding is extended by another word if Redzoning is enabled and
711 * object_size == inuse.
713 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
714 * 0xcc (RED_ACTIVE) for objects in use.
717 * Meta data starts here.
719 * A. Free pointer (if we cannot overwrite object on free)
720 * B. Tracking data for SLAB_STORE_USER
721 * C. Padding to reach required alignment boundary or at mininum
722 * one word if debugging is on to be able to detect writes
723 * before the word boundary.
725 * Padding is done using 0x5a (POISON_INUSE)
728 * Nothing is used beyond s->size.
730 * If slabcaches are merged then the object_size and inuse boundaries are mostly
731 * ignored. And therefore no slab options that rely on these boundaries
732 * may be used with merged slabcaches.
735 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
737 unsigned long off = s->inuse; /* The end of info */
740 /* Freepointer is placed after the object. */
741 off += sizeof(void *);
743 if (s->flags & SLAB_STORE_USER)
744 /* We also have user information there */
745 off += 2 * sizeof(struct track);
750 return check_bytes_and_report(s, page, p, "Object padding",
751 p + off, POISON_INUSE, s->size - off);
754 /* Check the pad bytes at the end of a slab page */
755 static int slab_pad_check(struct kmem_cache *s, struct page *page)
763 if (!(s->flags & SLAB_POISON))
766 start = page_address(page);
767 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
768 end = start + length;
769 remainder = length % s->size;
773 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
776 while (end > fault && end[-1] == POISON_INUSE)
779 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
780 print_section("Padding ", end - remainder, remainder);
782 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
786 static int check_object(struct kmem_cache *s, struct page *page,
787 void *object, u8 val)
790 u8 *endobject = object + s->object_size;
792 if (s->flags & SLAB_RED_ZONE) {
793 if (!check_bytes_and_report(s, page, object, "Redzone",
794 endobject, val, s->inuse - s->object_size))
797 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
798 check_bytes_and_report(s, page, p, "Alignment padding",
799 endobject, POISON_INUSE,
800 s->inuse - s->object_size);
804 if (s->flags & SLAB_POISON) {
805 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
806 (!check_bytes_and_report(s, page, p, "Poison", p,
807 POISON_FREE, s->object_size - 1) ||
808 !check_bytes_and_report(s, page, p, "Poison",
809 p + s->object_size - 1, POISON_END, 1)))
812 * check_pad_bytes cleans up on its own.
814 check_pad_bytes(s, page, p);
817 if (!s->offset && val == SLUB_RED_ACTIVE)
819 * Object and freepointer overlap. Cannot check
820 * freepointer while object is allocated.
824 /* Check free pointer validity */
825 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
826 object_err(s, page, p, "Freepointer corrupt");
828 * No choice but to zap it and thus lose the remainder
829 * of the free objects in this slab. May cause
830 * another error because the object count is now wrong.
832 set_freepointer(s, p, NULL);
838 static int check_slab(struct kmem_cache *s, struct page *page)
842 VM_BUG_ON(!irqs_disabled());
844 if (!PageSlab(page)) {
845 slab_err(s, page, "Not a valid slab page");
849 maxobj = order_objects(compound_order(page), s->size, s->reserved);
850 if (page->objects > maxobj) {
851 slab_err(s, page, "objects %u > max %u",
852 page->objects, maxobj);
855 if (page->inuse > page->objects) {
856 slab_err(s, page, "inuse %u > max %u",
857 page->inuse, page->objects);
860 /* Slab_pad_check fixes things up after itself */
861 slab_pad_check(s, page);
866 * Determine if a certain object on a page is on the freelist. Must hold the
867 * slab lock to guarantee that the chains are in a consistent state.
869 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
877 while (fp && nr <= page->objects) {
880 if (!check_valid_pointer(s, page, fp)) {
882 object_err(s, page, object,
883 "Freechain corrupt");
884 set_freepointer(s, object, NULL);
886 slab_err(s, page, "Freepointer corrupt");
887 page->freelist = NULL;
888 page->inuse = page->objects;
889 slab_fix(s, "Freelist cleared");
895 fp = get_freepointer(s, object);
899 max_objects = order_objects(compound_order(page), s->size, s->reserved);
900 if (max_objects > MAX_OBJS_PER_PAGE)
901 max_objects = MAX_OBJS_PER_PAGE;
903 if (page->objects != max_objects) {
904 slab_err(s, page, "Wrong number of objects. Found %d but "
905 "should be %d", page->objects, max_objects);
906 page->objects = max_objects;
907 slab_fix(s, "Number of objects adjusted.");
909 if (page->inuse != page->objects - nr) {
910 slab_err(s, page, "Wrong object count. Counter is %d but "
911 "counted were %d", page->inuse, page->objects - nr);
912 page->inuse = page->objects - nr;
913 slab_fix(s, "Object count adjusted.");
915 return search == NULL;
918 static void trace(struct kmem_cache *s, struct page *page, void *object,
921 if (s->flags & SLAB_TRACE) {
922 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
924 alloc ? "alloc" : "free",
929 print_section("Object ", (void *)object,
937 * Tracking of fully allocated slabs for debugging purposes.
939 static void add_full(struct kmem_cache *s,
940 struct kmem_cache_node *n, struct page *page)
942 if (!(s->flags & SLAB_STORE_USER))
945 lockdep_assert_held(&n->list_lock);
946 list_add(&page->lru, &n->full);
949 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
951 if (!(s->flags & SLAB_STORE_USER))
954 lockdep_assert_held(&n->list_lock);
955 list_del(&page->lru);
958 /* Tracking of the number of slabs for debugging purposes */
959 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
961 struct kmem_cache_node *n = get_node(s, node);
963 return atomic_long_read(&n->nr_slabs);
966 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
968 return atomic_long_read(&n->nr_slabs);
971 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
973 struct kmem_cache_node *n = get_node(s, node);
976 * May be called early in order to allocate a slab for the
977 * kmem_cache_node structure. Solve the chicken-egg
978 * dilemma by deferring the increment of the count during
979 * bootstrap (see early_kmem_cache_node_alloc).
982 atomic_long_inc(&n->nr_slabs);
983 atomic_long_add(objects, &n->total_objects);
986 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
988 struct kmem_cache_node *n = get_node(s, node);
990 atomic_long_dec(&n->nr_slabs);
991 atomic_long_sub(objects, &n->total_objects);
994 /* Object debug checks for alloc/free paths */
995 static void setup_object_debug(struct kmem_cache *s, struct page *page,
998 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1001 init_object(s, object, SLUB_RED_INACTIVE);
1002 init_tracking(s, object);
1005 static noinline int alloc_debug_processing(struct kmem_cache *s,
1007 void *object, unsigned long addr)
1009 if (!check_slab(s, page))
1012 if (!check_valid_pointer(s, page, object)) {
1013 object_err(s, page, object, "Freelist Pointer check fails");
1017 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1020 /* Success perform special debug activities for allocs */
1021 if (s->flags & SLAB_STORE_USER)
1022 set_track(s, object, TRACK_ALLOC, addr);
1023 trace(s, page, object, 1);
1024 init_object(s, object, SLUB_RED_ACTIVE);
1028 if (PageSlab(page)) {
1030 * If this is a slab page then lets do the best we can
1031 * to avoid issues in the future. Marking all objects
1032 * as used avoids touching the remaining objects.
1034 slab_fix(s, "Marking all objects used");
1035 page->inuse = page->objects;
1036 page->freelist = NULL;
1041 static noinline struct kmem_cache_node *free_debug_processing(
1042 struct kmem_cache *s, struct page *page, void *object,
1043 unsigned long addr, unsigned long *flags)
1045 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1047 spin_lock_irqsave(&n->list_lock, *flags);
1050 if (!check_slab(s, page))
1053 if (!check_valid_pointer(s, page, object)) {
1054 slab_err(s, page, "Invalid object pointer 0x%p", object);
1058 if (on_freelist(s, page, object)) {
1059 object_err(s, page, object, "Object already free");
1063 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1066 if (unlikely(s != page->slab_cache)) {
1067 if (!PageSlab(page)) {
1068 slab_err(s, page, "Attempt to free object(0x%p) "
1069 "outside of slab", object);
1070 } else if (!page->slab_cache) {
1071 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1075 object_err(s, page, object,
1076 "page slab pointer corrupt.");
1080 if (s->flags & SLAB_STORE_USER)
1081 set_track(s, object, TRACK_FREE, addr);
1082 trace(s, page, object, 0);
1083 init_object(s, object, SLUB_RED_INACTIVE);
1087 * Keep node_lock to preserve integrity
1088 * until the object is actually freed
1094 spin_unlock_irqrestore(&n->list_lock, *flags);
1095 slab_fix(s, "Object at 0x%p not freed", object);
1099 static int __init setup_slub_debug(char *str)
1101 slub_debug = DEBUG_DEFAULT_FLAGS;
1102 if (*str++ != '=' || !*str)
1104 * No options specified. Switch on full debugging.
1110 * No options but restriction on slabs. This means full
1111 * debugging for slabs matching a pattern.
1115 if (tolower(*str) == 'o') {
1117 * Avoid enabling debugging on caches if its minimum order
1118 * would increase as a result.
1120 disable_higher_order_debug = 1;
1127 * Switch off all debugging measures.
1132 * Determine which debug features should be switched on
1134 for (; *str && *str != ','; str++) {
1135 switch (tolower(*str)) {
1137 slub_debug |= SLAB_DEBUG_FREE;
1140 slub_debug |= SLAB_RED_ZONE;
1143 slub_debug |= SLAB_POISON;
1146 slub_debug |= SLAB_STORE_USER;
1149 slub_debug |= SLAB_TRACE;
1152 slub_debug |= SLAB_FAILSLAB;
1155 pr_err("slub_debug option '%c' unknown. skipped\n",
1162 slub_debug_slabs = str + 1;
1167 __setup("slub_debug", setup_slub_debug);
1169 unsigned long kmem_cache_flags(unsigned long object_size,
1170 unsigned long flags, const char *name,
1171 void (*ctor)(void *))
1174 * Enable debugging if selected on the kernel commandline.
1176 if (slub_debug && (!slub_debug_slabs || (name &&
1177 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1178 flags |= slub_debug;
1183 static inline void setup_object_debug(struct kmem_cache *s,
1184 struct page *page, void *object) {}
1186 static inline int alloc_debug_processing(struct kmem_cache *s,
1187 struct page *page, void *object, unsigned long addr) { return 0; }
1189 static inline struct kmem_cache_node *free_debug_processing(
1190 struct kmem_cache *s, struct page *page, void *object,
1191 unsigned long addr, unsigned long *flags) { return NULL; }
1193 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1195 static inline int check_object(struct kmem_cache *s, struct page *page,
1196 void *object, u8 val) { return 1; }
1197 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1198 struct page *page) {}
1199 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1200 struct page *page) {}
1201 unsigned long kmem_cache_flags(unsigned long object_size,
1202 unsigned long flags, const char *name,
1203 void (*ctor)(void *))
1207 #define slub_debug 0
1209 #define disable_higher_order_debug 0
1211 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1213 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1215 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1217 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1220 #endif /* CONFIG_SLUB_DEBUG */
1223 * Hooks for other subsystems that check memory allocations. In a typical
1224 * production configuration these hooks all should produce no code at all.
1226 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1228 kmemleak_alloc(ptr, size, 1, flags);
1231 static inline void kfree_hook(const void *x)
1236 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1238 flags &= gfp_allowed_mask;
1239 lockdep_trace_alloc(flags);
1240 might_sleep_if(flags & __GFP_WAIT);
1242 return should_failslab(s->object_size, flags, s->flags);
1245 static inline void slab_post_alloc_hook(struct kmem_cache *s,
1246 gfp_t flags, void *object)
1248 flags &= gfp_allowed_mask;
1249 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
1250 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
1253 static inline void slab_free_hook(struct kmem_cache *s, void *x)
1255 kmemleak_free_recursive(x, s->flags);
1258 * Trouble is that we may no longer disable interrupts in the fast path
1259 * So in order to make the debug calls that expect irqs to be
1260 * disabled we need to disable interrupts temporarily.
1262 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1264 unsigned long flags;
1266 local_irq_save(flags);
1267 kmemcheck_slab_free(s, x, s->object_size);
1268 debug_check_no_locks_freed(x, s->object_size);
1269 local_irq_restore(flags);
1272 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1273 debug_check_no_obj_freed(x, s->object_size);
1277 * Slab allocation and freeing
1279 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1280 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1283 int order = oo_order(oo);
1285 flags |= __GFP_NOTRACK;
1287 if (memcg_charge_slab(s, flags, order))
1290 if (node == NUMA_NO_NODE)
1291 page = alloc_pages(flags, order);
1293 page = alloc_pages_exact_node(node, flags, order);
1296 memcg_uncharge_slab(s, order);
1301 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1304 struct kmem_cache_order_objects oo = s->oo;
1307 flags &= gfp_allowed_mask;
1309 if (flags & __GFP_WAIT)
1312 flags |= s->allocflags;
1315 * Let the initial higher-order allocation fail under memory pressure
1316 * so we fall-back to the minimum order allocation.
1318 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1320 page = alloc_slab_page(s, alloc_gfp, node, oo);
1321 if (unlikely(!page)) {
1325 * Allocation may have failed due to fragmentation.
1326 * Try a lower order alloc if possible
1328 page = alloc_slab_page(s, alloc_gfp, node, oo);
1331 stat(s, ORDER_FALLBACK);
1334 if (kmemcheck_enabled && page
1335 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1336 int pages = 1 << oo_order(oo);
1338 kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node);
1341 * Objects from caches that have a constructor don't get
1342 * cleared when they're allocated, so we need to do it here.
1345 kmemcheck_mark_uninitialized_pages(page, pages);
1347 kmemcheck_mark_unallocated_pages(page, pages);
1350 if (flags & __GFP_WAIT)
1351 local_irq_disable();
1355 page->objects = oo_objects(oo);
1356 mod_zone_page_state(page_zone(page),
1357 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1358 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1364 static void setup_object(struct kmem_cache *s, struct page *page,
1367 setup_object_debug(s, page, object);
1368 if (unlikely(s->ctor))
1372 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1380 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1381 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
1385 page = allocate_slab(s,
1386 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1390 order = compound_order(page);
1391 inc_slabs_node(s, page_to_nid(page), page->objects);
1392 page->slab_cache = s;
1393 __SetPageSlab(page);
1394 if (page->pfmemalloc)
1395 SetPageSlabPfmemalloc(page);
1397 start = page_address(page);
1399 if (unlikely(s->flags & SLAB_POISON))
1400 memset(start, POISON_INUSE, PAGE_SIZE << order);
1402 for_each_object_idx(p, idx, s, start, page->objects) {
1403 setup_object(s, page, p);
1404 if (likely(idx < page->objects))
1405 set_freepointer(s, p, p + s->size);
1407 set_freepointer(s, p, NULL);
1410 page->freelist = start;
1411 page->inuse = page->objects;
1417 static void __free_slab(struct kmem_cache *s, struct page *page)
1419 int order = compound_order(page);
1420 int pages = 1 << order;
1422 if (kmem_cache_debug(s)) {
1425 slab_pad_check(s, page);
1426 for_each_object(p, s, page_address(page),
1428 check_object(s, page, p, SLUB_RED_INACTIVE);
1431 kmemcheck_free_shadow(page, compound_order(page));
1433 mod_zone_page_state(page_zone(page),
1434 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1435 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1438 __ClearPageSlabPfmemalloc(page);
1439 __ClearPageSlab(page);
1441 page_mapcount_reset(page);
1442 if (current->reclaim_state)
1443 current->reclaim_state->reclaimed_slab += pages;
1444 __free_pages(page, order);
1445 memcg_uncharge_slab(s, order);
1448 #define need_reserve_slab_rcu \
1449 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1451 static void rcu_free_slab(struct rcu_head *h)
1455 if (need_reserve_slab_rcu)
1456 page = virt_to_head_page(h);
1458 page = container_of((struct list_head *)h, struct page, lru);
1460 __free_slab(page->slab_cache, page);
1463 static void free_slab(struct kmem_cache *s, struct page *page)
1465 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1466 struct rcu_head *head;
1468 if (need_reserve_slab_rcu) {
1469 int order = compound_order(page);
1470 int offset = (PAGE_SIZE << order) - s->reserved;
1472 VM_BUG_ON(s->reserved != sizeof(*head));
1473 head = page_address(page) + offset;
1476 * RCU free overloads the RCU head over the LRU
1478 head = (void *)&page->lru;
1481 call_rcu(head, rcu_free_slab);
1483 __free_slab(s, page);
1486 static void discard_slab(struct kmem_cache *s, struct page *page)
1488 dec_slabs_node(s, page_to_nid(page), page->objects);
1493 * Management of partially allocated slabs.
1496 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1499 if (tail == DEACTIVATE_TO_TAIL)
1500 list_add_tail(&page->lru, &n->partial);
1502 list_add(&page->lru, &n->partial);
1505 static inline void add_partial(struct kmem_cache_node *n,
1506 struct page *page, int tail)
1508 lockdep_assert_held(&n->list_lock);
1509 __add_partial(n, page, tail);
1513 __remove_partial(struct kmem_cache_node *n, struct page *page)
1515 list_del(&page->lru);
1519 static inline void remove_partial(struct kmem_cache_node *n,
1522 lockdep_assert_held(&n->list_lock);
1523 __remove_partial(n, page);
1527 * Remove slab from the partial list, freeze it and
1528 * return the pointer to the freelist.
1530 * Returns a list of objects or NULL if it fails.
1532 static inline void *acquire_slab(struct kmem_cache *s,
1533 struct kmem_cache_node *n, struct page *page,
1534 int mode, int *objects)
1537 unsigned long counters;
1540 lockdep_assert_held(&n->list_lock);
1543 * Zap the freelist and set the frozen bit.
1544 * The old freelist is the list of objects for the
1545 * per cpu allocation list.
1547 freelist = page->freelist;
1548 counters = page->counters;
1549 new.counters = counters;
1550 *objects = new.objects - new.inuse;
1552 new.inuse = page->objects;
1553 new.freelist = NULL;
1555 new.freelist = freelist;
1558 VM_BUG_ON(new.frozen);
1561 if (!__cmpxchg_double_slab(s, page,
1563 new.freelist, new.counters,
1567 remove_partial(n, page);
1572 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1573 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1576 * Try to allocate a partial slab from a specific node.
1578 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1579 struct kmem_cache_cpu *c, gfp_t flags)
1581 struct page *page, *page2;
1582 void *object = NULL;
1587 * Racy check. If we mistakenly see no partial slabs then we
1588 * just allocate an empty slab. If we mistakenly try to get a
1589 * partial slab and there is none available then get_partials()
1592 if (!n || !n->nr_partial)
1595 spin_lock(&n->list_lock);
1596 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1599 if (!pfmemalloc_match(page, flags))
1602 t = acquire_slab(s, n, page, object == NULL, &objects);
1606 available += objects;
1609 stat(s, ALLOC_FROM_PARTIAL);
1612 put_cpu_partial(s, page, 0);
1613 stat(s, CPU_PARTIAL_NODE);
1615 if (!kmem_cache_has_cpu_partial(s)
1616 || available > s->cpu_partial / 2)
1620 spin_unlock(&n->list_lock);
1625 * Get a page from somewhere. Search in increasing NUMA distances.
1627 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1628 struct kmem_cache_cpu *c)
1631 struct zonelist *zonelist;
1634 enum zone_type high_zoneidx = gfp_zone(flags);
1636 unsigned int cpuset_mems_cookie;
1639 * The defrag ratio allows a configuration of the tradeoffs between
1640 * inter node defragmentation and node local allocations. A lower
1641 * defrag_ratio increases the tendency to do local allocations
1642 * instead of attempting to obtain partial slabs from other nodes.
1644 * If the defrag_ratio is set to 0 then kmalloc() always
1645 * returns node local objects. If the ratio is higher then kmalloc()
1646 * may return off node objects because partial slabs are obtained
1647 * from other nodes and filled up.
1649 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1650 * defrag_ratio = 1000) then every (well almost) allocation will
1651 * first attempt to defrag slab caches on other nodes. This means
1652 * scanning over all nodes to look for partial slabs which may be
1653 * expensive if we do it every time we are trying to find a slab
1654 * with available objects.
1656 if (!s->remote_node_defrag_ratio ||
1657 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1661 cpuset_mems_cookie = read_mems_allowed_begin();
1662 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1663 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1664 struct kmem_cache_node *n;
1666 n = get_node(s, zone_to_nid(zone));
1668 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1669 n->nr_partial > s->min_partial) {
1670 object = get_partial_node(s, n, c, flags);
1673 * Don't check read_mems_allowed_retry()
1674 * here - if mems_allowed was updated in
1675 * parallel, that was a harmless race
1676 * between allocation and the cpuset
1683 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1689 * Get a partial page, lock it and return it.
1691 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1692 struct kmem_cache_cpu *c)
1695 int searchnode = node;
1697 if (node == NUMA_NO_NODE)
1698 searchnode = numa_mem_id();
1699 else if (!node_present_pages(node))
1700 searchnode = node_to_mem_node(node);
1702 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1703 if (object || node != NUMA_NO_NODE)
1706 return get_any_partial(s, flags, c);
1709 #ifdef CONFIG_PREEMPT
1711 * Calculate the next globally unique transaction for disambiguiation
1712 * during cmpxchg. The transactions start with the cpu number and are then
1713 * incremented by CONFIG_NR_CPUS.
1715 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1718 * No preemption supported therefore also no need to check for
1724 static inline unsigned long next_tid(unsigned long tid)
1726 return tid + TID_STEP;
1729 static inline unsigned int tid_to_cpu(unsigned long tid)
1731 return tid % TID_STEP;
1734 static inline unsigned long tid_to_event(unsigned long tid)
1736 return tid / TID_STEP;
1739 static inline unsigned int init_tid(int cpu)
1744 static inline void note_cmpxchg_failure(const char *n,
1745 const struct kmem_cache *s, unsigned long tid)
1747 #ifdef SLUB_DEBUG_CMPXCHG
1748 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1750 pr_info("%s %s: cmpxchg redo ", n, s->name);
1752 #ifdef CONFIG_PREEMPT
1753 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1754 pr_warn("due to cpu change %d -> %d\n",
1755 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1758 if (tid_to_event(tid) != tid_to_event(actual_tid))
1759 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1760 tid_to_event(tid), tid_to_event(actual_tid));
1762 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1763 actual_tid, tid, next_tid(tid));
1765 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1768 static void init_kmem_cache_cpus(struct kmem_cache *s)
1772 for_each_possible_cpu(cpu)
1773 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1777 * Remove the cpu slab
1779 static void deactivate_slab(struct kmem_cache *s, struct page *page,
1782 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1783 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1785 enum slab_modes l = M_NONE, m = M_NONE;
1787 int tail = DEACTIVATE_TO_HEAD;
1791 if (page->freelist) {
1792 stat(s, DEACTIVATE_REMOTE_FREES);
1793 tail = DEACTIVATE_TO_TAIL;
1797 * Stage one: Free all available per cpu objects back
1798 * to the page freelist while it is still frozen. Leave the
1801 * There is no need to take the list->lock because the page
1804 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1806 unsigned long counters;
1809 prior = page->freelist;
1810 counters = page->counters;
1811 set_freepointer(s, freelist, prior);
1812 new.counters = counters;
1814 VM_BUG_ON(!new.frozen);
1816 } while (!__cmpxchg_double_slab(s, page,
1818 freelist, new.counters,
1819 "drain percpu freelist"));
1821 freelist = nextfree;
1825 * Stage two: Ensure that the page is unfrozen while the
1826 * list presence reflects the actual number of objects
1829 * We setup the list membership and then perform a cmpxchg
1830 * with the count. If there is a mismatch then the page
1831 * is not unfrozen but the page is on the wrong list.
1833 * Then we restart the process which may have to remove
1834 * the page from the list that we just put it on again
1835 * because the number of objects in the slab may have
1840 old.freelist = page->freelist;
1841 old.counters = page->counters;
1842 VM_BUG_ON(!old.frozen);
1844 /* Determine target state of the slab */
1845 new.counters = old.counters;
1848 set_freepointer(s, freelist, old.freelist);
1849 new.freelist = freelist;
1851 new.freelist = old.freelist;
1855 if (!new.inuse && n->nr_partial >= s->min_partial)
1857 else if (new.freelist) {
1862 * Taking the spinlock removes the possiblity
1863 * that acquire_slab() will see a slab page that
1866 spin_lock(&n->list_lock);
1870 if (kmem_cache_debug(s) && !lock) {
1873 * This also ensures that the scanning of full
1874 * slabs from diagnostic functions will not see
1877 spin_lock(&n->list_lock);
1885 remove_partial(n, page);
1887 else if (l == M_FULL)
1889 remove_full(s, n, page);
1891 if (m == M_PARTIAL) {
1893 add_partial(n, page, tail);
1896 } else if (m == M_FULL) {
1898 stat(s, DEACTIVATE_FULL);
1899 add_full(s, n, page);
1905 if (!__cmpxchg_double_slab(s, page,
1906 old.freelist, old.counters,
1907 new.freelist, new.counters,
1912 spin_unlock(&n->list_lock);
1915 stat(s, DEACTIVATE_EMPTY);
1916 discard_slab(s, page);
1922 * Unfreeze all the cpu partial slabs.
1924 * This function must be called with interrupts disabled
1925 * for the cpu using c (or some other guarantee must be there
1926 * to guarantee no concurrent accesses).
1928 static void unfreeze_partials(struct kmem_cache *s,
1929 struct kmem_cache_cpu *c)
1931 #ifdef CONFIG_SLUB_CPU_PARTIAL
1932 struct kmem_cache_node *n = NULL, *n2 = NULL;
1933 struct page *page, *discard_page = NULL;
1935 while ((page = c->partial)) {
1939 c->partial = page->next;
1941 n2 = get_node(s, page_to_nid(page));
1944 spin_unlock(&n->list_lock);
1947 spin_lock(&n->list_lock);
1952 old.freelist = page->freelist;
1953 old.counters = page->counters;
1954 VM_BUG_ON(!old.frozen);
1956 new.counters = old.counters;
1957 new.freelist = old.freelist;
1961 } while (!__cmpxchg_double_slab(s, page,
1962 old.freelist, old.counters,
1963 new.freelist, new.counters,
1964 "unfreezing slab"));
1966 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
1967 page->next = discard_page;
1968 discard_page = page;
1970 add_partial(n, page, DEACTIVATE_TO_TAIL);
1971 stat(s, FREE_ADD_PARTIAL);
1976 spin_unlock(&n->list_lock);
1978 while (discard_page) {
1979 page = discard_page;
1980 discard_page = discard_page->next;
1982 stat(s, DEACTIVATE_EMPTY);
1983 discard_slab(s, page);
1990 * Put a page that was just frozen (in __slab_free) into a partial page
1991 * slot if available. This is done without interrupts disabled and without
1992 * preemption disabled. The cmpxchg is racy and may put the partial page
1993 * onto a random cpus partial slot.
1995 * If we did not find a slot then simply move all the partials to the
1996 * per node partial list.
1998 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2000 #ifdef CONFIG_SLUB_CPU_PARTIAL
2001 struct page *oldpage;
2008 oldpage = this_cpu_read(s->cpu_slab->partial);
2011 pobjects = oldpage->pobjects;
2012 pages = oldpage->pages;
2013 if (drain && pobjects > s->cpu_partial) {
2014 unsigned long flags;
2016 * partial array is full. Move the existing
2017 * set to the per node partial list.
2019 local_irq_save(flags);
2020 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2021 local_irq_restore(flags);
2025 stat(s, CPU_PARTIAL_DRAIN);
2030 pobjects += page->objects - page->inuse;
2032 page->pages = pages;
2033 page->pobjects = pobjects;
2034 page->next = oldpage;
2036 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2041 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2043 stat(s, CPUSLAB_FLUSH);
2044 deactivate_slab(s, c->page, c->freelist);
2046 c->tid = next_tid(c->tid);
2054 * Called from IPI handler with interrupts disabled.
2056 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2058 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2064 unfreeze_partials(s, c);
2068 static void flush_cpu_slab(void *d)
2070 struct kmem_cache *s = d;
2072 __flush_cpu_slab(s, smp_processor_id());
2075 static bool has_cpu_slab(int cpu, void *info)
2077 struct kmem_cache *s = info;
2078 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2080 return c->page || c->partial;
2083 static void flush_all(struct kmem_cache *s)
2085 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2089 * Check if the objects in a per cpu structure fit numa
2090 * locality expectations.
2092 static inline int node_match(struct page *page, int node)
2095 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2101 #ifdef CONFIG_SLUB_DEBUG
2102 static int count_free(struct page *page)
2104 return page->objects - page->inuse;
2107 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2109 return atomic_long_read(&n->total_objects);
2111 #endif /* CONFIG_SLUB_DEBUG */
2113 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2114 static unsigned long count_partial(struct kmem_cache_node *n,
2115 int (*get_count)(struct page *))
2117 unsigned long flags;
2118 unsigned long x = 0;
2121 spin_lock_irqsave(&n->list_lock, flags);
2122 list_for_each_entry(page, &n->partial, lru)
2123 x += get_count(page);
2124 spin_unlock_irqrestore(&n->list_lock, flags);
2127 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2129 static noinline void
2130 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2132 #ifdef CONFIG_SLUB_DEBUG
2133 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2134 DEFAULT_RATELIMIT_BURST);
2136 struct kmem_cache_node *n;
2138 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2141 pr_warn("SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2143 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2144 s->name, s->object_size, s->size, oo_order(s->oo),
2147 if (oo_order(s->min) > get_order(s->object_size))
2148 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2151 for_each_kmem_cache_node(s, node, n) {
2152 unsigned long nr_slabs;
2153 unsigned long nr_objs;
2154 unsigned long nr_free;
2156 nr_free = count_partial(n, count_free);
2157 nr_slabs = node_nr_slabs(n);
2158 nr_objs = node_nr_objs(n);
2160 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2161 node, nr_slabs, nr_objs, nr_free);
2166 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2167 int node, struct kmem_cache_cpu **pc)
2170 struct kmem_cache_cpu *c = *pc;
2173 freelist = get_partial(s, flags, node, c);
2178 page = new_slab(s, flags, node);
2180 c = raw_cpu_ptr(s->cpu_slab);
2185 * No other reference to the page yet so we can
2186 * muck around with it freely without cmpxchg
2188 freelist = page->freelist;
2189 page->freelist = NULL;
2191 stat(s, ALLOC_SLAB);
2200 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2202 if (unlikely(PageSlabPfmemalloc(page)))
2203 return gfp_pfmemalloc_allowed(gfpflags);
2209 * Check the page->freelist of a page and either transfer the freelist to the
2210 * per cpu freelist or deactivate the page.
2212 * The page is still frozen if the return value is not NULL.
2214 * If this function returns NULL then the page has been unfrozen.
2216 * This function must be called with interrupt disabled.
2218 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2221 unsigned long counters;
2225 freelist = page->freelist;
2226 counters = page->counters;
2228 new.counters = counters;
2229 VM_BUG_ON(!new.frozen);
2231 new.inuse = page->objects;
2232 new.frozen = freelist != NULL;
2234 } while (!__cmpxchg_double_slab(s, page,
2243 * Slow path. The lockless freelist is empty or we need to perform
2246 * Processing is still very fast if new objects have been freed to the
2247 * regular freelist. In that case we simply take over the regular freelist
2248 * as the lockless freelist and zap the regular freelist.
2250 * If that is not working then we fall back to the partial lists. We take the
2251 * first element of the freelist as the object to allocate now and move the
2252 * rest of the freelist to the lockless freelist.
2254 * And if we were unable to get a new slab from the partial slab lists then
2255 * we need to allocate a new slab. This is the slowest path since it involves
2256 * a call to the page allocator and the setup of a new slab.
2258 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2259 unsigned long addr, struct kmem_cache_cpu *c)
2263 unsigned long flags;
2265 local_irq_save(flags);
2266 #ifdef CONFIG_PREEMPT
2268 * We may have been preempted and rescheduled on a different
2269 * cpu before disabling interrupts. Need to reload cpu area
2272 c = this_cpu_ptr(s->cpu_slab);
2280 if (unlikely(!node_match(page, node))) {
2281 int searchnode = node;
2283 if (node != NUMA_NO_NODE && !node_present_pages(node))
2284 searchnode = node_to_mem_node(node);
2286 if (unlikely(!node_match(page, searchnode))) {
2287 stat(s, ALLOC_NODE_MISMATCH);
2288 deactivate_slab(s, page, c->freelist);
2296 * By rights, we should be searching for a slab page that was
2297 * PFMEMALLOC but right now, we are losing the pfmemalloc
2298 * information when the page leaves the per-cpu allocator
2300 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2301 deactivate_slab(s, page, c->freelist);
2307 /* must check again c->freelist in case of cpu migration or IRQ */
2308 freelist = c->freelist;
2312 freelist = get_freelist(s, page);
2316 stat(s, DEACTIVATE_BYPASS);
2320 stat(s, ALLOC_REFILL);
2324 * freelist is pointing to the list of objects to be used.
2325 * page is pointing to the page from which the objects are obtained.
2326 * That page must be frozen for per cpu allocations to work.
2328 VM_BUG_ON(!c->page->frozen);
2329 c->freelist = get_freepointer(s, freelist);
2330 c->tid = next_tid(c->tid);
2331 local_irq_restore(flags);
2337 page = c->page = c->partial;
2338 c->partial = page->next;
2339 stat(s, CPU_PARTIAL_ALLOC);
2344 freelist = new_slab_objects(s, gfpflags, node, &c);
2346 if (unlikely(!freelist)) {
2347 slab_out_of_memory(s, gfpflags, node);
2348 local_irq_restore(flags);
2353 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2356 /* Only entered in the debug case */
2357 if (kmem_cache_debug(s) &&
2358 !alloc_debug_processing(s, page, freelist, addr))
2359 goto new_slab; /* Slab failed checks. Next slab needed */
2361 deactivate_slab(s, page, get_freepointer(s, freelist));
2364 local_irq_restore(flags);
2369 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2370 * have the fastpath folded into their functions. So no function call
2371 * overhead for requests that can be satisfied on the fastpath.
2373 * The fastpath works by first checking if the lockless freelist can be used.
2374 * If not then __slab_alloc is called for slow processing.
2376 * Otherwise we can simply pick the next object from the lockless free list.
2378 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2379 gfp_t gfpflags, int node, unsigned long addr)
2382 struct kmem_cache_cpu *c;
2386 if (slab_pre_alloc_hook(s, gfpflags))
2389 s = memcg_kmem_get_cache(s, gfpflags);
2392 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2393 * enabled. We may switch back and forth between cpus while
2394 * reading from one cpu area. That does not matter as long
2395 * as we end up on the original cpu again when doing the cmpxchg.
2397 * Preemption is disabled for the retrieval of the tid because that
2398 * must occur from the current processor. We cannot allow rescheduling
2399 * on a different processor between the determination of the pointer
2400 * and the retrieval of the tid.
2403 c = this_cpu_ptr(s->cpu_slab);
2406 * The transaction ids are globally unique per cpu and per operation on
2407 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2408 * occurs on the right processor and that there was no operation on the
2409 * linked list in between.
2414 object = c->freelist;
2416 if (unlikely(!object || !node_match(page, node))) {
2417 object = __slab_alloc(s, gfpflags, node, addr, c);
2418 stat(s, ALLOC_SLOWPATH);
2420 void *next_object = get_freepointer_safe(s, object);
2423 * The cmpxchg will only match if there was no additional
2424 * operation and if we are on the right processor.
2426 * The cmpxchg does the following atomically (without lock
2428 * 1. Relocate first pointer to the current per cpu area.
2429 * 2. Verify that tid and freelist have not been changed
2430 * 3. If they were not changed replace tid and freelist
2432 * Since this is without lock semantics the protection is only
2433 * against code executing on this cpu *not* from access by
2436 if (unlikely(!this_cpu_cmpxchg_double(
2437 s->cpu_slab->freelist, s->cpu_slab->tid,
2439 next_object, next_tid(tid)))) {
2441 note_cmpxchg_failure("slab_alloc", s, tid);
2444 prefetch_freepointer(s, next_object);
2445 stat(s, ALLOC_FASTPATH);
2448 if (unlikely(gfpflags & __GFP_ZERO) && object)
2449 memset(object, 0, s->object_size);
2451 slab_post_alloc_hook(s, gfpflags, object);
2456 static __always_inline void *slab_alloc(struct kmem_cache *s,
2457 gfp_t gfpflags, unsigned long addr)
2459 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2462 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2464 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2466 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2471 EXPORT_SYMBOL(kmem_cache_alloc);
2473 #ifdef CONFIG_TRACING
2474 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2476 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2477 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2480 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2484 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2486 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2488 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2489 s->object_size, s->size, gfpflags, node);
2493 EXPORT_SYMBOL(kmem_cache_alloc_node);
2495 #ifdef CONFIG_TRACING
2496 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2498 int node, size_t size)
2500 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2502 trace_kmalloc_node(_RET_IP_, ret,
2503 size, s->size, gfpflags, node);
2506 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2511 * Slow patch handling. This may still be called frequently since objects
2512 * have a longer lifetime than the cpu slabs in most processing loads.
2514 * So we still attempt to reduce cache line usage. Just take the slab
2515 * lock and free the item. If there is no additional partial page
2516 * handling required then we can return immediately.
2518 static void __slab_free(struct kmem_cache *s, struct page *page,
2519 void *x, unsigned long addr)
2522 void **object = (void *)x;
2525 unsigned long counters;
2526 struct kmem_cache_node *n = NULL;
2527 unsigned long uninitialized_var(flags);
2529 stat(s, FREE_SLOWPATH);
2531 if (kmem_cache_debug(s) &&
2532 !(n = free_debug_processing(s, page, x, addr, &flags)))
2537 spin_unlock_irqrestore(&n->list_lock, flags);
2540 prior = page->freelist;
2541 counters = page->counters;
2542 set_freepointer(s, object, prior);
2543 new.counters = counters;
2544 was_frozen = new.frozen;
2546 if ((!new.inuse || !prior) && !was_frozen) {
2548 if (kmem_cache_has_cpu_partial(s) && !prior) {
2551 * Slab was on no list before and will be
2553 * We can defer the list move and instead
2558 } else { /* Needs to be taken off a list */
2560 n = get_node(s, page_to_nid(page));
2562 * Speculatively acquire the list_lock.
2563 * If the cmpxchg does not succeed then we may
2564 * drop the list_lock without any processing.
2566 * Otherwise the list_lock will synchronize with
2567 * other processors updating the list of slabs.
2569 spin_lock_irqsave(&n->list_lock, flags);
2574 } while (!cmpxchg_double_slab(s, page,
2576 object, new.counters,
2582 * If we just froze the page then put it onto the
2583 * per cpu partial list.
2585 if (new.frozen && !was_frozen) {
2586 put_cpu_partial(s, page, 1);
2587 stat(s, CPU_PARTIAL_FREE);
2590 * The list lock was not taken therefore no list
2591 * activity can be necessary.
2594 stat(s, FREE_FROZEN);
2598 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2602 * Objects left in the slab. If it was not on the partial list before
2605 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2606 if (kmem_cache_debug(s))
2607 remove_full(s, n, page);
2608 add_partial(n, page, DEACTIVATE_TO_TAIL);
2609 stat(s, FREE_ADD_PARTIAL);
2611 spin_unlock_irqrestore(&n->list_lock, flags);
2617 * Slab on the partial list.
2619 remove_partial(n, page);
2620 stat(s, FREE_REMOVE_PARTIAL);
2622 /* Slab must be on the full list */
2623 remove_full(s, n, page);
2626 spin_unlock_irqrestore(&n->list_lock, flags);
2628 discard_slab(s, page);
2632 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2633 * can perform fastpath freeing without additional function calls.
2635 * The fastpath is only possible if we are freeing to the current cpu slab
2636 * of this processor. This typically the case if we have just allocated
2639 * If fastpath is not possible then fall back to __slab_free where we deal
2640 * with all sorts of special processing.
2642 static __always_inline void slab_free(struct kmem_cache *s,
2643 struct page *page, void *x, unsigned long addr)
2645 void **object = (void *)x;
2646 struct kmem_cache_cpu *c;
2649 slab_free_hook(s, x);
2653 * Determine the currently cpus per cpu slab.
2654 * The cpu may change afterward. However that does not matter since
2655 * data is retrieved via this pointer. If we are on the same cpu
2656 * during the cmpxchg then the free will succedd.
2659 c = this_cpu_ptr(s->cpu_slab);
2664 if (likely(page == c->page)) {
2665 set_freepointer(s, object, c->freelist);
2667 if (unlikely(!this_cpu_cmpxchg_double(
2668 s->cpu_slab->freelist, s->cpu_slab->tid,
2670 object, next_tid(tid)))) {
2672 note_cmpxchg_failure("slab_free", s, tid);
2675 stat(s, FREE_FASTPATH);
2677 __slab_free(s, page, x, addr);
2681 void kmem_cache_free(struct kmem_cache *s, void *x)
2683 s = cache_from_obj(s, x);
2686 slab_free(s, virt_to_head_page(x), x, _RET_IP_);
2687 trace_kmem_cache_free(_RET_IP_, x);
2689 EXPORT_SYMBOL(kmem_cache_free);
2692 * Object placement in a slab is made very easy because we always start at
2693 * offset 0. If we tune the size of the object to the alignment then we can
2694 * get the required alignment by putting one properly sized object after
2697 * Notice that the allocation order determines the sizes of the per cpu
2698 * caches. Each processor has always one slab available for allocations.
2699 * Increasing the allocation order reduces the number of times that slabs
2700 * must be moved on and off the partial lists and is therefore a factor in
2705 * Mininum / Maximum order of slab pages. This influences locking overhead
2706 * and slab fragmentation. A higher order reduces the number of partial slabs
2707 * and increases the number of allocations possible without having to
2708 * take the list_lock.
2710 static int slub_min_order;
2711 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2712 static int slub_min_objects;
2715 * Calculate the order of allocation given an slab object size.
2717 * The order of allocation has significant impact on performance and other
2718 * system components. Generally order 0 allocations should be preferred since
2719 * order 0 does not cause fragmentation in the page allocator. Larger objects
2720 * be problematic to put into order 0 slabs because there may be too much
2721 * unused space left. We go to a higher order if more than 1/16th of the slab
2724 * In order to reach satisfactory performance we must ensure that a minimum
2725 * number of objects is in one slab. Otherwise we may generate too much
2726 * activity on the partial lists which requires taking the list_lock. This is
2727 * less a concern for large slabs though which are rarely used.
2729 * slub_max_order specifies the order where we begin to stop considering the
2730 * number of objects in a slab as critical. If we reach slub_max_order then
2731 * we try to keep the page order as low as possible. So we accept more waste
2732 * of space in favor of a small page order.
2734 * Higher order allocations also allow the placement of more objects in a
2735 * slab and thereby reduce object handling overhead. If the user has
2736 * requested a higher mininum order then we start with that one instead of
2737 * the smallest order which will fit the object.
2739 static inline int slab_order(int size, int min_objects,
2740 int max_order, int fract_leftover, int reserved)
2744 int min_order = slub_min_order;
2746 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2747 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2749 for (order = max(min_order,
2750 fls(min_objects * size - 1) - PAGE_SHIFT);
2751 order <= max_order; order++) {
2753 unsigned long slab_size = PAGE_SIZE << order;
2755 if (slab_size < min_objects * size + reserved)
2758 rem = (slab_size - reserved) % size;
2760 if (rem <= slab_size / fract_leftover)
2768 static inline int calculate_order(int size, int reserved)
2776 * Attempt to find best configuration for a slab. This
2777 * works by first attempting to generate a layout with
2778 * the best configuration and backing off gradually.
2780 * First we reduce the acceptable waste in a slab. Then
2781 * we reduce the minimum objects required in a slab.
2783 min_objects = slub_min_objects;
2785 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2786 max_objects = order_objects(slub_max_order, size, reserved);
2787 min_objects = min(min_objects, max_objects);
2789 while (min_objects > 1) {
2791 while (fraction >= 4) {
2792 order = slab_order(size, min_objects,
2793 slub_max_order, fraction, reserved);
2794 if (order <= slub_max_order)
2802 * We were unable to place multiple objects in a slab. Now
2803 * lets see if we can place a single object there.
2805 order = slab_order(size, 1, slub_max_order, 1, reserved);
2806 if (order <= slub_max_order)
2810 * Doh this slab cannot be placed using slub_max_order.
2812 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2813 if (order < MAX_ORDER)
2819 init_kmem_cache_node(struct kmem_cache_node *n)
2822 spin_lock_init(&n->list_lock);
2823 INIT_LIST_HEAD(&n->partial);
2824 #ifdef CONFIG_SLUB_DEBUG
2825 atomic_long_set(&n->nr_slabs, 0);
2826 atomic_long_set(&n->total_objects, 0);
2827 INIT_LIST_HEAD(&n->full);
2831 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2833 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2834 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
2837 * Must align to double word boundary for the double cmpxchg
2838 * instructions to work; see __pcpu_double_call_return_bool().
2840 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2841 2 * sizeof(void *));
2846 init_kmem_cache_cpus(s);
2851 static struct kmem_cache *kmem_cache_node;
2854 * No kmalloc_node yet so do it by hand. We know that this is the first
2855 * slab on the node for this slabcache. There are no concurrent accesses
2858 * Note that this function only works on the kmem_cache_node
2859 * when allocating for the kmem_cache_node. This is used for bootstrapping
2860 * memory on a fresh node that has no slab structures yet.
2862 static void early_kmem_cache_node_alloc(int node)
2865 struct kmem_cache_node *n;
2867 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2869 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2872 if (page_to_nid(page) != node) {
2873 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
2874 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
2879 page->freelist = get_freepointer(kmem_cache_node, n);
2882 kmem_cache_node->node[node] = n;
2883 #ifdef CONFIG_SLUB_DEBUG
2884 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2885 init_tracking(kmem_cache_node, n);
2887 init_kmem_cache_node(n);
2888 inc_slabs_node(kmem_cache_node, node, page->objects);
2891 * No locks need to be taken here as it has just been
2892 * initialized and there is no concurrent access.
2894 __add_partial(n, page, DEACTIVATE_TO_HEAD);
2897 static void free_kmem_cache_nodes(struct kmem_cache *s)
2900 struct kmem_cache_node *n;
2902 for_each_kmem_cache_node(s, node, n) {
2903 kmem_cache_free(kmem_cache_node, n);
2904 s->node[node] = NULL;
2908 static int init_kmem_cache_nodes(struct kmem_cache *s)
2912 for_each_node_state(node, N_NORMAL_MEMORY) {
2913 struct kmem_cache_node *n;
2915 if (slab_state == DOWN) {
2916 early_kmem_cache_node_alloc(node);
2919 n = kmem_cache_alloc_node(kmem_cache_node,
2923 free_kmem_cache_nodes(s);
2928 init_kmem_cache_node(n);
2933 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2935 if (min < MIN_PARTIAL)
2937 else if (min > MAX_PARTIAL)
2939 s->min_partial = min;
2943 * calculate_sizes() determines the order and the distribution of data within
2946 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2948 unsigned long flags = s->flags;
2949 unsigned long size = s->object_size;
2953 * Round up object size to the next word boundary. We can only
2954 * place the free pointer at word boundaries and this determines
2955 * the possible location of the free pointer.
2957 size = ALIGN(size, sizeof(void *));
2959 #ifdef CONFIG_SLUB_DEBUG
2961 * Determine if we can poison the object itself. If the user of
2962 * the slab may touch the object after free or before allocation
2963 * then we should never poison the object itself.
2965 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2967 s->flags |= __OBJECT_POISON;
2969 s->flags &= ~__OBJECT_POISON;
2973 * If we are Redzoning then check if there is some space between the
2974 * end of the object and the free pointer. If not then add an
2975 * additional word to have some bytes to store Redzone information.
2977 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
2978 size += sizeof(void *);
2982 * With that we have determined the number of bytes in actual use
2983 * by the object. This is the potential offset to the free pointer.
2987 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2990 * Relocate free pointer after the object if it is not
2991 * permitted to overwrite the first word of the object on
2994 * This is the case if we do RCU, have a constructor or
2995 * destructor or are poisoning the objects.
2998 size += sizeof(void *);
3001 #ifdef CONFIG_SLUB_DEBUG
3002 if (flags & SLAB_STORE_USER)
3004 * Need to store information about allocs and frees after
3007 size += 2 * sizeof(struct track);
3009 if (flags & SLAB_RED_ZONE)
3011 * Add some empty padding so that we can catch
3012 * overwrites from earlier objects rather than let
3013 * tracking information or the free pointer be
3014 * corrupted if a user writes before the start
3017 size += sizeof(void *);
3021 * SLUB stores one object immediately after another beginning from
3022 * offset 0. In order to align the objects we have to simply size
3023 * each object to conform to the alignment.
3025 size = ALIGN(size, s->align);
3027 if (forced_order >= 0)
3028 order = forced_order;
3030 order = calculate_order(size, s->reserved);
3037 s->allocflags |= __GFP_COMP;
3039 if (s->flags & SLAB_CACHE_DMA)
3040 s->allocflags |= GFP_DMA;
3042 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3043 s->allocflags |= __GFP_RECLAIMABLE;
3046 * Determine the number of objects per slab
3048 s->oo = oo_make(order, size, s->reserved);
3049 s->min = oo_make(get_order(size), size, s->reserved);
3050 if (oo_objects(s->oo) > oo_objects(s->max))
3053 return !!oo_objects(s->oo);
3056 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3058 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3061 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3062 s->reserved = sizeof(struct rcu_head);
3064 if (!calculate_sizes(s, -1))
3066 if (disable_higher_order_debug) {
3068 * Disable debugging flags that store metadata if the min slab
3071 if (get_order(s->size) > get_order(s->object_size)) {
3072 s->flags &= ~DEBUG_METADATA_FLAGS;
3074 if (!calculate_sizes(s, -1))
3079 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3080 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3081 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3082 /* Enable fast mode */
3083 s->flags |= __CMPXCHG_DOUBLE;
3087 * The larger the object size is, the more pages we want on the partial
3088 * list to avoid pounding the page allocator excessively.
3090 set_min_partial(s, ilog2(s->size) / 2);
3093 * cpu_partial determined the maximum number of objects kept in the
3094 * per cpu partial lists of a processor.
3096 * Per cpu partial lists mainly contain slabs that just have one
3097 * object freed. If they are used for allocation then they can be
3098 * filled up again with minimal effort. The slab will never hit the
3099 * per node partial lists and therefore no locking will be required.
3101 * This setting also determines
3103 * A) The number of objects from per cpu partial slabs dumped to the
3104 * per node list when we reach the limit.
3105 * B) The number of objects in cpu partial slabs to extract from the
3106 * per node list when we run out of per cpu objects. We only fetch
3107 * 50% to keep some capacity around for frees.
3109 if (!kmem_cache_has_cpu_partial(s))
3111 else if (s->size >= PAGE_SIZE)
3113 else if (s->size >= 1024)
3115 else if (s->size >= 256)
3116 s->cpu_partial = 13;
3118 s->cpu_partial = 30;
3121 s->remote_node_defrag_ratio = 1000;
3123 if (!init_kmem_cache_nodes(s))
3126 if (alloc_kmem_cache_cpus(s))
3129 free_kmem_cache_nodes(s);
3131 if (flags & SLAB_PANIC)
3132 panic("Cannot create slab %s size=%lu realsize=%u "
3133 "order=%u offset=%u flags=%lx\n",
3134 s->name, (unsigned long)s->size, s->size,
3135 oo_order(s->oo), s->offset, flags);
3139 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3142 #ifdef CONFIG_SLUB_DEBUG
3143 void *addr = page_address(page);
3145 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3146 sizeof(long), GFP_ATOMIC);
3149 slab_err(s, page, text, s->name);
3152 get_map(s, page, map);
3153 for_each_object(p, s, addr, page->objects) {
3155 if (!test_bit(slab_index(p, s, addr), map)) {
3156 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3157 print_tracking(s, p);
3166 * Attempt to free all partial slabs on a node.
3167 * This is called from kmem_cache_close(). We must be the last thread
3168 * using the cache and therefore we do not need to lock anymore.
3170 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3172 struct page *page, *h;
3174 list_for_each_entry_safe(page, h, &n->partial, lru) {
3176 __remove_partial(n, page);
3177 discard_slab(s, page);
3179 list_slab_objects(s, page,
3180 "Objects remaining in %s on kmem_cache_close()");
3186 * Release all resources used by a slab cache.
3188 static inline int kmem_cache_close(struct kmem_cache *s)
3191 struct kmem_cache_node *n;
3194 /* Attempt to free all objects */
3195 for_each_kmem_cache_node(s, node, n) {
3197 if (n->nr_partial || slabs_node(s, node))
3200 free_percpu(s->cpu_slab);
3201 free_kmem_cache_nodes(s);
3205 int __kmem_cache_shutdown(struct kmem_cache *s)
3207 return kmem_cache_close(s);
3210 /********************************************************************
3212 *******************************************************************/
3214 static int __init setup_slub_min_order(char *str)
3216 get_option(&str, &slub_min_order);
3221 __setup("slub_min_order=", setup_slub_min_order);
3223 static int __init setup_slub_max_order(char *str)
3225 get_option(&str, &slub_max_order);
3226 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3231 __setup("slub_max_order=", setup_slub_max_order);
3233 static int __init setup_slub_min_objects(char *str)
3235 get_option(&str, &slub_min_objects);
3240 __setup("slub_min_objects=", setup_slub_min_objects);
3242 void *__kmalloc(size_t size, gfp_t flags)
3244 struct kmem_cache *s;
3247 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3248 return kmalloc_large(size, flags);
3250 s = kmalloc_slab(size, flags);
3252 if (unlikely(ZERO_OR_NULL_PTR(s)))
3255 ret = slab_alloc(s, flags, _RET_IP_);
3257 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3261 EXPORT_SYMBOL(__kmalloc);
3264 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3269 flags |= __GFP_COMP | __GFP_NOTRACK;
3270 page = alloc_kmem_pages_node(node, flags, get_order(size));
3272 ptr = page_address(page);
3274 kmalloc_large_node_hook(ptr, size, flags);
3278 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3280 struct kmem_cache *s;
3283 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3284 ret = kmalloc_large_node(size, flags, node);
3286 trace_kmalloc_node(_RET_IP_, ret,
3287 size, PAGE_SIZE << get_order(size),
3293 s = kmalloc_slab(size, flags);
3295 if (unlikely(ZERO_OR_NULL_PTR(s)))
3298 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3300 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3304 EXPORT_SYMBOL(__kmalloc_node);
3307 size_t ksize(const void *object)
3311 if (unlikely(object == ZERO_SIZE_PTR))
3314 page = virt_to_head_page(object);
3316 if (unlikely(!PageSlab(page))) {
3317 WARN_ON(!PageCompound(page));
3318 return PAGE_SIZE << compound_order(page);
3321 return slab_ksize(page->slab_cache);
3323 EXPORT_SYMBOL(ksize);
3325 void kfree(const void *x)
3328 void *object = (void *)x;
3330 trace_kfree(_RET_IP_, x);
3332 if (unlikely(ZERO_OR_NULL_PTR(x)))
3335 page = virt_to_head_page(x);
3336 if (unlikely(!PageSlab(page))) {
3337 BUG_ON(!PageCompound(page));
3339 __free_kmem_pages(page, compound_order(page));
3342 slab_free(page->slab_cache, page, object, _RET_IP_);
3344 EXPORT_SYMBOL(kfree);
3347 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3348 * the remaining slabs by the number of items in use. The slabs with the
3349 * most items in use come first. New allocations will then fill those up
3350 * and thus they can be removed from the partial lists.
3352 * The slabs with the least items are placed last. This results in them
3353 * being allocated from last increasing the chance that the last objects
3354 * are freed in them.
3356 int __kmem_cache_shrink(struct kmem_cache *s)
3360 struct kmem_cache_node *n;
3363 int objects = oo_objects(s->max);
3364 struct list_head *slabs_by_inuse =
3365 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3366 unsigned long flags;
3368 if (!slabs_by_inuse)
3372 for_each_kmem_cache_node(s, node, n) {
3376 for (i = 0; i < objects; i++)
3377 INIT_LIST_HEAD(slabs_by_inuse + i);
3379 spin_lock_irqsave(&n->list_lock, flags);
3382 * Build lists indexed by the items in use in each slab.
3384 * Note that concurrent frees may occur while we hold the
3385 * list_lock. page->inuse here is the upper limit.
3387 list_for_each_entry_safe(page, t, &n->partial, lru) {
3388 list_move(&page->lru, slabs_by_inuse + page->inuse);
3394 * Rebuild the partial list with the slabs filled up most
3395 * first and the least used slabs at the end.
3397 for (i = objects - 1; i > 0; i--)
3398 list_splice(slabs_by_inuse + i, n->partial.prev);
3400 spin_unlock_irqrestore(&n->list_lock, flags);
3402 /* Release empty slabs */
3403 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3404 discard_slab(s, page);
3407 kfree(slabs_by_inuse);
3411 static int slab_mem_going_offline_callback(void *arg)
3413 struct kmem_cache *s;
3415 mutex_lock(&slab_mutex);
3416 list_for_each_entry(s, &slab_caches, list)
3417 __kmem_cache_shrink(s);
3418 mutex_unlock(&slab_mutex);
3423 static void slab_mem_offline_callback(void *arg)
3425 struct kmem_cache_node *n;
3426 struct kmem_cache *s;
3427 struct memory_notify *marg = arg;
3430 offline_node = marg->status_change_nid_normal;
3433 * If the node still has available memory. we need kmem_cache_node
3436 if (offline_node < 0)
3439 mutex_lock(&slab_mutex);
3440 list_for_each_entry(s, &slab_caches, list) {
3441 n = get_node(s, offline_node);
3444 * if n->nr_slabs > 0, slabs still exist on the node
3445 * that is going down. We were unable to free them,
3446 * and offline_pages() function shouldn't call this
3447 * callback. So, we must fail.
3449 BUG_ON(slabs_node(s, offline_node));
3451 s->node[offline_node] = NULL;
3452 kmem_cache_free(kmem_cache_node, n);
3455 mutex_unlock(&slab_mutex);
3458 static int slab_mem_going_online_callback(void *arg)
3460 struct kmem_cache_node *n;
3461 struct kmem_cache *s;
3462 struct memory_notify *marg = arg;
3463 int nid = marg->status_change_nid_normal;
3467 * If the node's memory is already available, then kmem_cache_node is
3468 * already created. Nothing to do.
3474 * We are bringing a node online. No memory is available yet. We must
3475 * allocate a kmem_cache_node structure in order to bring the node
3478 mutex_lock(&slab_mutex);
3479 list_for_each_entry(s, &slab_caches, list) {
3481 * XXX: kmem_cache_alloc_node will fallback to other nodes
3482 * since memory is not yet available from the node that
3485 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3490 init_kmem_cache_node(n);
3494 mutex_unlock(&slab_mutex);
3498 static int slab_memory_callback(struct notifier_block *self,
3499 unsigned long action, void *arg)
3504 case MEM_GOING_ONLINE:
3505 ret = slab_mem_going_online_callback(arg);
3507 case MEM_GOING_OFFLINE:
3508 ret = slab_mem_going_offline_callback(arg);
3511 case MEM_CANCEL_ONLINE:
3512 slab_mem_offline_callback(arg);
3515 case MEM_CANCEL_OFFLINE:
3519 ret = notifier_from_errno(ret);
3525 static struct notifier_block slab_memory_callback_nb = {
3526 .notifier_call = slab_memory_callback,
3527 .priority = SLAB_CALLBACK_PRI,
3530 /********************************************************************
3531 * Basic setup of slabs
3532 *******************************************************************/
3535 * Used for early kmem_cache structures that were allocated using
3536 * the page allocator. Allocate them properly then fix up the pointers
3537 * that may be pointing to the wrong kmem_cache structure.
3540 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3543 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3544 struct kmem_cache_node *n;
3546 memcpy(s, static_cache, kmem_cache->object_size);
3549 * This runs very early, and only the boot processor is supposed to be
3550 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3553 __flush_cpu_slab(s, smp_processor_id());
3554 for_each_kmem_cache_node(s, node, n) {
3557 list_for_each_entry(p, &n->partial, lru)
3560 #ifdef CONFIG_SLUB_DEBUG
3561 list_for_each_entry(p, &n->full, lru)
3565 list_add(&s->list, &slab_caches);
3569 void __init kmem_cache_init(void)
3571 static __initdata struct kmem_cache boot_kmem_cache,
3572 boot_kmem_cache_node;
3574 if (debug_guardpage_minorder())
3577 kmem_cache_node = &boot_kmem_cache_node;
3578 kmem_cache = &boot_kmem_cache;
3580 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3581 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3583 register_hotmemory_notifier(&slab_memory_callback_nb);
3585 /* Able to allocate the per node structures */
3586 slab_state = PARTIAL;
3588 create_boot_cache(kmem_cache, "kmem_cache",
3589 offsetof(struct kmem_cache, node) +
3590 nr_node_ids * sizeof(struct kmem_cache_node *),
3591 SLAB_HWCACHE_ALIGN);
3593 kmem_cache = bootstrap(&boot_kmem_cache);
3596 * Allocate kmem_cache_node properly from the kmem_cache slab.
3597 * kmem_cache_node is separately allocated so no need to
3598 * update any list pointers.
3600 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3602 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3603 create_kmalloc_caches(0);
3606 register_cpu_notifier(&slab_notifier);
3609 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
3611 slub_min_order, slub_max_order, slub_min_objects,
3612 nr_cpu_ids, nr_node_ids);
3615 void __init kmem_cache_init_late(void)
3620 __kmem_cache_alias(const char *name, size_t size, size_t align,
3621 unsigned long flags, void (*ctor)(void *))
3623 struct kmem_cache *s;
3625 s = find_mergeable(size, align, flags, name, ctor);
3628 struct kmem_cache *c;
3633 * Adjust the object sizes so that we clear
3634 * the complete object on kzalloc.
3636 s->object_size = max(s->object_size, (int)size);
3637 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3639 for_each_memcg_cache_index(i) {
3640 c = cache_from_memcg_idx(s, i);
3643 c->object_size = s->object_size;
3644 c->inuse = max_t(int, c->inuse,
3645 ALIGN(size, sizeof(void *)));
3648 if (sysfs_slab_alias(s, name)) {
3657 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3661 err = kmem_cache_open(s, flags);
3665 /* Mutex is not taken during early boot */
3666 if (slab_state <= UP)
3669 memcg_propagate_slab_attrs(s);
3670 err = sysfs_slab_add(s);
3672 kmem_cache_close(s);
3679 * Use the cpu notifier to insure that the cpu slabs are flushed when
3682 static int slab_cpuup_callback(struct notifier_block *nfb,
3683 unsigned long action, void *hcpu)
3685 long cpu = (long)hcpu;
3686 struct kmem_cache *s;
3687 unsigned long flags;
3690 case CPU_UP_CANCELED:
3691 case CPU_UP_CANCELED_FROZEN:
3693 case CPU_DEAD_FROZEN:
3694 mutex_lock(&slab_mutex);
3695 list_for_each_entry(s, &slab_caches, list) {
3696 local_irq_save(flags);
3697 __flush_cpu_slab(s, cpu);
3698 local_irq_restore(flags);
3700 mutex_unlock(&slab_mutex);
3708 static struct notifier_block slab_notifier = {
3709 .notifier_call = slab_cpuup_callback
3714 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3716 struct kmem_cache *s;
3719 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3720 return kmalloc_large(size, gfpflags);
3722 s = kmalloc_slab(size, gfpflags);
3724 if (unlikely(ZERO_OR_NULL_PTR(s)))
3727 ret = slab_alloc(s, gfpflags, caller);
3729 /* Honor the call site pointer we received. */
3730 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3736 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3737 int node, unsigned long caller)
3739 struct kmem_cache *s;
3742 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3743 ret = kmalloc_large_node(size, gfpflags, node);
3745 trace_kmalloc_node(caller, ret,
3746 size, PAGE_SIZE << get_order(size),
3752 s = kmalloc_slab(size, gfpflags);
3754 if (unlikely(ZERO_OR_NULL_PTR(s)))
3757 ret = slab_alloc_node(s, gfpflags, node, caller);
3759 /* Honor the call site pointer we received. */
3760 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3767 static int count_inuse(struct page *page)
3772 static int count_total(struct page *page)
3774 return page->objects;
3778 #ifdef CONFIG_SLUB_DEBUG
3779 static int validate_slab(struct kmem_cache *s, struct page *page,
3783 void *addr = page_address(page);
3785 if (!check_slab(s, page) ||
3786 !on_freelist(s, page, NULL))
3789 /* Now we know that a valid freelist exists */
3790 bitmap_zero(map, page->objects);
3792 get_map(s, page, map);
3793 for_each_object(p, s, addr, page->objects) {
3794 if (test_bit(slab_index(p, s, addr), map))
3795 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3799 for_each_object(p, s, addr, page->objects)
3800 if (!test_bit(slab_index(p, s, addr), map))
3801 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3806 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3810 validate_slab(s, page, map);
3814 static int validate_slab_node(struct kmem_cache *s,
3815 struct kmem_cache_node *n, unsigned long *map)
3817 unsigned long count = 0;
3819 unsigned long flags;
3821 spin_lock_irqsave(&n->list_lock, flags);
3823 list_for_each_entry(page, &n->partial, lru) {
3824 validate_slab_slab(s, page, map);
3827 if (count != n->nr_partial)
3828 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
3829 s->name, count, n->nr_partial);
3831 if (!(s->flags & SLAB_STORE_USER))
3834 list_for_each_entry(page, &n->full, lru) {
3835 validate_slab_slab(s, page, map);
3838 if (count != atomic_long_read(&n->nr_slabs))
3839 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
3840 s->name, count, atomic_long_read(&n->nr_slabs));
3843 spin_unlock_irqrestore(&n->list_lock, flags);
3847 static long validate_slab_cache(struct kmem_cache *s)
3850 unsigned long count = 0;
3851 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3852 sizeof(unsigned long), GFP_KERNEL);
3853 struct kmem_cache_node *n;
3859 for_each_kmem_cache_node(s, node, n)
3860 count += validate_slab_node(s, n, map);
3865 * Generate lists of code addresses where slabcache objects are allocated
3870 unsigned long count;
3877 DECLARE_BITMAP(cpus, NR_CPUS);
3883 unsigned long count;
3884 struct location *loc;
3887 static void free_loc_track(struct loc_track *t)
3890 free_pages((unsigned long)t->loc,
3891 get_order(sizeof(struct location) * t->max));
3894 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3899 order = get_order(sizeof(struct location) * max);
3901 l = (void *)__get_free_pages(flags, order);
3906 memcpy(l, t->loc, sizeof(struct location) * t->count);
3914 static int add_location(struct loc_track *t, struct kmem_cache *s,
3915 const struct track *track)
3917 long start, end, pos;
3919 unsigned long caddr;
3920 unsigned long age = jiffies - track->when;
3926 pos = start + (end - start + 1) / 2;
3929 * There is nothing at "end". If we end up there
3930 * we need to add something to before end.
3935 caddr = t->loc[pos].addr;
3936 if (track->addr == caddr) {
3942 if (age < l->min_time)
3944 if (age > l->max_time)
3947 if (track->pid < l->min_pid)
3948 l->min_pid = track->pid;
3949 if (track->pid > l->max_pid)
3950 l->max_pid = track->pid;
3952 cpumask_set_cpu(track->cpu,
3953 to_cpumask(l->cpus));
3955 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3959 if (track->addr < caddr)
3966 * Not found. Insert new tracking element.
3968 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3974 (t->count - pos) * sizeof(struct location));
3977 l->addr = track->addr;
3981 l->min_pid = track->pid;
3982 l->max_pid = track->pid;
3983 cpumask_clear(to_cpumask(l->cpus));
3984 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
3985 nodes_clear(l->nodes);
3986 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3990 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3991 struct page *page, enum track_item alloc,
3994 void *addr = page_address(page);
3997 bitmap_zero(map, page->objects);
3998 get_map(s, page, map);
4000 for_each_object(p, s, addr, page->objects)
4001 if (!test_bit(slab_index(p, s, addr), map))
4002 add_location(t, s, get_track(s, p, alloc));
4005 static int list_locations(struct kmem_cache *s, char *buf,
4006 enum track_item alloc)
4010 struct loc_track t = { 0, 0, NULL };
4012 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4013 sizeof(unsigned long), GFP_KERNEL);
4014 struct kmem_cache_node *n;
4016 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4019 return sprintf(buf, "Out of memory\n");
4021 /* Push back cpu slabs */
4024 for_each_kmem_cache_node(s, node, n) {
4025 unsigned long flags;
4028 if (!atomic_long_read(&n->nr_slabs))
4031 spin_lock_irqsave(&n->list_lock, flags);
4032 list_for_each_entry(page, &n->partial, lru)
4033 process_slab(&t, s, page, alloc, map);
4034 list_for_each_entry(page, &n->full, lru)
4035 process_slab(&t, s, page, alloc, map);
4036 spin_unlock_irqrestore(&n->list_lock, flags);
4039 for (i = 0; i < t.count; i++) {
4040 struct location *l = &t.loc[i];
4042 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4044 len += sprintf(buf + len, "%7ld ", l->count);
4047 len += sprintf(buf + len, "%pS", (void *)l->addr);
4049 len += sprintf(buf + len, "<not-available>");
4051 if (l->sum_time != l->min_time) {
4052 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4054 (long)div_u64(l->sum_time, l->count),
4057 len += sprintf(buf + len, " age=%ld",
4060 if (l->min_pid != l->max_pid)
4061 len += sprintf(buf + len, " pid=%ld-%ld",
4062 l->min_pid, l->max_pid);
4064 len += sprintf(buf + len, " pid=%ld",
4067 if (num_online_cpus() > 1 &&
4068 !cpumask_empty(to_cpumask(l->cpus)) &&
4069 len < PAGE_SIZE - 60) {
4070 len += sprintf(buf + len, " cpus=");
4071 len += cpulist_scnprintf(buf + len,
4072 PAGE_SIZE - len - 50,
4073 to_cpumask(l->cpus));
4076 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4077 len < PAGE_SIZE - 60) {
4078 len += sprintf(buf + len, " nodes=");
4079 len += nodelist_scnprintf(buf + len,
4080 PAGE_SIZE - len - 50,
4084 len += sprintf(buf + len, "\n");
4090 len += sprintf(buf, "No data\n");
4095 #ifdef SLUB_RESILIENCY_TEST
4096 static void __init resiliency_test(void)
4100 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4102 pr_err("SLUB resiliency testing\n");
4103 pr_err("-----------------------\n");
4104 pr_err("A. Corruption after allocation\n");
4106 p = kzalloc(16, GFP_KERNEL);
4108 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4111 validate_slab_cache(kmalloc_caches[4]);
4113 /* Hmmm... The next two are dangerous */
4114 p = kzalloc(32, GFP_KERNEL);
4115 p[32 + sizeof(void *)] = 0x34;
4116 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4118 pr_err("If allocated object is overwritten then not detectable\n\n");
4120 validate_slab_cache(kmalloc_caches[5]);
4121 p = kzalloc(64, GFP_KERNEL);
4122 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4124 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4126 pr_err("If allocated object is overwritten then not detectable\n\n");
4127 validate_slab_cache(kmalloc_caches[6]);
4129 pr_err("\nB. Corruption after free\n");
4130 p = kzalloc(128, GFP_KERNEL);
4133 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4134 validate_slab_cache(kmalloc_caches[7]);
4136 p = kzalloc(256, GFP_KERNEL);
4139 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4140 validate_slab_cache(kmalloc_caches[8]);
4142 p = kzalloc(512, GFP_KERNEL);
4145 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4146 validate_slab_cache(kmalloc_caches[9]);
4150 static void resiliency_test(void) {};
4155 enum slab_stat_type {
4156 SL_ALL, /* All slabs */
4157 SL_PARTIAL, /* Only partially allocated slabs */
4158 SL_CPU, /* Only slabs used for cpu caches */
4159 SL_OBJECTS, /* Determine allocated objects not slabs */
4160 SL_TOTAL /* Determine object capacity not slabs */
4163 #define SO_ALL (1 << SL_ALL)
4164 #define SO_PARTIAL (1 << SL_PARTIAL)
4165 #define SO_CPU (1 << SL_CPU)
4166 #define SO_OBJECTS (1 << SL_OBJECTS)
4167 #define SO_TOTAL (1 << SL_TOTAL)
4169 static ssize_t show_slab_objects(struct kmem_cache *s,
4170 char *buf, unsigned long flags)
4172 unsigned long total = 0;
4175 unsigned long *nodes;
4177 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4181 if (flags & SO_CPU) {
4184 for_each_possible_cpu(cpu) {
4185 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4190 page = ACCESS_ONCE(c->page);
4194 node = page_to_nid(page);
4195 if (flags & SO_TOTAL)
4197 else if (flags & SO_OBJECTS)
4205 page = ACCESS_ONCE(c->partial);
4207 node = page_to_nid(page);
4208 if (flags & SO_TOTAL)
4210 else if (flags & SO_OBJECTS)
4221 #ifdef CONFIG_SLUB_DEBUG
4222 if (flags & SO_ALL) {
4223 struct kmem_cache_node *n;
4225 for_each_kmem_cache_node(s, node, n) {
4227 if (flags & SO_TOTAL)
4228 x = atomic_long_read(&n->total_objects);
4229 else if (flags & SO_OBJECTS)
4230 x = atomic_long_read(&n->total_objects) -
4231 count_partial(n, count_free);
4233 x = atomic_long_read(&n->nr_slabs);
4240 if (flags & SO_PARTIAL) {
4241 struct kmem_cache_node *n;
4243 for_each_kmem_cache_node(s, node, n) {
4244 if (flags & SO_TOTAL)
4245 x = count_partial(n, count_total);
4246 else if (flags & SO_OBJECTS)
4247 x = count_partial(n, count_inuse);
4254 x = sprintf(buf, "%lu", total);
4256 for (node = 0; node < nr_node_ids; node++)
4258 x += sprintf(buf + x, " N%d=%lu",
4263 return x + sprintf(buf + x, "\n");
4266 #ifdef CONFIG_SLUB_DEBUG
4267 static int any_slab_objects(struct kmem_cache *s)
4270 struct kmem_cache_node *n;
4272 for_each_kmem_cache_node(s, node, n)
4273 if (atomic_long_read(&n->total_objects))
4280 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4281 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4283 struct slab_attribute {
4284 struct attribute attr;
4285 ssize_t (*show)(struct kmem_cache *s, char *buf);
4286 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4289 #define SLAB_ATTR_RO(_name) \
4290 static struct slab_attribute _name##_attr = \
4291 __ATTR(_name, 0400, _name##_show, NULL)
4293 #define SLAB_ATTR(_name) \
4294 static struct slab_attribute _name##_attr = \
4295 __ATTR(_name, 0600, _name##_show, _name##_store)
4297 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4299 return sprintf(buf, "%d\n", s->size);
4301 SLAB_ATTR_RO(slab_size);
4303 static ssize_t align_show(struct kmem_cache *s, char *buf)
4305 return sprintf(buf, "%d\n", s->align);
4307 SLAB_ATTR_RO(align);
4309 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4311 return sprintf(buf, "%d\n", s->object_size);
4313 SLAB_ATTR_RO(object_size);
4315 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4317 return sprintf(buf, "%d\n", oo_objects(s->oo));
4319 SLAB_ATTR_RO(objs_per_slab);
4321 static ssize_t order_store(struct kmem_cache *s,
4322 const char *buf, size_t length)
4324 unsigned long order;
4327 err = kstrtoul(buf, 10, &order);
4331 if (order > slub_max_order || order < slub_min_order)
4334 calculate_sizes(s, order);
4338 static ssize_t order_show(struct kmem_cache *s, char *buf)
4340 return sprintf(buf, "%d\n", oo_order(s->oo));
4344 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4346 return sprintf(buf, "%lu\n", s->min_partial);
4349 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4355 err = kstrtoul(buf, 10, &min);
4359 set_min_partial(s, min);
4362 SLAB_ATTR(min_partial);
4364 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4366 return sprintf(buf, "%u\n", s->cpu_partial);
4369 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4372 unsigned long objects;
4375 err = kstrtoul(buf, 10, &objects);
4378 if (objects && !kmem_cache_has_cpu_partial(s))
4381 s->cpu_partial = objects;
4385 SLAB_ATTR(cpu_partial);
4387 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4391 return sprintf(buf, "%pS\n", s->ctor);
4395 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4397 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
4399 SLAB_ATTR_RO(aliases);
4401 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4403 return show_slab_objects(s, buf, SO_PARTIAL);
4405 SLAB_ATTR_RO(partial);
4407 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4409 return show_slab_objects(s, buf, SO_CPU);
4411 SLAB_ATTR_RO(cpu_slabs);
4413 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4415 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4417 SLAB_ATTR_RO(objects);
4419 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4421 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4423 SLAB_ATTR_RO(objects_partial);
4425 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4432 for_each_online_cpu(cpu) {
4433 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4436 pages += page->pages;
4437 objects += page->pobjects;
4441 len = sprintf(buf, "%d(%d)", objects, pages);
4444 for_each_online_cpu(cpu) {
4445 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4447 if (page && len < PAGE_SIZE - 20)
4448 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4449 page->pobjects, page->pages);
4452 return len + sprintf(buf + len, "\n");
4454 SLAB_ATTR_RO(slabs_cpu_partial);
4456 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4458 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4461 static ssize_t reclaim_account_store(struct kmem_cache *s,
4462 const char *buf, size_t length)
4464 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4466 s->flags |= SLAB_RECLAIM_ACCOUNT;
4469 SLAB_ATTR(reclaim_account);
4471 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4473 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4475 SLAB_ATTR_RO(hwcache_align);
4477 #ifdef CONFIG_ZONE_DMA
4478 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4480 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4482 SLAB_ATTR_RO(cache_dma);
4485 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4487 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4489 SLAB_ATTR_RO(destroy_by_rcu);
4491 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4493 return sprintf(buf, "%d\n", s->reserved);
4495 SLAB_ATTR_RO(reserved);
4497 #ifdef CONFIG_SLUB_DEBUG
4498 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4500 return show_slab_objects(s, buf, SO_ALL);
4502 SLAB_ATTR_RO(slabs);
4504 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4506 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4508 SLAB_ATTR_RO(total_objects);
4510 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4512 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4515 static ssize_t sanity_checks_store(struct kmem_cache *s,
4516 const char *buf, size_t length)
4518 s->flags &= ~SLAB_DEBUG_FREE;
4519 if (buf[0] == '1') {
4520 s->flags &= ~__CMPXCHG_DOUBLE;
4521 s->flags |= SLAB_DEBUG_FREE;
4525 SLAB_ATTR(sanity_checks);
4527 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4529 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4532 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4536 * Tracing a merged cache is going to give confusing results
4537 * as well as cause other issues like converting a mergeable
4538 * cache into an umergeable one.
4540 if (s->refcount > 1)
4543 s->flags &= ~SLAB_TRACE;
4544 if (buf[0] == '1') {
4545 s->flags &= ~__CMPXCHG_DOUBLE;
4546 s->flags |= SLAB_TRACE;
4552 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4554 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4557 static ssize_t red_zone_store(struct kmem_cache *s,
4558 const char *buf, size_t length)
4560 if (any_slab_objects(s))
4563 s->flags &= ~SLAB_RED_ZONE;
4564 if (buf[0] == '1') {
4565 s->flags &= ~__CMPXCHG_DOUBLE;
4566 s->flags |= SLAB_RED_ZONE;
4568 calculate_sizes(s, -1);
4571 SLAB_ATTR(red_zone);
4573 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4575 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4578 static ssize_t poison_store(struct kmem_cache *s,
4579 const char *buf, size_t length)
4581 if (any_slab_objects(s))
4584 s->flags &= ~SLAB_POISON;
4585 if (buf[0] == '1') {
4586 s->flags &= ~__CMPXCHG_DOUBLE;
4587 s->flags |= SLAB_POISON;
4589 calculate_sizes(s, -1);
4594 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4596 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4599 static ssize_t store_user_store(struct kmem_cache *s,
4600 const char *buf, size_t length)
4602 if (any_slab_objects(s))
4605 s->flags &= ~SLAB_STORE_USER;
4606 if (buf[0] == '1') {
4607 s->flags &= ~__CMPXCHG_DOUBLE;
4608 s->flags |= SLAB_STORE_USER;
4610 calculate_sizes(s, -1);
4613 SLAB_ATTR(store_user);
4615 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4620 static ssize_t validate_store(struct kmem_cache *s,
4621 const char *buf, size_t length)
4625 if (buf[0] == '1') {
4626 ret = validate_slab_cache(s);
4632 SLAB_ATTR(validate);
4634 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4636 if (!(s->flags & SLAB_STORE_USER))
4638 return list_locations(s, buf, TRACK_ALLOC);
4640 SLAB_ATTR_RO(alloc_calls);
4642 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4644 if (!(s->flags & SLAB_STORE_USER))
4646 return list_locations(s, buf, TRACK_FREE);
4648 SLAB_ATTR_RO(free_calls);
4649 #endif /* CONFIG_SLUB_DEBUG */
4651 #ifdef CONFIG_FAILSLAB
4652 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4654 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4657 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4660 if (s->refcount > 1)
4663 s->flags &= ~SLAB_FAILSLAB;
4665 s->flags |= SLAB_FAILSLAB;
4668 SLAB_ATTR(failslab);
4671 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4676 static ssize_t shrink_store(struct kmem_cache *s,
4677 const char *buf, size_t length)
4679 if (buf[0] == '1') {
4680 int rc = kmem_cache_shrink(s);
4691 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4693 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4696 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4697 const char *buf, size_t length)
4699 unsigned long ratio;
4702 err = kstrtoul(buf, 10, &ratio);
4707 s->remote_node_defrag_ratio = ratio * 10;
4711 SLAB_ATTR(remote_node_defrag_ratio);
4714 #ifdef CONFIG_SLUB_STATS
4715 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4717 unsigned long sum = 0;
4720 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4725 for_each_online_cpu(cpu) {
4726 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4732 len = sprintf(buf, "%lu", sum);
4735 for_each_online_cpu(cpu) {
4736 if (data[cpu] && len < PAGE_SIZE - 20)
4737 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4741 return len + sprintf(buf + len, "\n");
4744 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4748 for_each_online_cpu(cpu)
4749 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4752 #define STAT_ATTR(si, text) \
4753 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4755 return show_stat(s, buf, si); \
4757 static ssize_t text##_store(struct kmem_cache *s, \
4758 const char *buf, size_t length) \
4760 if (buf[0] != '0') \
4762 clear_stat(s, si); \
4767 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4768 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4769 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4770 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4771 STAT_ATTR(FREE_FROZEN, free_frozen);
4772 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4773 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4774 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4775 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4776 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4777 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
4778 STAT_ATTR(FREE_SLAB, free_slab);
4779 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4780 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4781 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4782 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4783 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4784 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4785 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
4786 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4787 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4788 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4789 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
4790 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
4791 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
4792 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
4795 static struct attribute *slab_attrs[] = {
4796 &slab_size_attr.attr,
4797 &object_size_attr.attr,
4798 &objs_per_slab_attr.attr,
4800 &min_partial_attr.attr,
4801 &cpu_partial_attr.attr,
4803 &objects_partial_attr.attr,
4805 &cpu_slabs_attr.attr,
4809 &hwcache_align_attr.attr,
4810 &reclaim_account_attr.attr,
4811 &destroy_by_rcu_attr.attr,
4813 &reserved_attr.attr,
4814 &slabs_cpu_partial_attr.attr,
4815 #ifdef CONFIG_SLUB_DEBUG
4816 &total_objects_attr.attr,
4818 &sanity_checks_attr.attr,
4820 &red_zone_attr.attr,
4822 &store_user_attr.attr,
4823 &validate_attr.attr,
4824 &alloc_calls_attr.attr,
4825 &free_calls_attr.attr,
4827 #ifdef CONFIG_ZONE_DMA
4828 &cache_dma_attr.attr,
4831 &remote_node_defrag_ratio_attr.attr,
4833 #ifdef CONFIG_SLUB_STATS
4834 &alloc_fastpath_attr.attr,
4835 &alloc_slowpath_attr.attr,
4836 &free_fastpath_attr.attr,
4837 &free_slowpath_attr.attr,
4838 &free_frozen_attr.attr,
4839 &free_add_partial_attr.attr,
4840 &free_remove_partial_attr.attr,
4841 &alloc_from_partial_attr.attr,
4842 &alloc_slab_attr.attr,
4843 &alloc_refill_attr.attr,
4844 &alloc_node_mismatch_attr.attr,
4845 &free_slab_attr.attr,
4846 &cpuslab_flush_attr.attr,
4847 &deactivate_full_attr.attr,
4848 &deactivate_empty_attr.attr,
4849 &deactivate_to_head_attr.attr,
4850 &deactivate_to_tail_attr.attr,
4851 &deactivate_remote_frees_attr.attr,
4852 &deactivate_bypass_attr.attr,
4853 &order_fallback_attr.attr,
4854 &cmpxchg_double_fail_attr.attr,
4855 &cmpxchg_double_cpu_fail_attr.attr,
4856 &cpu_partial_alloc_attr.attr,
4857 &cpu_partial_free_attr.attr,
4858 &cpu_partial_node_attr.attr,
4859 &cpu_partial_drain_attr.attr,
4861 #ifdef CONFIG_FAILSLAB
4862 &failslab_attr.attr,
4868 static struct attribute_group slab_attr_group = {
4869 .attrs = slab_attrs,
4872 static ssize_t slab_attr_show(struct kobject *kobj,
4873 struct attribute *attr,
4876 struct slab_attribute *attribute;
4877 struct kmem_cache *s;
4880 attribute = to_slab_attr(attr);
4883 if (!attribute->show)
4886 err = attribute->show(s, buf);
4891 static ssize_t slab_attr_store(struct kobject *kobj,
4892 struct attribute *attr,
4893 const char *buf, size_t len)
4895 struct slab_attribute *attribute;
4896 struct kmem_cache *s;
4899 attribute = to_slab_attr(attr);
4902 if (!attribute->store)
4905 err = attribute->store(s, buf, len);
4906 #ifdef CONFIG_MEMCG_KMEM
4907 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
4910 mutex_lock(&slab_mutex);
4911 if (s->max_attr_size < len)
4912 s->max_attr_size = len;
4915 * This is a best effort propagation, so this function's return
4916 * value will be determined by the parent cache only. This is
4917 * basically because not all attributes will have a well
4918 * defined semantics for rollbacks - most of the actions will
4919 * have permanent effects.
4921 * Returning the error value of any of the children that fail
4922 * is not 100 % defined, in the sense that users seeing the
4923 * error code won't be able to know anything about the state of
4926 * Only returning the error code for the parent cache at least
4927 * has well defined semantics. The cache being written to
4928 * directly either failed or succeeded, in which case we loop
4929 * through the descendants with best-effort propagation.
4931 for_each_memcg_cache_index(i) {
4932 struct kmem_cache *c = cache_from_memcg_idx(s, i);
4934 attribute->store(c, buf, len);
4936 mutex_unlock(&slab_mutex);
4942 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
4944 #ifdef CONFIG_MEMCG_KMEM
4946 char *buffer = NULL;
4947 struct kmem_cache *root_cache;
4949 if (is_root_cache(s))
4952 root_cache = s->memcg_params->root_cache;
4955 * This mean this cache had no attribute written. Therefore, no point
4956 * in copying default values around
4958 if (!root_cache->max_attr_size)
4961 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
4964 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
4966 if (!attr || !attr->store || !attr->show)
4970 * It is really bad that we have to allocate here, so we will
4971 * do it only as a fallback. If we actually allocate, though,
4972 * we can just use the allocated buffer until the end.
4974 * Most of the slub attributes will tend to be very small in
4975 * size, but sysfs allows buffers up to a page, so they can
4976 * theoretically happen.
4980 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
4983 buffer = (char *) get_zeroed_page(GFP_KERNEL);
4984 if (WARN_ON(!buffer))
4989 attr->show(root_cache, buf);
4990 attr->store(s, buf, strlen(buf));
4994 free_page((unsigned long)buffer);
4998 static void kmem_cache_release(struct kobject *k)
5000 slab_kmem_cache_release(to_slab(k));
5003 static const struct sysfs_ops slab_sysfs_ops = {
5004 .show = slab_attr_show,
5005 .store = slab_attr_store,
5008 static struct kobj_type slab_ktype = {
5009 .sysfs_ops = &slab_sysfs_ops,
5010 .release = kmem_cache_release,
5013 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5015 struct kobj_type *ktype = get_ktype(kobj);
5017 if (ktype == &slab_ktype)
5022 static const struct kset_uevent_ops slab_uevent_ops = {
5023 .filter = uevent_filter,
5026 static struct kset *slab_kset;
5028 static inline struct kset *cache_kset(struct kmem_cache *s)
5030 #ifdef CONFIG_MEMCG_KMEM
5031 if (!is_root_cache(s))
5032 return s->memcg_params->root_cache->memcg_kset;
5037 #define ID_STR_LENGTH 64
5039 /* Create a unique string id for a slab cache:
5041 * Format :[flags-]size
5043 static char *create_unique_id(struct kmem_cache *s)
5045 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5052 * First flags affecting slabcache operations. We will only
5053 * get here for aliasable slabs so we do not need to support
5054 * too many flags. The flags here must cover all flags that
5055 * are matched during merging to guarantee that the id is
5058 if (s->flags & SLAB_CACHE_DMA)
5060 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5062 if (s->flags & SLAB_DEBUG_FREE)
5064 if (!(s->flags & SLAB_NOTRACK))
5068 p += sprintf(p, "%07d", s->size);
5070 BUG_ON(p > name + ID_STR_LENGTH - 1);
5074 static int sysfs_slab_add(struct kmem_cache *s)
5078 int unmergeable = slab_unmergeable(s);
5082 * Slabcache can never be merged so we can use the name proper.
5083 * This is typically the case for debug situations. In that
5084 * case we can catch duplicate names easily.
5086 sysfs_remove_link(&slab_kset->kobj, s->name);
5090 * Create a unique name for the slab as a target
5093 name = create_unique_id(s);
5096 s->kobj.kset = cache_kset(s);
5097 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5101 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5105 #ifdef CONFIG_MEMCG_KMEM
5106 if (is_root_cache(s)) {
5107 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5108 if (!s->memcg_kset) {
5115 kobject_uevent(&s->kobj, KOBJ_ADD);
5117 /* Setup first alias */
5118 sysfs_slab_alias(s, s->name);
5125 kobject_del(&s->kobj);
5127 kobject_put(&s->kobj);
5131 void sysfs_slab_remove(struct kmem_cache *s)
5133 if (slab_state < FULL)
5135 * Sysfs has not been setup yet so no need to remove the
5140 #ifdef CONFIG_MEMCG_KMEM
5141 kset_unregister(s->memcg_kset);
5143 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5144 kobject_del(&s->kobj);
5145 kobject_put(&s->kobj);
5149 * Need to buffer aliases during bootup until sysfs becomes
5150 * available lest we lose that information.
5152 struct saved_alias {
5153 struct kmem_cache *s;
5155 struct saved_alias *next;
5158 static struct saved_alias *alias_list;
5160 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5162 struct saved_alias *al;
5164 if (slab_state == FULL) {
5166 * If we have a leftover link then remove it.
5168 sysfs_remove_link(&slab_kset->kobj, name);
5169 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5172 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5178 al->next = alias_list;
5183 static int __init slab_sysfs_init(void)
5185 struct kmem_cache *s;
5188 mutex_lock(&slab_mutex);
5190 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5192 mutex_unlock(&slab_mutex);
5193 pr_err("Cannot register slab subsystem.\n");
5199 list_for_each_entry(s, &slab_caches, list) {
5200 err = sysfs_slab_add(s);
5202 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5206 while (alias_list) {
5207 struct saved_alias *al = alias_list;
5209 alias_list = alias_list->next;
5210 err = sysfs_slab_alias(al->s, al->name);
5212 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5217 mutex_unlock(&slab_mutex);
5222 __initcall(slab_sysfs_init);
5223 #endif /* CONFIG_SYSFS */
5226 * The /proc/slabinfo ABI
5228 #ifdef CONFIG_SLABINFO
5229 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5231 unsigned long nr_slabs = 0;
5232 unsigned long nr_objs = 0;
5233 unsigned long nr_free = 0;
5235 struct kmem_cache_node *n;
5237 for_each_kmem_cache_node(s, node, n) {
5238 nr_slabs += node_nr_slabs(n);
5239 nr_objs += node_nr_objs(n);
5240 nr_free += count_partial(n, count_free);
5243 sinfo->active_objs = nr_objs - nr_free;
5244 sinfo->num_objs = nr_objs;
5245 sinfo->active_slabs = nr_slabs;
5246 sinfo->num_slabs = nr_slabs;
5247 sinfo->objects_per_slab = oo_objects(s->oo);
5248 sinfo->cache_order = oo_order(s->oo);
5251 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5255 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5256 size_t count, loff_t *ppos)
5260 #endif /* CONFIG_SLABINFO */