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 * Set of flags that will prevent slab merging
175 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
176 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
179 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
180 SLAB_CACHE_DMA | SLAB_NOTRACK)
183 #define OO_MASK ((1 << OO_SHIFT) - 1)
184 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
186 /* Internal SLUB flags */
187 #define __OBJECT_POISON 0x80000000UL /* Poison object */
188 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
191 static struct notifier_block slab_notifier;
195 * Tracking user of a slab.
197 #define TRACK_ADDRS_COUNT 16
199 unsigned long addr; /* Called from address */
200 #ifdef CONFIG_STACKTRACE
201 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
203 int cpu; /* Was running on cpu */
204 int pid; /* Pid context */
205 unsigned long when; /* When did the operation occur */
208 enum track_item { TRACK_ALLOC, TRACK_FREE };
211 static int sysfs_slab_add(struct kmem_cache *);
212 static int sysfs_slab_alias(struct kmem_cache *, const char *);
213 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
215 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
216 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
218 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
221 static inline void stat(const struct kmem_cache *s, enum stat_item si)
223 #ifdef CONFIG_SLUB_STATS
225 * The rmw is racy on a preemptible kernel but this is acceptable, so
226 * avoid this_cpu_add()'s irq-disable overhead.
228 raw_cpu_inc(s->cpu_slab->stat[si]);
232 /********************************************************************
233 * Core slab cache functions
234 *******************************************************************/
236 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
238 return s->node[node];
241 /* Verify that a pointer has an address that is valid within a slab page */
242 static inline int check_valid_pointer(struct kmem_cache *s,
243 struct page *page, const void *object)
250 base = page_address(page);
251 if (object < base || object >= base + page->objects * s->size ||
252 (object - base) % s->size) {
259 static inline void *get_freepointer(struct kmem_cache *s, void *object)
261 return *(void **)(object + s->offset);
264 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
266 prefetch(object + s->offset);
269 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
273 #ifdef CONFIG_DEBUG_PAGEALLOC
274 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
276 p = get_freepointer(s, object);
281 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
283 *(void **)(object + s->offset) = fp;
286 /* Loop over all objects in a slab */
287 #define for_each_object(__p, __s, __addr, __objects) \
288 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
291 /* Determine object index from a given position */
292 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
294 return (p - addr) / s->size;
297 static inline size_t slab_ksize(const struct kmem_cache *s)
299 #ifdef CONFIG_SLUB_DEBUG
301 * Debugging requires use of the padding between object
302 * and whatever may come after it.
304 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
305 return s->object_size;
309 * If we have the need to store the freelist pointer
310 * back there or track user information then we can
311 * only use the space before that information.
313 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
316 * Else we can use all the padding etc for the allocation
321 static inline int order_objects(int order, unsigned long size, int reserved)
323 return ((PAGE_SIZE << order) - reserved) / size;
326 static inline struct kmem_cache_order_objects oo_make(int order,
327 unsigned long size, int reserved)
329 struct kmem_cache_order_objects x = {
330 (order << OO_SHIFT) + order_objects(order, size, reserved)
336 static inline int oo_order(struct kmem_cache_order_objects x)
338 return x.x >> OO_SHIFT;
341 static inline int oo_objects(struct kmem_cache_order_objects x)
343 return x.x & OO_MASK;
347 * Per slab locking using the pagelock
349 static __always_inline void slab_lock(struct page *page)
351 bit_spin_lock(PG_locked, &page->flags);
354 static __always_inline void slab_unlock(struct page *page)
356 __bit_spin_unlock(PG_locked, &page->flags);
359 static inline void set_page_slub_counters(struct page *page, unsigned long counters_new)
362 tmp.counters = counters_new;
364 * page->counters can cover frozen/inuse/objects as well
365 * as page->_count. If we assign to ->counters directly
366 * we run the risk of losing updates to page->_count, so
367 * be careful and only assign to the fields we need.
369 page->frozen = tmp.frozen;
370 page->inuse = tmp.inuse;
371 page->objects = tmp.objects;
374 /* Interrupts must be disabled (for the fallback code to work right) */
375 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
376 void *freelist_old, unsigned long counters_old,
377 void *freelist_new, unsigned long counters_new,
380 VM_BUG_ON(!irqs_disabled());
381 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
382 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
383 if (s->flags & __CMPXCHG_DOUBLE) {
384 if (cmpxchg_double(&page->freelist, &page->counters,
385 freelist_old, counters_old,
386 freelist_new, counters_new))
392 if (page->freelist == freelist_old &&
393 page->counters == counters_old) {
394 page->freelist = freelist_new;
395 set_page_slub_counters(page, counters_new);
403 stat(s, CMPXCHG_DOUBLE_FAIL);
405 #ifdef SLUB_DEBUG_CMPXCHG
406 pr_info("%s %s: cmpxchg double redo ", n, s->name);
412 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
413 void *freelist_old, unsigned long counters_old,
414 void *freelist_new, unsigned long counters_new,
417 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
418 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
419 if (s->flags & __CMPXCHG_DOUBLE) {
420 if (cmpxchg_double(&page->freelist, &page->counters,
421 freelist_old, counters_old,
422 freelist_new, counters_new))
429 local_irq_save(flags);
431 if (page->freelist == freelist_old &&
432 page->counters == counters_old) {
433 page->freelist = freelist_new;
434 set_page_slub_counters(page, counters_new);
436 local_irq_restore(flags);
440 local_irq_restore(flags);
444 stat(s, CMPXCHG_DOUBLE_FAIL);
446 #ifdef SLUB_DEBUG_CMPXCHG
447 pr_info("%s %s: cmpxchg double redo ", n, s->name);
453 #ifdef CONFIG_SLUB_DEBUG
455 * Determine a map of object in use on a page.
457 * Node listlock must be held to guarantee that the page does
458 * not vanish from under us.
460 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
463 void *addr = page_address(page);
465 for (p = page->freelist; p; p = get_freepointer(s, p))
466 set_bit(slab_index(p, s, addr), map);
472 #ifdef CONFIG_SLUB_DEBUG_ON
473 static int slub_debug = DEBUG_DEFAULT_FLAGS;
475 static int slub_debug;
478 static char *slub_debug_slabs;
479 static int disable_higher_order_debug;
484 static void print_section(char *text, u8 *addr, unsigned int length)
486 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
490 static struct track *get_track(struct kmem_cache *s, void *object,
491 enum track_item alloc)
496 p = object + s->offset + sizeof(void *);
498 p = object + s->inuse;
503 static void set_track(struct kmem_cache *s, void *object,
504 enum track_item alloc, unsigned long addr)
506 struct track *p = get_track(s, object, alloc);
509 #ifdef CONFIG_STACKTRACE
510 struct stack_trace trace;
513 trace.nr_entries = 0;
514 trace.max_entries = TRACK_ADDRS_COUNT;
515 trace.entries = p->addrs;
517 save_stack_trace(&trace);
519 /* See rant in lockdep.c */
520 if (trace.nr_entries != 0 &&
521 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
524 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
528 p->cpu = smp_processor_id();
529 p->pid = current->pid;
532 memset(p, 0, sizeof(struct track));
535 static void init_tracking(struct kmem_cache *s, void *object)
537 if (!(s->flags & SLAB_STORE_USER))
540 set_track(s, object, TRACK_FREE, 0UL);
541 set_track(s, object, TRACK_ALLOC, 0UL);
544 static void print_track(const char *s, struct track *t)
549 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
550 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
551 #ifdef CONFIG_STACKTRACE
554 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
556 pr_err("\t%pS\n", (void *)t->addrs[i]);
563 static void print_tracking(struct kmem_cache *s, void *object)
565 if (!(s->flags & SLAB_STORE_USER))
568 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
569 print_track("Freed", get_track(s, object, TRACK_FREE));
572 static void print_page_info(struct page *page)
574 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
575 page, page->objects, page->inuse, page->freelist, page->flags);
579 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
585 vsnprintf(buf, sizeof(buf), fmt, args);
587 pr_err("=============================================================================\n");
588 pr_err("BUG %s (%s): %s\n", s->name, print_tainted(), buf);
589 pr_err("-----------------------------------------------------------------------------\n\n");
591 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
594 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
600 vsnprintf(buf, sizeof(buf), fmt, args);
602 pr_err("FIX %s: %s\n", s->name, buf);
605 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
607 unsigned int off; /* Offset of last byte */
608 u8 *addr = page_address(page);
610 print_tracking(s, p);
612 print_page_info(page);
614 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
615 p, p - addr, get_freepointer(s, p));
618 print_section("Bytes b4 ", p - 16, 16);
620 print_section("Object ", p, min_t(unsigned long, s->object_size,
622 if (s->flags & SLAB_RED_ZONE)
623 print_section("Redzone ", p + s->object_size,
624 s->inuse - s->object_size);
627 off = s->offset + sizeof(void *);
631 if (s->flags & SLAB_STORE_USER)
632 off += 2 * sizeof(struct track);
635 /* Beginning of the filler is the free pointer */
636 print_section("Padding ", p + off, s->size - off);
641 static void object_err(struct kmem_cache *s, struct page *page,
642 u8 *object, char *reason)
644 slab_bug(s, "%s", reason);
645 print_trailer(s, page, object);
648 static void slab_err(struct kmem_cache *s, struct page *page,
649 const char *fmt, ...)
655 vsnprintf(buf, sizeof(buf), fmt, args);
657 slab_bug(s, "%s", buf);
658 print_page_info(page);
662 static void init_object(struct kmem_cache *s, void *object, u8 val)
666 if (s->flags & __OBJECT_POISON) {
667 memset(p, POISON_FREE, s->object_size - 1);
668 p[s->object_size - 1] = POISON_END;
671 if (s->flags & SLAB_RED_ZONE)
672 memset(p + s->object_size, val, s->inuse - s->object_size);
675 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
676 void *from, void *to)
678 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
679 memset(from, data, to - from);
682 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
683 u8 *object, char *what,
684 u8 *start, unsigned int value, unsigned int bytes)
689 fault = memchr_inv(start, value, bytes);
694 while (end > fault && end[-1] == value)
697 slab_bug(s, "%s overwritten", what);
698 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
699 fault, end - 1, fault[0], value);
700 print_trailer(s, page, object);
702 restore_bytes(s, what, value, fault, end);
710 * Bytes of the object to be managed.
711 * If the freepointer may overlay the object then the free
712 * pointer is the first word of the object.
714 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
717 * object + s->object_size
718 * Padding to reach word boundary. This is also used for Redzoning.
719 * Padding is extended by another word if Redzoning is enabled and
720 * object_size == inuse.
722 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
723 * 0xcc (RED_ACTIVE) for objects in use.
726 * Meta data starts here.
728 * A. Free pointer (if we cannot overwrite object on free)
729 * B. Tracking data for SLAB_STORE_USER
730 * C. Padding to reach required alignment boundary or at mininum
731 * one word if debugging is on to be able to detect writes
732 * before the word boundary.
734 * Padding is done using 0x5a (POISON_INUSE)
737 * Nothing is used beyond s->size.
739 * If slabcaches are merged then the object_size and inuse boundaries are mostly
740 * ignored. And therefore no slab options that rely on these boundaries
741 * may be used with merged slabcaches.
744 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
746 unsigned long off = s->inuse; /* The end of info */
749 /* Freepointer is placed after the object. */
750 off += sizeof(void *);
752 if (s->flags & SLAB_STORE_USER)
753 /* We also have user information there */
754 off += 2 * sizeof(struct track);
759 return check_bytes_and_report(s, page, p, "Object padding",
760 p + off, POISON_INUSE, s->size - off);
763 /* Check the pad bytes at the end of a slab page */
764 static int slab_pad_check(struct kmem_cache *s, struct page *page)
772 if (!(s->flags & SLAB_POISON))
775 start = page_address(page);
776 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
777 end = start + length;
778 remainder = length % s->size;
782 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
785 while (end > fault && end[-1] == POISON_INUSE)
788 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
789 print_section("Padding ", end - remainder, remainder);
791 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
795 static int check_object(struct kmem_cache *s, struct page *page,
796 void *object, u8 val)
799 u8 *endobject = object + s->object_size;
801 if (s->flags & SLAB_RED_ZONE) {
802 if (!check_bytes_and_report(s, page, object, "Redzone",
803 endobject, val, s->inuse - s->object_size))
806 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
807 check_bytes_and_report(s, page, p, "Alignment padding",
808 endobject, POISON_INUSE,
809 s->inuse - s->object_size);
813 if (s->flags & SLAB_POISON) {
814 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
815 (!check_bytes_and_report(s, page, p, "Poison", p,
816 POISON_FREE, s->object_size - 1) ||
817 !check_bytes_and_report(s, page, p, "Poison",
818 p + s->object_size - 1, POISON_END, 1)))
821 * check_pad_bytes cleans up on its own.
823 check_pad_bytes(s, page, p);
826 if (!s->offset && val == SLUB_RED_ACTIVE)
828 * Object and freepointer overlap. Cannot check
829 * freepointer while object is allocated.
833 /* Check free pointer validity */
834 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
835 object_err(s, page, p, "Freepointer corrupt");
837 * No choice but to zap it and thus lose the remainder
838 * of the free objects in this slab. May cause
839 * another error because the object count is now wrong.
841 set_freepointer(s, p, NULL);
847 static int check_slab(struct kmem_cache *s, struct page *page)
851 VM_BUG_ON(!irqs_disabled());
853 if (!PageSlab(page)) {
854 slab_err(s, page, "Not a valid slab page");
858 maxobj = order_objects(compound_order(page), s->size, s->reserved);
859 if (page->objects > maxobj) {
860 slab_err(s, page, "objects %u > max %u",
861 s->name, page->objects, maxobj);
864 if (page->inuse > page->objects) {
865 slab_err(s, page, "inuse %u > max %u",
866 s->name, page->inuse, page->objects);
869 /* Slab_pad_check fixes things up after itself */
870 slab_pad_check(s, page);
875 * Determine if a certain object on a page is on the freelist. Must hold the
876 * slab lock to guarantee that the chains are in a consistent state.
878 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
883 unsigned long max_objects;
886 while (fp && nr <= page->objects) {
889 if (!check_valid_pointer(s, page, fp)) {
891 object_err(s, page, object,
892 "Freechain corrupt");
893 set_freepointer(s, object, NULL);
895 slab_err(s, page, "Freepointer corrupt");
896 page->freelist = NULL;
897 page->inuse = page->objects;
898 slab_fix(s, "Freelist cleared");
904 fp = get_freepointer(s, object);
908 max_objects = order_objects(compound_order(page), s->size, s->reserved);
909 if (max_objects > MAX_OBJS_PER_PAGE)
910 max_objects = MAX_OBJS_PER_PAGE;
912 if (page->objects != max_objects) {
913 slab_err(s, page, "Wrong number of objects. Found %d but "
914 "should be %d", page->objects, max_objects);
915 page->objects = max_objects;
916 slab_fix(s, "Number of objects adjusted.");
918 if (page->inuse != page->objects - nr) {
919 slab_err(s, page, "Wrong object count. Counter is %d but "
920 "counted were %d", page->inuse, page->objects - nr);
921 page->inuse = page->objects - nr;
922 slab_fix(s, "Object count adjusted.");
924 return search == NULL;
927 static void trace(struct kmem_cache *s, struct page *page, void *object,
930 if (s->flags & SLAB_TRACE) {
931 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
933 alloc ? "alloc" : "free",
938 print_section("Object ", (void *)object,
946 * Hooks for other subsystems that check memory allocations. In a typical
947 * production configuration these hooks all should produce no code at all.
949 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
951 kmemleak_alloc(ptr, size, 1, flags);
954 static inline void kfree_hook(const void *x)
959 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
961 flags &= gfp_allowed_mask;
962 lockdep_trace_alloc(flags);
963 might_sleep_if(flags & __GFP_WAIT);
965 return should_failslab(s->object_size, flags, s->flags);
968 static inline void slab_post_alloc_hook(struct kmem_cache *s,
969 gfp_t flags, void *object)
971 flags &= gfp_allowed_mask;
972 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
973 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
976 static inline void slab_free_hook(struct kmem_cache *s, void *x)
978 kmemleak_free_recursive(x, s->flags);
981 * Trouble is that we may no longer disable interrupts in the fast path
982 * So in order to make the debug calls that expect irqs to be
983 * disabled we need to disable interrupts temporarily.
985 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
989 local_irq_save(flags);
990 kmemcheck_slab_free(s, x, s->object_size);
991 debug_check_no_locks_freed(x, s->object_size);
992 local_irq_restore(flags);
995 if (!(s->flags & SLAB_DEBUG_OBJECTS))
996 debug_check_no_obj_freed(x, s->object_size);
1000 * Tracking of fully allocated slabs for debugging purposes.
1002 static void add_full(struct kmem_cache *s,
1003 struct kmem_cache_node *n, struct page *page)
1005 if (!(s->flags & SLAB_STORE_USER))
1008 lockdep_assert_held(&n->list_lock);
1009 list_add(&page->lru, &n->full);
1012 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
1014 if (!(s->flags & SLAB_STORE_USER))
1017 lockdep_assert_held(&n->list_lock);
1018 list_del(&page->lru);
1021 /* Tracking of the number of slabs for debugging purposes */
1022 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1024 struct kmem_cache_node *n = get_node(s, node);
1026 return atomic_long_read(&n->nr_slabs);
1029 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1031 return atomic_long_read(&n->nr_slabs);
1034 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1036 struct kmem_cache_node *n = get_node(s, node);
1039 * May be called early in order to allocate a slab for the
1040 * kmem_cache_node structure. Solve the chicken-egg
1041 * dilemma by deferring the increment of the count during
1042 * bootstrap (see early_kmem_cache_node_alloc).
1045 atomic_long_inc(&n->nr_slabs);
1046 atomic_long_add(objects, &n->total_objects);
1049 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1051 struct kmem_cache_node *n = get_node(s, node);
1053 atomic_long_dec(&n->nr_slabs);
1054 atomic_long_sub(objects, &n->total_objects);
1057 /* Object debug checks for alloc/free paths */
1058 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1061 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1064 init_object(s, object, SLUB_RED_INACTIVE);
1065 init_tracking(s, object);
1068 static noinline int alloc_debug_processing(struct kmem_cache *s,
1070 void *object, unsigned long addr)
1072 if (!check_slab(s, page))
1075 if (!check_valid_pointer(s, page, object)) {
1076 object_err(s, page, object, "Freelist Pointer check fails");
1080 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1083 /* Success perform special debug activities for allocs */
1084 if (s->flags & SLAB_STORE_USER)
1085 set_track(s, object, TRACK_ALLOC, addr);
1086 trace(s, page, object, 1);
1087 init_object(s, object, SLUB_RED_ACTIVE);
1091 if (PageSlab(page)) {
1093 * If this is a slab page then lets do the best we can
1094 * to avoid issues in the future. Marking all objects
1095 * as used avoids touching the remaining objects.
1097 slab_fix(s, "Marking all objects used");
1098 page->inuse = page->objects;
1099 page->freelist = NULL;
1104 static noinline struct kmem_cache_node *free_debug_processing(
1105 struct kmem_cache *s, struct page *page, void *object,
1106 unsigned long addr, unsigned long *flags)
1108 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1110 spin_lock_irqsave(&n->list_lock, *flags);
1113 if (!check_slab(s, page))
1116 if (!check_valid_pointer(s, page, object)) {
1117 slab_err(s, page, "Invalid object pointer 0x%p", object);
1121 if (on_freelist(s, page, object)) {
1122 object_err(s, page, object, "Object already free");
1126 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1129 if (unlikely(s != page->slab_cache)) {
1130 if (!PageSlab(page)) {
1131 slab_err(s, page, "Attempt to free object(0x%p) "
1132 "outside of slab", object);
1133 } else if (!page->slab_cache) {
1134 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1138 object_err(s, page, object,
1139 "page slab pointer corrupt.");
1143 if (s->flags & SLAB_STORE_USER)
1144 set_track(s, object, TRACK_FREE, addr);
1145 trace(s, page, object, 0);
1146 init_object(s, object, SLUB_RED_INACTIVE);
1150 * Keep node_lock to preserve integrity
1151 * until the object is actually freed
1157 spin_unlock_irqrestore(&n->list_lock, *flags);
1158 slab_fix(s, "Object at 0x%p not freed", object);
1162 static int __init setup_slub_debug(char *str)
1164 slub_debug = DEBUG_DEFAULT_FLAGS;
1165 if (*str++ != '=' || !*str)
1167 * No options specified. Switch on full debugging.
1173 * No options but restriction on slabs. This means full
1174 * debugging for slabs matching a pattern.
1178 if (tolower(*str) == 'o') {
1180 * Avoid enabling debugging on caches if its minimum order
1181 * would increase as a result.
1183 disable_higher_order_debug = 1;
1190 * Switch off all debugging measures.
1195 * Determine which debug features should be switched on
1197 for (; *str && *str != ','; str++) {
1198 switch (tolower(*str)) {
1200 slub_debug |= SLAB_DEBUG_FREE;
1203 slub_debug |= SLAB_RED_ZONE;
1206 slub_debug |= SLAB_POISON;
1209 slub_debug |= SLAB_STORE_USER;
1212 slub_debug |= SLAB_TRACE;
1215 slub_debug |= SLAB_FAILSLAB;
1218 pr_err("slub_debug option '%c' unknown. skipped\n",
1225 slub_debug_slabs = str + 1;
1230 __setup("slub_debug", setup_slub_debug);
1232 static unsigned long kmem_cache_flags(unsigned long object_size,
1233 unsigned long flags, const char *name,
1234 void (*ctor)(void *))
1237 * Enable debugging if selected on the kernel commandline.
1239 if (slub_debug && (!slub_debug_slabs || (name &&
1240 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1241 flags |= slub_debug;
1246 static inline void setup_object_debug(struct kmem_cache *s,
1247 struct page *page, void *object) {}
1249 static inline int alloc_debug_processing(struct kmem_cache *s,
1250 struct page *page, void *object, unsigned long addr) { return 0; }
1252 static inline struct kmem_cache_node *free_debug_processing(
1253 struct kmem_cache *s, struct page *page, void *object,
1254 unsigned long addr, unsigned long *flags) { return NULL; }
1256 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1258 static inline int check_object(struct kmem_cache *s, struct page *page,
1259 void *object, u8 val) { return 1; }
1260 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1261 struct page *page) {}
1262 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1263 struct page *page) {}
1264 static inline unsigned long kmem_cache_flags(unsigned long object_size,
1265 unsigned long flags, const char *name,
1266 void (*ctor)(void *))
1270 #define slub_debug 0
1272 #define disable_higher_order_debug 0
1274 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1276 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1278 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1280 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1283 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1285 kmemleak_alloc(ptr, size, 1, flags);
1288 static inline void kfree_hook(const void *x)
1293 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1296 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1299 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags,
1300 flags & gfp_allowed_mask);
1303 static inline void slab_free_hook(struct kmem_cache *s, void *x)
1305 kmemleak_free_recursive(x, s->flags);
1308 #endif /* CONFIG_SLUB_DEBUG */
1311 * Slab allocation and freeing
1313 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1314 struct kmem_cache_order_objects oo)
1316 int order = oo_order(oo);
1318 flags |= __GFP_NOTRACK;
1320 if (node == NUMA_NO_NODE)
1321 return alloc_pages(flags, order);
1323 return alloc_pages_exact_node(node, flags, order);
1326 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1329 struct kmem_cache_order_objects oo = s->oo;
1332 flags &= gfp_allowed_mask;
1334 if (flags & __GFP_WAIT)
1337 flags |= s->allocflags;
1340 * Let the initial higher-order allocation fail under memory pressure
1341 * so we fall-back to the minimum order allocation.
1343 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1345 page = alloc_slab_page(alloc_gfp, node, oo);
1346 if (unlikely(!page)) {
1350 * Allocation may have failed due to fragmentation.
1351 * Try a lower order alloc if possible
1353 page = alloc_slab_page(alloc_gfp, node, oo);
1356 stat(s, ORDER_FALLBACK);
1359 if (kmemcheck_enabled && page
1360 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1361 int pages = 1 << oo_order(oo);
1363 kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node);
1366 * Objects from caches that have a constructor don't get
1367 * cleared when they're allocated, so we need to do it here.
1370 kmemcheck_mark_uninitialized_pages(page, pages);
1372 kmemcheck_mark_unallocated_pages(page, pages);
1375 if (flags & __GFP_WAIT)
1376 local_irq_disable();
1380 page->objects = oo_objects(oo);
1381 mod_zone_page_state(page_zone(page),
1382 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1383 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1389 static void setup_object(struct kmem_cache *s, struct page *page,
1392 setup_object_debug(s, page, object);
1393 if (unlikely(s->ctor))
1397 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1405 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1407 page = allocate_slab(s,
1408 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1412 order = compound_order(page);
1413 inc_slabs_node(s, page_to_nid(page), page->objects);
1414 memcg_bind_pages(s, order);
1415 page->slab_cache = s;
1416 __SetPageSlab(page);
1417 if (page->pfmemalloc)
1418 SetPageSlabPfmemalloc(page);
1420 start = page_address(page);
1422 if (unlikely(s->flags & SLAB_POISON))
1423 memset(start, POISON_INUSE, PAGE_SIZE << order);
1426 for_each_object(p, s, start, page->objects) {
1427 setup_object(s, page, last);
1428 set_freepointer(s, last, p);
1431 setup_object(s, page, last);
1432 set_freepointer(s, last, NULL);
1434 page->freelist = start;
1435 page->inuse = page->objects;
1441 static void __free_slab(struct kmem_cache *s, struct page *page)
1443 int order = compound_order(page);
1444 int pages = 1 << order;
1446 if (kmem_cache_debug(s)) {
1449 slab_pad_check(s, page);
1450 for_each_object(p, s, page_address(page),
1452 check_object(s, page, p, SLUB_RED_INACTIVE);
1455 kmemcheck_free_shadow(page, compound_order(page));
1457 mod_zone_page_state(page_zone(page),
1458 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1459 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1462 __ClearPageSlabPfmemalloc(page);
1463 __ClearPageSlab(page);
1465 memcg_release_pages(s, order);
1466 page_mapcount_reset(page);
1467 if (current->reclaim_state)
1468 current->reclaim_state->reclaimed_slab += pages;
1469 __free_memcg_kmem_pages(page, order);
1472 #define need_reserve_slab_rcu \
1473 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1475 static void rcu_free_slab(struct rcu_head *h)
1479 if (need_reserve_slab_rcu)
1480 page = virt_to_head_page(h);
1482 page = container_of((struct list_head *)h, struct page, lru);
1484 __free_slab(page->slab_cache, page);
1487 static void free_slab(struct kmem_cache *s, struct page *page)
1489 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1490 struct rcu_head *head;
1492 if (need_reserve_slab_rcu) {
1493 int order = compound_order(page);
1494 int offset = (PAGE_SIZE << order) - s->reserved;
1496 VM_BUG_ON(s->reserved != sizeof(*head));
1497 head = page_address(page) + offset;
1500 * RCU free overloads the RCU head over the LRU
1502 head = (void *)&page->lru;
1505 call_rcu(head, rcu_free_slab);
1507 __free_slab(s, page);
1510 static void discard_slab(struct kmem_cache *s, struct page *page)
1512 dec_slabs_node(s, page_to_nid(page), page->objects);
1517 * Management of partially allocated slabs.
1520 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1523 if (tail == DEACTIVATE_TO_TAIL)
1524 list_add_tail(&page->lru, &n->partial);
1526 list_add(&page->lru, &n->partial);
1529 static inline void add_partial(struct kmem_cache_node *n,
1530 struct page *page, int tail)
1532 lockdep_assert_held(&n->list_lock);
1533 __add_partial(n, page, tail);
1537 __remove_partial(struct kmem_cache_node *n, struct page *page)
1539 list_del(&page->lru);
1543 static inline void remove_partial(struct kmem_cache_node *n,
1546 lockdep_assert_held(&n->list_lock);
1547 __remove_partial(n, page);
1551 * Remove slab from the partial list, freeze it and
1552 * return the pointer to the freelist.
1554 * Returns a list of objects or NULL if it fails.
1556 static inline void *acquire_slab(struct kmem_cache *s,
1557 struct kmem_cache_node *n, struct page *page,
1558 int mode, int *objects)
1561 unsigned long counters;
1564 lockdep_assert_held(&n->list_lock);
1567 * Zap the freelist and set the frozen bit.
1568 * The old freelist is the list of objects for the
1569 * per cpu allocation list.
1571 freelist = page->freelist;
1572 counters = page->counters;
1573 new.counters = counters;
1574 *objects = new.objects - new.inuse;
1576 new.inuse = page->objects;
1577 new.freelist = NULL;
1579 new.freelist = freelist;
1582 VM_BUG_ON(new.frozen);
1585 if (!__cmpxchg_double_slab(s, page,
1587 new.freelist, new.counters,
1591 remove_partial(n, page);
1596 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1597 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1600 * Try to allocate a partial slab from a specific node.
1602 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1603 struct kmem_cache_cpu *c, gfp_t flags)
1605 struct page *page, *page2;
1606 void *object = NULL;
1611 * Racy check. If we mistakenly see no partial slabs then we
1612 * just allocate an empty slab. If we mistakenly try to get a
1613 * partial slab and there is none available then get_partials()
1616 if (!n || !n->nr_partial)
1619 spin_lock(&n->list_lock);
1620 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1623 if (!pfmemalloc_match(page, flags))
1626 t = acquire_slab(s, n, page, object == NULL, &objects);
1630 available += objects;
1633 stat(s, ALLOC_FROM_PARTIAL);
1636 put_cpu_partial(s, page, 0);
1637 stat(s, CPU_PARTIAL_NODE);
1639 if (!kmem_cache_has_cpu_partial(s)
1640 || available > s->cpu_partial / 2)
1644 spin_unlock(&n->list_lock);
1649 * Get a page from somewhere. Search in increasing NUMA distances.
1651 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1652 struct kmem_cache_cpu *c)
1655 struct zonelist *zonelist;
1658 enum zone_type high_zoneidx = gfp_zone(flags);
1660 unsigned int cpuset_mems_cookie;
1663 * The defrag ratio allows a configuration of the tradeoffs between
1664 * inter node defragmentation and node local allocations. A lower
1665 * defrag_ratio increases the tendency to do local allocations
1666 * instead of attempting to obtain partial slabs from other nodes.
1668 * If the defrag_ratio is set to 0 then kmalloc() always
1669 * returns node local objects. If the ratio is higher then kmalloc()
1670 * may return off node objects because partial slabs are obtained
1671 * from other nodes and filled up.
1673 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1674 * defrag_ratio = 1000) then every (well almost) allocation will
1675 * first attempt to defrag slab caches on other nodes. This means
1676 * scanning over all nodes to look for partial slabs which may be
1677 * expensive if we do it every time we are trying to find a slab
1678 * with available objects.
1680 if (!s->remote_node_defrag_ratio ||
1681 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1685 cpuset_mems_cookie = read_mems_allowed_begin();
1686 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1687 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1688 struct kmem_cache_node *n;
1690 n = get_node(s, zone_to_nid(zone));
1692 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1693 n->nr_partial > s->min_partial) {
1694 object = get_partial_node(s, n, c, flags);
1697 * Don't check read_mems_allowed_retry()
1698 * here - if mems_allowed was updated in
1699 * parallel, that was a harmless race
1700 * between allocation and the cpuset
1707 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1713 * Get a partial page, lock it and return it.
1715 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1716 struct kmem_cache_cpu *c)
1719 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1721 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1722 if (object || node != NUMA_NO_NODE)
1725 return get_any_partial(s, flags, c);
1728 #ifdef CONFIG_PREEMPT
1730 * Calculate the next globally unique transaction for disambiguiation
1731 * during cmpxchg. The transactions start with the cpu number and are then
1732 * incremented by CONFIG_NR_CPUS.
1734 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1737 * No preemption supported therefore also no need to check for
1743 static inline unsigned long next_tid(unsigned long tid)
1745 return tid + TID_STEP;
1748 static inline unsigned int tid_to_cpu(unsigned long tid)
1750 return tid % TID_STEP;
1753 static inline unsigned long tid_to_event(unsigned long tid)
1755 return tid / TID_STEP;
1758 static inline unsigned int init_tid(int cpu)
1763 static inline void note_cmpxchg_failure(const char *n,
1764 const struct kmem_cache *s, unsigned long tid)
1766 #ifdef SLUB_DEBUG_CMPXCHG
1767 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1769 pr_info("%s %s: cmpxchg redo ", n, s->name);
1771 #ifdef CONFIG_PREEMPT
1772 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1773 pr_warn("due to cpu change %d -> %d\n",
1774 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1777 if (tid_to_event(tid) != tid_to_event(actual_tid))
1778 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1779 tid_to_event(tid), tid_to_event(actual_tid));
1781 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1782 actual_tid, tid, next_tid(tid));
1784 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1787 static void init_kmem_cache_cpus(struct kmem_cache *s)
1791 for_each_possible_cpu(cpu)
1792 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1796 * Remove the cpu slab
1798 static void deactivate_slab(struct kmem_cache *s, struct page *page,
1801 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1802 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1804 enum slab_modes l = M_NONE, m = M_NONE;
1806 int tail = DEACTIVATE_TO_HEAD;
1810 if (page->freelist) {
1811 stat(s, DEACTIVATE_REMOTE_FREES);
1812 tail = DEACTIVATE_TO_TAIL;
1816 * Stage one: Free all available per cpu objects back
1817 * to the page freelist while it is still frozen. Leave the
1820 * There is no need to take the list->lock because the page
1823 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1825 unsigned long counters;
1828 prior = page->freelist;
1829 counters = page->counters;
1830 set_freepointer(s, freelist, prior);
1831 new.counters = counters;
1833 VM_BUG_ON(!new.frozen);
1835 } while (!__cmpxchg_double_slab(s, page,
1837 freelist, new.counters,
1838 "drain percpu freelist"));
1840 freelist = nextfree;
1844 * Stage two: Ensure that the page is unfrozen while the
1845 * list presence reflects the actual number of objects
1848 * We setup the list membership and then perform a cmpxchg
1849 * with the count. If there is a mismatch then the page
1850 * is not unfrozen but the page is on the wrong list.
1852 * Then we restart the process which may have to remove
1853 * the page from the list that we just put it on again
1854 * because the number of objects in the slab may have
1859 old.freelist = page->freelist;
1860 old.counters = page->counters;
1861 VM_BUG_ON(!old.frozen);
1863 /* Determine target state of the slab */
1864 new.counters = old.counters;
1867 set_freepointer(s, freelist, old.freelist);
1868 new.freelist = freelist;
1870 new.freelist = old.freelist;
1874 if (!new.inuse && n->nr_partial > s->min_partial)
1876 else if (new.freelist) {
1881 * Taking the spinlock removes the possiblity
1882 * that acquire_slab() will see a slab page that
1885 spin_lock(&n->list_lock);
1889 if (kmem_cache_debug(s) && !lock) {
1892 * This also ensures that the scanning of full
1893 * slabs from diagnostic functions will not see
1896 spin_lock(&n->list_lock);
1904 remove_partial(n, page);
1906 else if (l == M_FULL)
1908 remove_full(s, n, page);
1910 if (m == M_PARTIAL) {
1912 add_partial(n, page, tail);
1915 } else if (m == M_FULL) {
1917 stat(s, DEACTIVATE_FULL);
1918 add_full(s, n, page);
1924 if (!__cmpxchg_double_slab(s, page,
1925 old.freelist, old.counters,
1926 new.freelist, new.counters,
1931 spin_unlock(&n->list_lock);
1934 stat(s, DEACTIVATE_EMPTY);
1935 discard_slab(s, page);
1941 * Unfreeze all the cpu partial slabs.
1943 * This function must be called with interrupts disabled
1944 * for the cpu using c (or some other guarantee must be there
1945 * to guarantee no concurrent accesses).
1947 static void unfreeze_partials(struct kmem_cache *s,
1948 struct kmem_cache_cpu *c)
1950 #ifdef CONFIG_SLUB_CPU_PARTIAL
1951 struct kmem_cache_node *n = NULL, *n2 = NULL;
1952 struct page *page, *discard_page = NULL;
1954 while ((page = c->partial)) {
1958 c->partial = page->next;
1960 n2 = get_node(s, page_to_nid(page));
1963 spin_unlock(&n->list_lock);
1966 spin_lock(&n->list_lock);
1971 old.freelist = page->freelist;
1972 old.counters = page->counters;
1973 VM_BUG_ON(!old.frozen);
1975 new.counters = old.counters;
1976 new.freelist = old.freelist;
1980 } while (!__cmpxchg_double_slab(s, page,
1981 old.freelist, old.counters,
1982 new.freelist, new.counters,
1983 "unfreezing slab"));
1985 if (unlikely(!new.inuse && n->nr_partial > s->min_partial)) {
1986 page->next = discard_page;
1987 discard_page = page;
1989 add_partial(n, page, DEACTIVATE_TO_TAIL);
1990 stat(s, FREE_ADD_PARTIAL);
1995 spin_unlock(&n->list_lock);
1997 while (discard_page) {
1998 page = discard_page;
1999 discard_page = discard_page->next;
2001 stat(s, DEACTIVATE_EMPTY);
2002 discard_slab(s, page);
2009 * Put a page that was just frozen (in __slab_free) into a partial page
2010 * slot if available. This is done without interrupts disabled and without
2011 * preemption disabled. The cmpxchg is racy and may put the partial page
2012 * onto a random cpus partial slot.
2014 * If we did not find a slot then simply move all the partials to the
2015 * per node partial list.
2017 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2019 #ifdef CONFIG_SLUB_CPU_PARTIAL
2020 struct page *oldpage;
2027 oldpage = this_cpu_read(s->cpu_slab->partial);
2030 pobjects = oldpage->pobjects;
2031 pages = oldpage->pages;
2032 if (drain && pobjects > s->cpu_partial) {
2033 unsigned long flags;
2035 * partial array is full. Move the existing
2036 * set to the per node partial list.
2038 local_irq_save(flags);
2039 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2040 local_irq_restore(flags);
2044 stat(s, CPU_PARTIAL_DRAIN);
2049 pobjects += page->objects - page->inuse;
2051 page->pages = pages;
2052 page->pobjects = pobjects;
2053 page->next = oldpage;
2055 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2060 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2062 stat(s, CPUSLAB_FLUSH);
2063 deactivate_slab(s, c->page, c->freelist);
2065 c->tid = next_tid(c->tid);
2073 * Called from IPI handler with interrupts disabled.
2075 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2077 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2083 unfreeze_partials(s, c);
2087 static void flush_cpu_slab(void *d)
2089 struct kmem_cache *s = d;
2091 __flush_cpu_slab(s, smp_processor_id());
2094 static bool has_cpu_slab(int cpu, void *info)
2096 struct kmem_cache *s = info;
2097 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2099 return c->page || c->partial;
2102 static void flush_all(struct kmem_cache *s)
2104 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2108 * Check if the objects in a per cpu structure fit numa
2109 * locality expectations.
2111 static inline int node_match(struct page *page, int node)
2114 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2120 static int count_free(struct page *page)
2122 return page->objects - page->inuse;
2125 static unsigned long count_partial(struct kmem_cache_node *n,
2126 int (*get_count)(struct page *))
2128 unsigned long flags;
2129 unsigned long x = 0;
2132 spin_lock_irqsave(&n->list_lock, flags);
2133 list_for_each_entry(page, &n->partial, lru)
2134 x += get_count(page);
2135 spin_unlock_irqrestore(&n->list_lock, flags);
2139 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2141 #ifdef CONFIG_SLUB_DEBUG
2142 return atomic_long_read(&n->total_objects);
2148 static noinline void
2149 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2153 pr_warn("SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2155 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2156 s->name, s->object_size, s->size, oo_order(s->oo),
2159 if (oo_order(s->min) > get_order(s->object_size))
2160 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2163 for_each_online_node(node) {
2164 struct kmem_cache_node *n = get_node(s, node);
2165 unsigned long nr_slabs;
2166 unsigned long nr_objs;
2167 unsigned long nr_free;
2172 nr_free = count_partial(n, count_free);
2173 nr_slabs = node_nr_slabs(n);
2174 nr_objs = node_nr_objs(n);
2176 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2177 node, nr_slabs, nr_objs, nr_free);
2181 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2182 int node, struct kmem_cache_cpu **pc)
2185 struct kmem_cache_cpu *c = *pc;
2188 freelist = get_partial(s, flags, node, c);
2193 page = new_slab(s, flags, node);
2195 c = __this_cpu_ptr(s->cpu_slab);
2200 * No other reference to the page yet so we can
2201 * muck around with it freely without cmpxchg
2203 freelist = page->freelist;
2204 page->freelist = NULL;
2206 stat(s, ALLOC_SLAB);
2215 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2217 if (unlikely(PageSlabPfmemalloc(page)))
2218 return gfp_pfmemalloc_allowed(gfpflags);
2224 * Check the page->freelist of a page and either transfer the freelist to the
2225 * per cpu freelist or deactivate the page.
2227 * The page is still frozen if the return value is not NULL.
2229 * If this function returns NULL then the page has been unfrozen.
2231 * This function must be called with interrupt disabled.
2233 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2236 unsigned long counters;
2240 freelist = page->freelist;
2241 counters = page->counters;
2243 new.counters = counters;
2244 VM_BUG_ON(!new.frozen);
2246 new.inuse = page->objects;
2247 new.frozen = freelist != NULL;
2249 } while (!__cmpxchg_double_slab(s, page,
2258 * Slow path. The lockless freelist is empty or we need to perform
2261 * Processing is still very fast if new objects have been freed to the
2262 * regular freelist. In that case we simply take over the regular freelist
2263 * as the lockless freelist and zap the regular freelist.
2265 * If that is not working then we fall back to the partial lists. We take the
2266 * first element of the freelist as the object to allocate now and move the
2267 * rest of the freelist to the lockless freelist.
2269 * And if we were unable to get a new slab from the partial slab lists then
2270 * we need to allocate a new slab. This is the slowest path since it involves
2271 * a call to the page allocator and the setup of a new slab.
2273 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2274 unsigned long addr, struct kmem_cache_cpu *c)
2278 unsigned long flags;
2280 local_irq_save(flags);
2281 #ifdef CONFIG_PREEMPT
2283 * We may have been preempted and rescheduled on a different
2284 * cpu before disabling interrupts. Need to reload cpu area
2287 c = this_cpu_ptr(s->cpu_slab);
2295 if (unlikely(!node_match(page, node))) {
2296 stat(s, ALLOC_NODE_MISMATCH);
2297 deactivate_slab(s, page, c->freelist);
2304 * By rights, we should be searching for a slab page that was
2305 * PFMEMALLOC but right now, we are losing the pfmemalloc
2306 * information when the page leaves the per-cpu allocator
2308 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2309 deactivate_slab(s, page, c->freelist);
2315 /* must check again c->freelist in case of cpu migration or IRQ */
2316 freelist = c->freelist;
2320 stat(s, ALLOC_SLOWPATH);
2322 freelist = get_freelist(s, page);
2326 stat(s, DEACTIVATE_BYPASS);
2330 stat(s, ALLOC_REFILL);
2334 * freelist is pointing to the list of objects to be used.
2335 * page is pointing to the page from which the objects are obtained.
2336 * That page must be frozen for per cpu allocations to work.
2338 VM_BUG_ON(!c->page->frozen);
2339 c->freelist = get_freepointer(s, freelist);
2340 c->tid = next_tid(c->tid);
2341 local_irq_restore(flags);
2347 page = c->page = c->partial;
2348 c->partial = page->next;
2349 stat(s, CPU_PARTIAL_ALLOC);
2354 freelist = new_slab_objects(s, gfpflags, node, &c);
2356 if (unlikely(!freelist)) {
2357 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2358 slab_out_of_memory(s, gfpflags, node);
2360 local_irq_restore(flags);
2365 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2368 /* Only entered in the debug case */
2369 if (kmem_cache_debug(s) &&
2370 !alloc_debug_processing(s, page, freelist, addr))
2371 goto new_slab; /* Slab failed checks. Next slab needed */
2373 deactivate_slab(s, page, get_freepointer(s, freelist));
2376 local_irq_restore(flags);
2381 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2382 * have the fastpath folded into their functions. So no function call
2383 * overhead for requests that can be satisfied on the fastpath.
2385 * The fastpath works by first checking if the lockless freelist can be used.
2386 * If not then __slab_alloc is called for slow processing.
2388 * Otherwise we can simply pick the next object from the lockless free list.
2390 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2391 gfp_t gfpflags, int node, unsigned long addr)
2394 struct kmem_cache_cpu *c;
2398 if (slab_pre_alloc_hook(s, gfpflags))
2401 s = memcg_kmem_get_cache(s, gfpflags);
2404 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2405 * enabled. We may switch back and forth between cpus while
2406 * reading from one cpu area. That does not matter as long
2407 * as we end up on the original cpu again when doing the cmpxchg.
2409 * Preemption is disabled for the retrieval of the tid because that
2410 * must occur from the current processor. We cannot allow rescheduling
2411 * on a different processor between the determination of the pointer
2412 * and the retrieval of the tid.
2415 c = __this_cpu_ptr(s->cpu_slab);
2418 * The transaction ids are globally unique per cpu and per operation on
2419 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2420 * occurs on the right processor and that there was no operation on the
2421 * linked list in between.
2426 object = c->freelist;
2428 if (unlikely(!object || !node_match(page, node)))
2429 object = __slab_alloc(s, gfpflags, node, addr, c);
2432 void *next_object = get_freepointer_safe(s, object);
2435 * The cmpxchg will only match if there was no additional
2436 * operation and if we are on the right processor.
2438 * The cmpxchg does the following atomically (without lock
2440 * 1. Relocate first pointer to the current per cpu area.
2441 * 2. Verify that tid and freelist have not been changed
2442 * 3. If they were not changed replace tid and freelist
2444 * Since this is without lock semantics the protection is only
2445 * against code executing on this cpu *not* from access by
2448 if (unlikely(!this_cpu_cmpxchg_double(
2449 s->cpu_slab->freelist, s->cpu_slab->tid,
2451 next_object, next_tid(tid)))) {
2453 note_cmpxchg_failure("slab_alloc", s, tid);
2456 prefetch_freepointer(s, next_object);
2457 stat(s, ALLOC_FASTPATH);
2460 if (unlikely(gfpflags & __GFP_ZERO) && object)
2461 memset(object, 0, s->object_size);
2463 slab_post_alloc_hook(s, gfpflags, object);
2468 static __always_inline void *slab_alloc(struct kmem_cache *s,
2469 gfp_t gfpflags, unsigned long addr)
2471 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2474 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2476 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2478 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2483 EXPORT_SYMBOL(kmem_cache_alloc);
2485 #ifdef CONFIG_TRACING
2486 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2488 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2489 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2492 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2496 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2498 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2500 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2501 s->object_size, s->size, gfpflags, node);
2505 EXPORT_SYMBOL(kmem_cache_alloc_node);
2507 #ifdef CONFIG_TRACING
2508 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2510 int node, size_t size)
2512 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2514 trace_kmalloc_node(_RET_IP_, ret,
2515 size, s->size, gfpflags, node);
2518 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2523 * Slow patch handling. This may still be called frequently since objects
2524 * have a longer lifetime than the cpu slabs in most processing loads.
2526 * So we still attempt to reduce cache line usage. Just take the slab
2527 * lock and free the item. If there is no additional partial page
2528 * handling required then we can return immediately.
2530 static void __slab_free(struct kmem_cache *s, struct page *page,
2531 void *x, unsigned long addr)
2534 void **object = (void *)x;
2537 unsigned long counters;
2538 struct kmem_cache_node *n = NULL;
2539 unsigned long uninitialized_var(flags);
2541 stat(s, FREE_SLOWPATH);
2543 if (kmem_cache_debug(s) &&
2544 !(n = free_debug_processing(s, page, x, addr, &flags)))
2549 spin_unlock_irqrestore(&n->list_lock, flags);
2552 prior = page->freelist;
2553 counters = page->counters;
2554 set_freepointer(s, object, prior);
2555 new.counters = counters;
2556 was_frozen = new.frozen;
2558 if ((!new.inuse || !prior) && !was_frozen) {
2560 if (kmem_cache_has_cpu_partial(s) && !prior) {
2563 * Slab was on no list before and will be
2565 * We can defer the list move and instead
2570 } else { /* Needs to be taken off a list */
2572 n = get_node(s, page_to_nid(page));
2574 * Speculatively acquire the list_lock.
2575 * If the cmpxchg does not succeed then we may
2576 * drop the list_lock without any processing.
2578 * Otherwise the list_lock will synchronize with
2579 * other processors updating the list of slabs.
2581 spin_lock_irqsave(&n->list_lock, flags);
2586 } while (!cmpxchg_double_slab(s, page,
2588 object, new.counters,
2594 * If we just froze the page then put it onto the
2595 * per cpu partial list.
2597 if (new.frozen && !was_frozen) {
2598 put_cpu_partial(s, page, 1);
2599 stat(s, CPU_PARTIAL_FREE);
2602 * The list lock was not taken therefore no list
2603 * activity can be necessary.
2606 stat(s, FREE_FROZEN);
2610 if (unlikely(!new.inuse && n->nr_partial > s->min_partial))
2614 * Objects left in the slab. If it was not on the partial list before
2617 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2618 if (kmem_cache_debug(s))
2619 remove_full(s, n, page);
2620 add_partial(n, page, DEACTIVATE_TO_TAIL);
2621 stat(s, FREE_ADD_PARTIAL);
2623 spin_unlock_irqrestore(&n->list_lock, flags);
2629 * Slab on the partial list.
2631 remove_partial(n, page);
2632 stat(s, FREE_REMOVE_PARTIAL);
2634 /* Slab must be on the full list */
2635 remove_full(s, n, page);
2638 spin_unlock_irqrestore(&n->list_lock, flags);
2640 discard_slab(s, page);
2644 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2645 * can perform fastpath freeing without additional function calls.
2647 * The fastpath is only possible if we are freeing to the current cpu slab
2648 * of this processor. This typically the case if we have just allocated
2651 * If fastpath is not possible then fall back to __slab_free where we deal
2652 * with all sorts of special processing.
2654 static __always_inline void slab_free(struct kmem_cache *s,
2655 struct page *page, void *x, unsigned long addr)
2657 void **object = (void *)x;
2658 struct kmem_cache_cpu *c;
2661 slab_free_hook(s, x);
2665 * Determine the currently cpus per cpu slab.
2666 * The cpu may change afterward. However that does not matter since
2667 * data is retrieved via this pointer. If we are on the same cpu
2668 * during the cmpxchg then the free will succedd.
2671 c = __this_cpu_ptr(s->cpu_slab);
2676 if (likely(page == c->page)) {
2677 set_freepointer(s, object, c->freelist);
2679 if (unlikely(!this_cpu_cmpxchg_double(
2680 s->cpu_slab->freelist, s->cpu_slab->tid,
2682 object, next_tid(tid)))) {
2684 note_cmpxchg_failure("slab_free", s, tid);
2687 stat(s, FREE_FASTPATH);
2689 __slab_free(s, page, x, addr);
2693 void kmem_cache_free(struct kmem_cache *s, void *x)
2695 s = cache_from_obj(s, x);
2698 slab_free(s, virt_to_head_page(x), x, _RET_IP_);
2699 trace_kmem_cache_free(_RET_IP_, x);
2701 EXPORT_SYMBOL(kmem_cache_free);
2704 * Object placement in a slab is made very easy because we always start at
2705 * offset 0. If we tune the size of the object to the alignment then we can
2706 * get the required alignment by putting one properly sized object after
2709 * Notice that the allocation order determines the sizes of the per cpu
2710 * caches. Each processor has always one slab available for allocations.
2711 * Increasing the allocation order reduces the number of times that slabs
2712 * must be moved on and off the partial lists and is therefore a factor in
2717 * Mininum / Maximum order of slab pages. This influences locking overhead
2718 * and slab fragmentation. A higher order reduces the number of partial slabs
2719 * and increases the number of allocations possible without having to
2720 * take the list_lock.
2722 static int slub_min_order;
2723 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2724 static int slub_min_objects;
2727 * Merge control. If this is set then no merging of slab caches will occur.
2728 * (Could be removed. This was introduced to pacify the merge skeptics.)
2730 static int slub_nomerge;
2733 * Calculate the order of allocation given an slab object size.
2735 * The order of allocation has significant impact on performance and other
2736 * system components. Generally order 0 allocations should be preferred since
2737 * order 0 does not cause fragmentation in the page allocator. Larger objects
2738 * be problematic to put into order 0 slabs because there may be too much
2739 * unused space left. We go to a higher order if more than 1/16th of the slab
2742 * In order to reach satisfactory performance we must ensure that a minimum
2743 * number of objects is in one slab. Otherwise we may generate too much
2744 * activity on the partial lists which requires taking the list_lock. This is
2745 * less a concern for large slabs though which are rarely used.
2747 * slub_max_order specifies the order where we begin to stop considering the
2748 * number of objects in a slab as critical. If we reach slub_max_order then
2749 * we try to keep the page order as low as possible. So we accept more waste
2750 * of space in favor of a small page order.
2752 * Higher order allocations also allow the placement of more objects in a
2753 * slab and thereby reduce object handling overhead. If the user has
2754 * requested a higher mininum order then we start with that one instead of
2755 * the smallest order which will fit the object.
2757 static inline int slab_order(int size, int min_objects,
2758 int max_order, int fract_leftover, int reserved)
2762 int min_order = slub_min_order;
2764 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2765 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2767 for (order = max(min_order,
2768 fls(min_objects * size - 1) - PAGE_SHIFT);
2769 order <= max_order; order++) {
2771 unsigned long slab_size = PAGE_SIZE << order;
2773 if (slab_size < min_objects * size + reserved)
2776 rem = (slab_size - reserved) % size;
2778 if (rem <= slab_size / fract_leftover)
2786 static inline int calculate_order(int size, int reserved)
2794 * Attempt to find best configuration for a slab. This
2795 * works by first attempting to generate a layout with
2796 * the best configuration and backing off gradually.
2798 * First we reduce the acceptable waste in a slab. Then
2799 * we reduce the minimum objects required in a slab.
2801 min_objects = slub_min_objects;
2803 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2804 max_objects = order_objects(slub_max_order, size, reserved);
2805 min_objects = min(min_objects, max_objects);
2807 while (min_objects > 1) {
2809 while (fraction >= 4) {
2810 order = slab_order(size, min_objects,
2811 slub_max_order, fraction, reserved);
2812 if (order <= slub_max_order)
2820 * We were unable to place multiple objects in a slab. Now
2821 * lets see if we can place a single object there.
2823 order = slab_order(size, 1, slub_max_order, 1, reserved);
2824 if (order <= slub_max_order)
2828 * Doh this slab cannot be placed using slub_max_order.
2830 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2831 if (order < MAX_ORDER)
2837 init_kmem_cache_node(struct kmem_cache_node *n)
2840 spin_lock_init(&n->list_lock);
2841 INIT_LIST_HEAD(&n->partial);
2842 #ifdef CONFIG_SLUB_DEBUG
2843 atomic_long_set(&n->nr_slabs, 0);
2844 atomic_long_set(&n->total_objects, 0);
2845 INIT_LIST_HEAD(&n->full);
2849 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2851 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2852 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
2855 * Must align to double word boundary for the double cmpxchg
2856 * instructions to work; see __pcpu_double_call_return_bool().
2858 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2859 2 * sizeof(void *));
2864 init_kmem_cache_cpus(s);
2869 static struct kmem_cache *kmem_cache_node;
2872 * No kmalloc_node yet so do it by hand. We know that this is the first
2873 * slab on the node for this slabcache. There are no concurrent accesses
2876 * Note that this function only works on the kmem_cache_node
2877 * when allocating for the kmem_cache_node. This is used for bootstrapping
2878 * memory on a fresh node that has no slab structures yet.
2880 static void early_kmem_cache_node_alloc(int node)
2883 struct kmem_cache_node *n;
2885 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2887 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2890 if (page_to_nid(page) != node) {
2891 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
2892 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
2897 page->freelist = get_freepointer(kmem_cache_node, n);
2900 kmem_cache_node->node[node] = n;
2901 #ifdef CONFIG_SLUB_DEBUG
2902 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2903 init_tracking(kmem_cache_node, n);
2905 init_kmem_cache_node(n);
2906 inc_slabs_node(kmem_cache_node, node, page->objects);
2909 * No locks need to be taken here as it has just been
2910 * initialized and there is no concurrent access.
2912 __add_partial(n, page, DEACTIVATE_TO_HEAD);
2915 static void free_kmem_cache_nodes(struct kmem_cache *s)
2919 for_each_node_state(node, N_NORMAL_MEMORY) {
2920 struct kmem_cache_node *n = s->node[node];
2923 kmem_cache_free(kmem_cache_node, n);
2925 s->node[node] = NULL;
2929 static int init_kmem_cache_nodes(struct kmem_cache *s)
2933 for_each_node_state(node, N_NORMAL_MEMORY) {
2934 struct kmem_cache_node *n;
2936 if (slab_state == DOWN) {
2937 early_kmem_cache_node_alloc(node);
2940 n = kmem_cache_alloc_node(kmem_cache_node,
2944 free_kmem_cache_nodes(s);
2949 init_kmem_cache_node(n);
2954 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2956 if (min < MIN_PARTIAL)
2958 else if (min > MAX_PARTIAL)
2960 s->min_partial = min;
2964 * calculate_sizes() determines the order and the distribution of data within
2967 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2969 unsigned long flags = s->flags;
2970 unsigned long size = s->object_size;
2974 * Round up object size to the next word boundary. We can only
2975 * place the free pointer at word boundaries and this determines
2976 * the possible location of the free pointer.
2978 size = ALIGN(size, sizeof(void *));
2980 #ifdef CONFIG_SLUB_DEBUG
2982 * Determine if we can poison the object itself. If the user of
2983 * the slab may touch the object after free or before allocation
2984 * then we should never poison the object itself.
2986 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2988 s->flags |= __OBJECT_POISON;
2990 s->flags &= ~__OBJECT_POISON;
2994 * If we are Redzoning then check if there is some space between the
2995 * end of the object and the free pointer. If not then add an
2996 * additional word to have some bytes to store Redzone information.
2998 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
2999 size += sizeof(void *);
3003 * With that we have determined the number of bytes in actual use
3004 * by the object. This is the potential offset to the free pointer.
3008 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
3011 * Relocate free pointer after the object if it is not
3012 * permitted to overwrite the first word of the object on
3015 * This is the case if we do RCU, have a constructor or
3016 * destructor or are poisoning the objects.
3019 size += sizeof(void *);
3022 #ifdef CONFIG_SLUB_DEBUG
3023 if (flags & SLAB_STORE_USER)
3025 * Need to store information about allocs and frees after
3028 size += 2 * sizeof(struct track);
3030 if (flags & SLAB_RED_ZONE)
3032 * Add some empty padding so that we can catch
3033 * overwrites from earlier objects rather than let
3034 * tracking information or the free pointer be
3035 * corrupted if a user writes before the start
3038 size += sizeof(void *);
3042 * SLUB stores one object immediately after another beginning from
3043 * offset 0. In order to align the objects we have to simply size
3044 * each object to conform to the alignment.
3046 size = ALIGN(size, s->align);
3048 if (forced_order >= 0)
3049 order = forced_order;
3051 order = calculate_order(size, s->reserved);
3058 s->allocflags |= __GFP_COMP;
3060 if (s->flags & SLAB_CACHE_DMA)
3061 s->allocflags |= GFP_DMA;
3063 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3064 s->allocflags |= __GFP_RECLAIMABLE;
3067 * Determine the number of objects per slab
3069 s->oo = oo_make(order, size, s->reserved);
3070 s->min = oo_make(get_order(size), size, s->reserved);
3071 if (oo_objects(s->oo) > oo_objects(s->max))
3074 return !!oo_objects(s->oo);
3077 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3079 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3082 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3083 s->reserved = sizeof(struct rcu_head);
3085 if (!calculate_sizes(s, -1))
3087 if (disable_higher_order_debug) {
3089 * Disable debugging flags that store metadata if the min slab
3092 if (get_order(s->size) > get_order(s->object_size)) {
3093 s->flags &= ~DEBUG_METADATA_FLAGS;
3095 if (!calculate_sizes(s, -1))
3100 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3101 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3102 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3103 /* Enable fast mode */
3104 s->flags |= __CMPXCHG_DOUBLE;
3108 * The larger the object size is, the more pages we want on the partial
3109 * list to avoid pounding the page allocator excessively.
3111 set_min_partial(s, ilog2(s->size) / 2);
3114 * cpu_partial determined the maximum number of objects kept in the
3115 * per cpu partial lists of a processor.
3117 * Per cpu partial lists mainly contain slabs that just have one
3118 * object freed. If they are used for allocation then they can be
3119 * filled up again with minimal effort. The slab will never hit the
3120 * per node partial lists and therefore no locking will be required.
3122 * This setting also determines
3124 * A) The number of objects from per cpu partial slabs dumped to the
3125 * per node list when we reach the limit.
3126 * B) The number of objects in cpu partial slabs to extract from the
3127 * per node list when we run out of per cpu objects. We only fetch
3128 * 50% to keep some capacity around for frees.
3130 if (!kmem_cache_has_cpu_partial(s))
3132 else if (s->size >= PAGE_SIZE)
3134 else if (s->size >= 1024)
3136 else if (s->size >= 256)
3137 s->cpu_partial = 13;
3139 s->cpu_partial = 30;
3142 s->remote_node_defrag_ratio = 1000;
3144 if (!init_kmem_cache_nodes(s))
3147 if (alloc_kmem_cache_cpus(s))
3150 free_kmem_cache_nodes(s);
3152 if (flags & SLAB_PANIC)
3153 panic("Cannot create slab %s size=%lu realsize=%u "
3154 "order=%u offset=%u flags=%lx\n",
3155 s->name, (unsigned long)s->size, s->size,
3156 oo_order(s->oo), s->offset, flags);
3160 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3163 #ifdef CONFIG_SLUB_DEBUG
3164 void *addr = page_address(page);
3166 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3167 sizeof(long), GFP_ATOMIC);
3170 slab_err(s, page, text, s->name);
3173 get_map(s, page, map);
3174 for_each_object(p, s, addr, page->objects) {
3176 if (!test_bit(slab_index(p, s, addr), map)) {
3177 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3178 print_tracking(s, p);
3187 * Attempt to free all partial slabs on a node.
3188 * This is called from kmem_cache_close(). We must be the last thread
3189 * using the cache and therefore we do not need to lock anymore.
3191 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3193 struct page *page, *h;
3195 list_for_each_entry_safe(page, h, &n->partial, lru) {
3197 __remove_partial(n, page);
3198 discard_slab(s, page);
3200 list_slab_objects(s, page,
3201 "Objects remaining in %s on kmem_cache_close()");
3207 * Release all resources used by a slab cache.
3209 static inline int kmem_cache_close(struct kmem_cache *s)
3214 /* Attempt to free all objects */
3215 for_each_node_state(node, N_NORMAL_MEMORY) {
3216 struct kmem_cache_node *n = get_node(s, node);
3219 if (n->nr_partial || slabs_node(s, node))
3222 free_percpu(s->cpu_slab);
3223 free_kmem_cache_nodes(s);
3227 int __kmem_cache_shutdown(struct kmem_cache *s)
3229 return kmem_cache_close(s);
3232 /********************************************************************
3234 *******************************************************************/
3236 static int __init setup_slub_min_order(char *str)
3238 get_option(&str, &slub_min_order);
3243 __setup("slub_min_order=", setup_slub_min_order);
3245 static int __init setup_slub_max_order(char *str)
3247 get_option(&str, &slub_max_order);
3248 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3253 __setup("slub_max_order=", setup_slub_max_order);
3255 static int __init setup_slub_min_objects(char *str)
3257 get_option(&str, &slub_min_objects);
3262 __setup("slub_min_objects=", setup_slub_min_objects);
3264 static int __init setup_slub_nomerge(char *str)
3270 __setup("slub_nomerge", setup_slub_nomerge);
3272 void *__kmalloc(size_t size, gfp_t flags)
3274 struct kmem_cache *s;
3277 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3278 return kmalloc_large(size, flags);
3280 s = kmalloc_slab(size, flags);
3282 if (unlikely(ZERO_OR_NULL_PTR(s)))
3285 ret = slab_alloc(s, flags, _RET_IP_);
3287 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3291 EXPORT_SYMBOL(__kmalloc);
3294 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3299 flags |= __GFP_COMP | __GFP_NOTRACK | __GFP_KMEMCG;
3300 page = alloc_pages_node(node, flags, get_order(size));
3302 ptr = page_address(page);
3304 kmalloc_large_node_hook(ptr, size, flags);
3308 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3310 struct kmem_cache *s;
3313 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3314 ret = kmalloc_large_node(size, flags, node);
3316 trace_kmalloc_node(_RET_IP_, ret,
3317 size, PAGE_SIZE << get_order(size),
3323 s = kmalloc_slab(size, flags);
3325 if (unlikely(ZERO_OR_NULL_PTR(s)))
3328 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3330 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3334 EXPORT_SYMBOL(__kmalloc_node);
3337 size_t ksize(const void *object)
3341 if (unlikely(object == ZERO_SIZE_PTR))
3344 page = virt_to_head_page(object);
3346 if (unlikely(!PageSlab(page))) {
3347 WARN_ON(!PageCompound(page));
3348 return PAGE_SIZE << compound_order(page);
3351 return slab_ksize(page->slab_cache);
3353 EXPORT_SYMBOL(ksize);
3355 void kfree(const void *x)
3358 void *object = (void *)x;
3360 trace_kfree(_RET_IP_, x);
3362 if (unlikely(ZERO_OR_NULL_PTR(x)))
3365 page = virt_to_head_page(x);
3366 if (unlikely(!PageSlab(page))) {
3367 BUG_ON(!PageCompound(page));
3369 __free_memcg_kmem_pages(page, compound_order(page));
3372 slab_free(page->slab_cache, page, object, _RET_IP_);
3374 EXPORT_SYMBOL(kfree);
3377 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3378 * the remaining slabs by the number of items in use. The slabs with the
3379 * most items in use come first. New allocations will then fill those up
3380 * and thus they can be removed from the partial lists.
3382 * The slabs with the least items are placed last. This results in them
3383 * being allocated from last increasing the chance that the last objects
3384 * are freed in them.
3386 int kmem_cache_shrink(struct kmem_cache *s)
3390 struct kmem_cache_node *n;
3393 int objects = oo_objects(s->max);
3394 struct list_head *slabs_by_inuse =
3395 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3396 unsigned long flags;
3398 if (!slabs_by_inuse)
3402 for_each_node_state(node, N_NORMAL_MEMORY) {
3403 n = get_node(s, node);
3408 for (i = 0; i < objects; i++)
3409 INIT_LIST_HEAD(slabs_by_inuse + i);
3411 spin_lock_irqsave(&n->list_lock, flags);
3414 * Build lists indexed by the items in use in each slab.
3416 * Note that concurrent frees may occur while we hold the
3417 * list_lock. page->inuse here is the upper limit.
3419 list_for_each_entry_safe(page, t, &n->partial, lru) {
3420 list_move(&page->lru, slabs_by_inuse + page->inuse);
3426 * Rebuild the partial list with the slabs filled up most
3427 * first and the least used slabs at the end.
3429 for (i = objects - 1; i > 0; i--)
3430 list_splice(slabs_by_inuse + i, n->partial.prev);
3432 spin_unlock_irqrestore(&n->list_lock, flags);
3434 /* Release empty slabs */
3435 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3436 discard_slab(s, page);
3439 kfree(slabs_by_inuse);
3442 EXPORT_SYMBOL(kmem_cache_shrink);
3444 static int slab_mem_going_offline_callback(void *arg)
3446 struct kmem_cache *s;
3448 mutex_lock(&slab_mutex);
3449 list_for_each_entry(s, &slab_caches, list)
3450 kmem_cache_shrink(s);
3451 mutex_unlock(&slab_mutex);
3456 static void slab_mem_offline_callback(void *arg)
3458 struct kmem_cache_node *n;
3459 struct kmem_cache *s;
3460 struct memory_notify *marg = arg;
3463 offline_node = marg->status_change_nid_normal;
3466 * If the node still has available memory. we need kmem_cache_node
3469 if (offline_node < 0)
3472 mutex_lock(&slab_mutex);
3473 list_for_each_entry(s, &slab_caches, list) {
3474 n = get_node(s, offline_node);
3477 * if n->nr_slabs > 0, slabs still exist on the node
3478 * that is going down. We were unable to free them,
3479 * and offline_pages() function shouldn't call this
3480 * callback. So, we must fail.
3482 BUG_ON(slabs_node(s, offline_node));
3484 s->node[offline_node] = NULL;
3485 kmem_cache_free(kmem_cache_node, n);
3488 mutex_unlock(&slab_mutex);
3491 static int slab_mem_going_online_callback(void *arg)
3493 struct kmem_cache_node *n;
3494 struct kmem_cache *s;
3495 struct memory_notify *marg = arg;
3496 int nid = marg->status_change_nid_normal;
3500 * If the node's memory is already available, then kmem_cache_node is
3501 * already created. Nothing to do.
3507 * We are bringing a node online. No memory is available yet. We must
3508 * allocate a kmem_cache_node structure in order to bring the node
3511 mutex_lock(&slab_mutex);
3512 list_for_each_entry(s, &slab_caches, list) {
3514 * XXX: kmem_cache_alloc_node will fallback to other nodes
3515 * since memory is not yet available from the node that
3518 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3523 init_kmem_cache_node(n);
3527 mutex_unlock(&slab_mutex);
3531 static int slab_memory_callback(struct notifier_block *self,
3532 unsigned long action, void *arg)
3537 case MEM_GOING_ONLINE:
3538 ret = slab_mem_going_online_callback(arg);
3540 case MEM_GOING_OFFLINE:
3541 ret = slab_mem_going_offline_callback(arg);
3544 case MEM_CANCEL_ONLINE:
3545 slab_mem_offline_callback(arg);
3548 case MEM_CANCEL_OFFLINE:
3552 ret = notifier_from_errno(ret);
3558 static struct notifier_block slab_memory_callback_nb = {
3559 .notifier_call = slab_memory_callback,
3560 .priority = SLAB_CALLBACK_PRI,
3563 /********************************************************************
3564 * Basic setup of slabs
3565 *******************************************************************/
3568 * Used for early kmem_cache structures that were allocated using
3569 * the page allocator. Allocate them properly then fix up the pointers
3570 * that may be pointing to the wrong kmem_cache structure.
3573 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3576 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3578 memcpy(s, static_cache, kmem_cache->object_size);
3581 * This runs very early, and only the boot processor is supposed to be
3582 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3585 __flush_cpu_slab(s, smp_processor_id());
3586 for_each_node_state(node, N_NORMAL_MEMORY) {
3587 struct kmem_cache_node *n = get_node(s, node);
3591 list_for_each_entry(p, &n->partial, lru)
3594 #ifdef CONFIG_SLUB_DEBUG
3595 list_for_each_entry(p, &n->full, lru)
3600 list_add(&s->list, &slab_caches);
3604 void __init kmem_cache_init(void)
3606 static __initdata struct kmem_cache boot_kmem_cache,
3607 boot_kmem_cache_node;
3609 if (debug_guardpage_minorder())
3612 kmem_cache_node = &boot_kmem_cache_node;
3613 kmem_cache = &boot_kmem_cache;
3615 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3616 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3618 register_hotmemory_notifier(&slab_memory_callback_nb);
3620 /* Able to allocate the per node structures */
3621 slab_state = PARTIAL;
3623 create_boot_cache(kmem_cache, "kmem_cache",
3624 offsetof(struct kmem_cache, node) +
3625 nr_node_ids * sizeof(struct kmem_cache_node *),
3626 SLAB_HWCACHE_ALIGN);
3628 kmem_cache = bootstrap(&boot_kmem_cache);
3631 * Allocate kmem_cache_node properly from the kmem_cache slab.
3632 * kmem_cache_node is separately allocated so no need to
3633 * update any list pointers.
3635 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3637 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3638 create_kmalloc_caches(0);
3641 register_cpu_notifier(&slab_notifier);
3644 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
3646 slub_min_order, slub_max_order, slub_min_objects,
3647 nr_cpu_ids, nr_node_ids);
3650 void __init kmem_cache_init_late(void)
3655 * Find a mergeable slab cache
3657 static int slab_unmergeable(struct kmem_cache *s)
3659 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3662 if (!is_root_cache(s))
3669 * We may have set a slab to be unmergeable during bootstrap.
3671 if (s->refcount < 0)
3677 static struct kmem_cache *find_mergeable(size_t size, size_t align,
3678 unsigned long flags, const char *name, void (*ctor)(void *))
3680 struct kmem_cache *s;
3682 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3688 size = ALIGN(size, sizeof(void *));
3689 align = calculate_alignment(flags, align, size);
3690 size = ALIGN(size, align);
3691 flags = kmem_cache_flags(size, flags, name, NULL);
3693 list_for_each_entry(s, &slab_caches, list) {
3694 if (slab_unmergeable(s))
3700 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3703 * Check if alignment is compatible.
3704 * Courtesy of Adrian Drzewiecki
3706 if ((s->size & ~(align - 1)) != s->size)
3709 if (s->size - size >= sizeof(void *))
3718 __kmem_cache_alias(const char *name, size_t size, size_t align,
3719 unsigned long flags, void (*ctor)(void *))
3721 struct kmem_cache *s;
3723 s = find_mergeable(size, align, flags, name, ctor);
3726 struct kmem_cache *c;
3731 * Adjust the object sizes so that we clear
3732 * the complete object on kzalloc.
3734 s->object_size = max(s->object_size, (int)size);
3735 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3737 for_each_memcg_cache_index(i) {
3738 c = cache_from_memcg_idx(s, i);
3741 c->object_size = s->object_size;
3742 c->inuse = max_t(int, c->inuse,
3743 ALIGN(size, sizeof(void *)));
3746 if (sysfs_slab_alias(s, name)) {
3755 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3759 err = kmem_cache_open(s, flags);
3763 /* Mutex is not taken during early boot */
3764 if (slab_state <= UP)
3767 memcg_propagate_slab_attrs(s);
3768 err = sysfs_slab_add(s);
3770 kmem_cache_close(s);
3777 * Use the cpu notifier to insure that the cpu slabs are flushed when
3780 static int slab_cpuup_callback(struct notifier_block *nfb,
3781 unsigned long action, void *hcpu)
3783 long cpu = (long)hcpu;
3784 struct kmem_cache *s;
3785 unsigned long flags;
3788 case CPU_UP_CANCELED:
3789 case CPU_UP_CANCELED_FROZEN:
3791 case CPU_DEAD_FROZEN:
3792 mutex_lock(&slab_mutex);
3793 list_for_each_entry(s, &slab_caches, list) {
3794 local_irq_save(flags);
3795 __flush_cpu_slab(s, cpu);
3796 local_irq_restore(flags);
3798 mutex_unlock(&slab_mutex);
3806 static struct notifier_block slab_notifier = {
3807 .notifier_call = slab_cpuup_callback
3812 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3814 struct kmem_cache *s;
3817 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3818 return kmalloc_large(size, gfpflags);
3820 s = kmalloc_slab(size, gfpflags);
3822 if (unlikely(ZERO_OR_NULL_PTR(s)))
3825 ret = slab_alloc(s, gfpflags, caller);
3827 /* Honor the call site pointer we received. */
3828 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3834 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3835 int node, unsigned long caller)
3837 struct kmem_cache *s;
3840 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3841 ret = kmalloc_large_node(size, gfpflags, node);
3843 trace_kmalloc_node(caller, ret,
3844 size, PAGE_SIZE << get_order(size),
3850 s = kmalloc_slab(size, gfpflags);
3852 if (unlikely(ZERO_OR_NULL_PTR(s)))
3855 ret = slab_alloc_node(s, gfpflags, node, caller);
3857 /* Honor the call site pointer we received. */
3858 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3865 static int count_inuse(struct page *page)
3870 static int count_total(struct page *page)
3872 return page->objects;
3876 #ifdef CONFIG_SLUB_DEBUG
3877 static int validate_slab(struct kmem_cache *s, struct page *page,
3881 void *addr = page_address(page);
3883 if (!check_slab(s, page) ||
3884 !on_freelist(s, page, NULL))
3887 /* Now we know that a valid freelist exists */
3888 bitmap_zero(map, page->objects);
3890 get_map(s, page, map);
3891 for_each_object(p, s, addr, page->objects) {
3892 if (test_bit(slab_index(p, s, addr), map))
3893 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3897 for_each_object(p, s, addr, page->objects)
3898 if (!test_bit(slab_index(p, s, addr), map))
3899 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3904 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3908 validate_slab(s, page, map);
3912 static int validate_slab_node(struct kmem_cache *s,
3913 struct kmem_cache_node *n, unsigned long *map)
3915 unsigned long count = 0;
3917 unsigned long flags;
3919 spin_lock_irqsave(&n->list_lock, flags);
3921 list_for_each_entry(page, &n->partial, lru) {
3922 validate_slab_slab(s, page, map);
3925 if (count != n->nr_partial)
3926 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
3927 s->name, count, n->nr_partial);
3929 if (!(s->flags & SLAB_STORE_USER))
3932 list_for_each_entry(page, &n->full, lru) {
3933 validate_slab_slab(s, page, map);
3936 if (count != atomic_long_read(&n->nr_slabs))
3937 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
3938 s->name, count, atomic_long_read(&n->nr_slabs));
3941 spin_unlock_irqrestore(&n->list_lock, flags);
3945 static long validate_slab_cache(struct kmem_cache *s)
3948 unsigned long count = 0;
3949 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3950 sizeof(unsigned long), GFP_KERNEL);
3956 for_each_node_state(node, N_NORMAL_MEMORY) {
3957 struct kmem_cache_node *n = get_node(s, node);
3959 count += validate_slab_node(s, n, map);
3965 * Generate lists of code addresses where slabcache objects are allocated
3970 unsigned long count;
3977 DECLARE_BITMAP(cpus, NR_CPUS);
3983 unsigned long count;
3984 struct location *loc;
3987 static void free_loc_track(struct loc_track *t)
3990 free_pages((unsigned long)t->loc,
3991 get_order(sizeof(struct location) * t->max));
3994 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3999 order = get_order(sizeof(struct location) * max);
4001 l = (void *)__get_free_pages(flags, order);
4006 memcpy(l, t->loc, sizeof(struct location) * t->count);
4014 static int add_location(struct loc_track *t, struct kmem_cache *s,
4015 const struct track *track)
4017 long start, end, pos;
4019 unsigned long caddr;
4020 unsigned long age = jiffies - track->when;
4026 pos = start + (end - start + 1) / 2;
4029 * There is nothing at "end". If we end up there
4030 * we need to add something to before end.
4035 caddr = t->loc[pos].addr;
4036 if (track->addr == caddr) {
4042 if (age < l->min_time)
4044 if (age > l->max_time)
4047 if (track->pid < l->min_pid)
4048 l->min_pid = track->pid;
4049 if (track->pid > l->max_pid)
4050 l->max_pid = track->pid;
4052 cpumask_set_cpu(track->cpu,
4053 to_cpumask(l->cpus));
4055 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4059 if (track->addr < caddr)
4066 * Not found. Insert new tracking element.
4068 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4074 (t->count - pos) * sizeof(struct location));
4077 l->addr = track->addr;
4081 l->min_pid = track->pid;
4082 l->max_pid = track->pid;
4083 cpumask_clear(to_cpumask(l->cpus));
4084 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4085 nodes_clear(l->nodes);
4086 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4090 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4091 struct page *page, enum track_item alloc,
4094 void *addr = page_address(page);
4097 bitmap_zero(map, page->objects);
4098 get_map(s, page, map);
4100 for_each_object(p, s, addr, page->objects)
4101 if (!test_bit(slab_index(p, s, addr), map))
4102 add_location(t, s, get_track(s, p, alloc));
4105 static int list_locations(struct kmem_cache *s, char *buf,
4106 enum track_item alloc)
4110 struct loc_track t = { 0, 0, NULL };
4112 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4113 sizeof(unsigned long), GFP_KERNEL);
4115 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4118 return sprintf(buf, "Out of memory\n");
4120 /* Push back cpu slabs */
4123 for_each_node_state(node, N_NORMAL_MEMORY) {
4124 struct kmem_cache_node *n = get_node(s, node);
4125 unsigned long flags;
4128 if (!atomic_long_read(&n->nr_slabs))
4131 spin_lock_irqsave(&n->list_lock, flags);
4132 list_for_each_entry(page, &n->partial, lru)
4133 process_slab(&t, s, page, alloc, map);
4134 list_for_each_entry(page, &n->full, lru)
4135 process_slab(&t, s, page, alloc, map);
4136 spin_unlock_irqrestore(&n->list_lock, flags);
4139 for (i = 0; i < t.count; i++) {
4140 struct location *l = &t.loc[i];
4142 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4144 len += sprintf(buf + len, "%7ld ", l->count);
4147 len += sprintf(buf + len, "%pS", (void *)l->addr);
4149 len += sprintf(buf + len, "<not-available>");
4151 if (l->sum_time != l->min_time) {
4152 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4154 (long)div_u64(l->sum_time, l->count),
4157 len += sprintf(buf + len, " age=%ld",
4160 if (l->min_pid != l->max_pid)
4161 len += sprintf(buf + len, " pid=%ld-%ld",
4162 l->min_pid, l->max_pid);
4164 len += sprintf(buf + len, " pid=%ld",
4167 if (num_online_cpus() > 1 &&
4168 !cpumask_empty(to_cpumask(l->cpus)) &&
4169 len < PAGE_SIZE - 60) {
4170 len += sprintf(buf + len, " cpus=");
4171 len += cpulist_scnprintf(buf + len,
4172 PAGE_SIZE - len - 50,
4173 to_cpumask(l->cpus));
4176 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4177 len < PAGE_SIZE - 60) {
4178 len += sprintf(buf + len, " nodes=");
4179 len += nodelist_scnprintf(buf + len,
4180 PAGE_SIZE - len - 50,
4184 len += sprintf(buf + len, "\n");
4190 len += sprintf(buf, "No data\n");
4195 #ifdef SLUB_RESILIENCY_TEST
4196 static void resiliency_test(void)
4200 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4202 pr_err("SLUB resiliency testing\n");
4203 pr_err("-----------------------\n");
4204 pr_err("A. Corruption after allocation\n");
4206 p = kzalloc(16, GFP_KERNEL);
4208 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4211 validate_slab_cache(kmalloc_caches[4]);
4213 /* Hmmm... The next two are dangerous */
4214 p = kzalloc(32, GFP_KERNEL);
4215 p[32 + sizeof(void *)] = 0x34;
4216 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4218 pr_err("If allocated object is overwritten then not detectable\n\n");
4220 validate_slab_cache(kmalloc_caches[5]);
4221 p = kzalloc(64, GFP_KERNEL);
4222 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4224 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4226 pr_err("If allocated object is overwritten then not detectable\n\n");
4227 validate_slab_cache(kmalloc_caches[6]);
4229 pr_err("\nB. Corruption after free\n");
4230 p = kzalloc(128, GFP_KERNEL);
4233 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4234 validate_slab_cache(kmalloc_caches[7]);
4236 p = kzalloc(256, GFP_KERNEL);
4239 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4240 validate_slab_cache(kmalloc_caches[8]);
4242 p = kzalloc(512, GFP_KERNEL);
4245 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4246 validate_slab_cache(kmalloc_caches[9]);
4250 static void resiliency_test(void) {};
4255 enum slab_stat_type {
4256 SL_ALL, /* All slabs */
4257 SL_PARTIAL, /* Only partially allocated slabs */
4258 SL_CPU, /* Only slabs used for cpu caches */
4259 SL_OBJECTS, /* Determine allocated objects not slabs */
4260 SL_TOTAL /* Determine object capacity not slabs */
4263 #define SO_ALL (1 << SL_ALL)
4264 #define SO_PARTIAL (1 << SL_PARTIAL)
4265 #define SO_CPU (1 << SL_CPU)
4266 #define SO_OBJECTS (1 << SL_OBJECTS)
4267 #define SO_TOTAL (1 << SL_TOTAL)
4269 static ssize_t show_slab_objects(struct kmem_cache *s,
4270 char *buf, unsigned long flags)
4272 unsigned long total = 0;
4275 unsigned long *nodes;
4277 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4281 if (flags & SO_CPU) {
4284 for_each_possible_cpu(cpu) {
4285 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4290 page = ACCESS_ONCE(c->page);
4294 node = page_to_nid(page);
4295 if (flags & SO_TOTAL)
4297 else if (flags & SO_OBJECTS)
4305 page = ACCESS_ONCE(c->partial);
4307 node = page_to_nid(page);
4308 if (flags & SO_TOTAL)
4310 else if (flags & SO_OBJECTS)
4320 lock_memory_hotplug();
4321 #ifdef CONFIG_SLUB_DEBUG
4322 if (flags & SO_ALL) {
4323 for_each_node_state(node, N_NORMAL_MEMORY) {
4324 struct kmem_cache_node *n = get_node(s, node);
4326 if (flags & SO_TOTAL)
4327 x = atomic_long_read(&n->total_objects);
4328 else if (flags & SO_OBJECTS)
4329 x = atomic_long_read(&n->total_objects) -
4330 count_partial(n, count_free);
4332 x = atomic_long_read(&n->nr_slabs);
4339 if (flags & SO_PARTIAL) {
4340 for_each_node_state(node, N_NORMAL_MEMORY) {
4341 struct kmem_cache_node *n = get_node(s, node);
4343 if (flags & SO_TOTAL)
4344 x = count_partial(n, count_total);
4345 else if (flags & SO_OBJECTS)
4346 x = count_partial(n, count_inuse);
4353 x = sprintf(buf, "%lu", total);
4355 for_each_node_state(node, N_NORMAL_MEMORY)
4357 x += sprintf(buf + x, " N%d=%lu",
4360 unlock_memory_hotplug();
4362 return x + sprintf(buf + x, "\n");
4365 #ifdef CONFIG_SLUB_DEBUG
4366 static int any_slab_objects(struct kmem_cache *s)
4370 for_each_online_node(node) {
4371 struct kmem_cache_node *n = get_node(s, node);
4376 if (atomic_long_read(&n->total_objects))
4383 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4384 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4386 struct slab_attribute {
4387 struct attribute attr;
4388 ssize_t (*show)(struct kmem_cache *s, char *buf);
4389 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4392 #define SLAB_ATTR_RO(_name) \
4393 static struct slab_attribute _name##_attr = \
4394 __ATTR(_name, 0400, _name##_show, NULL)
4396 #define SLAB_ATTR(_name) \
4397 static struct slab_attribute _name##_attr = \
4398 __ATTR(_name, 0600, _name##_show, _name##_store)
4400 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4402 return sprintf(buf, "%d\n", s->size);
4404 SLAB_ATTR_RO(slab_size);
4406 static ssize_t align_show(struct kmem_cache *s, char *buf)
4408 return sprintf(buf, "%d\n", s->align);
4410 SLAB_ATTR_RO(align);
4412 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4414 return sprintf(buf, "%d\n", s->object_size);
4416 SLAB_ATTR_RO(object_size);
4418 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4420 return sprintf(buf, "%d\n", oo_objects(s->oo));
4422 SLAB_ATTR_RO(objs_per_slab);
4424 static ssize_t order_store(struct kmem_cache *s,
4425 const char *buf, size_t length)
4427 unsigned long order;
4430 err = kstrtoul(buf, 10, &order);
4434 if (order > slub_max_order || order < slub_min_order)
4437 calculate_sizes(s, order);
4441 static ssize_t order_show(struct kmem_cache *s, char *buf)
4443 return sprintf(buf, "%d\n", oo_order(s->oo));
4447 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4449 return sprintf(buf, "%lu\n", s->min_partial);
4452 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4458 err = kstrtoul(buf, 10, &min);
4462 set_min_partial(s, min);
4465 SLAB_ATTR(min_partial);
4467 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4469 return sprintf(buf, "%u\n", s->cpu_partial);
4472 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4475 unsigned long objects;
4478 err = kstrtoul(buf, 10, &objects);
4481 if (objects && !kmem_cache_has_cpu_partial(s))
4484 s->cpu_partial = objects;
4488 SLAB_ATTR(cpu_partial);
4490 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4494 return sprintf(buf, "%pS\n", s->ctor);
4498 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4500 return sprintf(buf, "%d\n", s->refcount - 1);
4502 SLAB_ATTR_RO(aliases);
4504 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4506 return show_slab_objects(s, buf, SO_PARTIAL);
4508 SLAB_ATTR_RO(partial);
4510 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4512 return show_slab_objects(s, buf, SO_CPU);
4514 SLAB_ATTR_RO(cpu_slabs);
4516 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4518 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4520 SLAB_ATTR_RO(objects);
4522 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4524 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4526 SLAB_ATTR_RO(objects_partial);
4528 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4535 for_each_online_cpu(cpu) {
4536 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4539 pages += page->pages;
4540 objects += page->pobjects;
4544 len = sprintf(buf, "%d(%d)", objects, pages);
4547 for_each_online_cpu(cpu) {
4548 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4550 if (page && len < PAGE_SIZE - 20)
4551 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4552 page->pobjects, page->pages);
4555 return len + sprintf(buf + len, "\n");
4557 SLAB_ATTR_RO(slabs_cpu_partial);
4559 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4561 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4564 static ssize_t reclaim_account_store(struct kmem_cache *s,
4565 const char *buf, size_t length)
4567 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4569 s->flags |= SLAB_RECLAIM_ACCOUNT;
4572 SLAB_ATTR(reclaim_account);
4574 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4576 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4578 SLAB_ATTR_RO(hwcache_align);
4580 #ifdef CONFIG_ZONE_DMA
4581 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4583 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4585 SLAB_ATTR_RO(cache_dma);
4588 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4590 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4592 SLAB_ATTR_RO(destroy_by_rcu);
4594 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4596 return sprintf(buf, "%d\n", s->reserved);
4598 SLAB_ATTR_RO(reserved);
4600 #ifdef CONFIG_SLUB_DEBUG
4601 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4603 return show_slab_objects(s, buf, SO_ALL);
4605 SLAB_ATTR_RO(slabs);
4607 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4609 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4611 SLAB_ATTR_RO(total_objects);
4613 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4615 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4618 static ssize_t sanity_checks_store(struct kmem_cache *s,
4619 const char *buf, size_t length)
4621 s->flags &= ~SLAB_DEBUG_FREE;
4622 if (buf[0] == '1') {
4623 s->flags &= ~__CMPXCHG_DOUBLE;
4624 s->flags |= SLAB_DEBUG_FREE;
4628 SLAB_ATTR(sanity_checks);
4630 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4632 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4635 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4638 s->flags &= ~SLAB_TRACE;
4639 if (buf[0] == '1') {
4640 s->flags &= ~__CMPXCHG_DOUBLE;
4641 s->flags |= SLAB_TRACE;
4647 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4649 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4652 static ssize_t red_zone_store(struct kmem_cache *s,
4653 const char *buf, size_t length)
4655 if (any_slab_objects(s))
4658 s->flags &= ~SLAB_RED_ZONE;
4659 if (buf[0] == '1') {
4660 s->flags &= ~__CMPXCHG_DOUBLE;
4661 s->flags |= SLAB_RED_ZONE;
4663 calculate_sizes(s, -1);
4666 SLAB_ATTR(red_zone);
4668 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4670 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4673 static ssize_t poison_store(struct kmem_cache *s,
4674 const char *buf, size_t length)
4676 if (any_slab_objects(s))
4679 s->flags &= ~SLAB_POISON;
4680 if (buf[0] == '1') {
4681 s->flags &= ~__CMPXCHG_DOUBLE;
4682 s->flags |= SLAB_POISON;
4684 calculate_sizes(s, -1);
4689 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4691 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4694 static ssize_t store_user_store(struct kmem_cache *s,
4695 const char *buf, size_t length)
4697 if (any_slab_objects(s))
4700 s->flags &= ~SLAB_STORE_USER;
4701 if (buf[0] == '1') {
4702 s->flags &= ~__CMPXCHG_DOUBLE;
4703 s->flags |= SLAB_STORE_USER;
4705 calculate_sizes(s, -1);
4708 SLAB_ATTR(store_user);
4710 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4715 static ssize_t validate_store(struct kmem_cache *s,
4716 const char *buf, size_t length)
4720 if (buf[0] == '1') {
4721 ret = validate_slab_cache(s);
4727 SLAB_ATTR(validate);
4729 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4731 if (!(s->flags & SLAB_STORE_USER))
4733 return list_locations(s, buf, TRACK_ALLOC);
4735 SLAB_ATTR_RO(alloc_calls);
4737 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4739 if (!(s->flags & SLAB_STORE_USER))
4741 return list_locations(s, buf, TRACK_FREE);
4743 SLAB_ATTR_RO(free_calls);
4744 #endif /* CONFIG_SLUB_DEBUG */
4746 #ifdef CONFIG_FAILSLAB
4747 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4749 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4752 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4755 s->flags &= ~SLAB_FAILSLAB;
4757 s->flags |= SLAB_FAILSLAB;
4760 SLAB_ATTR(failslab);
4763 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4768 static ssize_t shrink_store(struct kmem_cache *s,
4769 const char *buf, size_t length)
4771 if (buf[0] == '1') {
4772 int rc = kmem_cache_shrink(s);
4783 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4785 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4788 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4789 const char *buf, size_t length)
4791 unsigned long ratio;
4794 err = kstrtoul(buf, 10, &ratio);
4799 s->remote_node_defrag_ratio = ratio * 10;
4803 SLAB_ATTR(remote_node_defrag_ratio);
4806 #ifdef CONFIG_SLUB_STATS
4807 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4809 unsigned long sum = 0;
4812 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4817 for_each_online_cpu(cpu) {
4818 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4824 len = sprintf(buf, "%lu", sum);
4827 for_each_online_cpu(cpu) {
4828 if (data[cpu] && len < PAGE_SIZE - 20)
4829 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4833 return len + sprintf(buf + len, "\n");
4836 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4840 for_each_online_cpu(cpu)
4841 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4844 #define STAT_ATTR(si, text) \
4845 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4847 return show_stat(s, buf, si); \
4849 static ssize_t text##_store(struct kmem_cache *s, \
4850 const char *buf, size_t length) \
4852 if (buf[0] != '0') \
4854 clear_stat(s, si); \
4859 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4860 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4861 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4862 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4863 STAT_ATTR(FREE_FROZEN, free_frozen);
4864 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4865 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4866 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4867 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4868 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4869 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
4870 STAT_ATTR(FREE_SLAB, free_slab);
4871 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4872 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4873 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4874 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4875 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4876 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4877 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
4878 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4879 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4880 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4881 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
4882 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
4883 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
4884 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
4887 static struct attribute *slab_attrs[] = {
4888 &slab_size_attr.attr,
4889 &object_size_attr.attr,
4890 &objs_per_slab_attr.attr,
4892 &min_partial_attr.attr,
4893 &cpu_partial_attr.attr,
4895 &objects_partial_attr.attr,
4897 &cpu_slabs_attr.attr,
4901 &hwcache_align_attr.attr,
4902 &reclaim_account_attr.attr,
4903 &destroy_by_rcu_attr.attr,
4905 &reserved_attr.attr,
4906 &slabs_cpu_partial_attr.attr,
4907 #ifdef CONFIG_SLUB_DEBUG
4908 &total_objects_attr.attr,
4910 &sanity_checks_attr.attr,
4912 &red_zone_attr.attr,
4914 &store_user_attr.attr,
4915 &validate_attr.attr,
4916 &alloc_calls_attr.attr,
4917 &free_calls_attr.attr,
4919 #ifdef CONFIG_ZONE_DMA
4920 &cache_dma_attr.attr,
4923 &remote_node_defrag_ratio_attr.attr,
4925 #ifdef CONFIG_SLUB_STATS
4926 &alloc_fastpath_attr.attr,
4927 &alloc_slowpath_attr.attr,
4928 &free_fastpath_attr.attr,
4929 &free_slowpath_attr.attr,
4930 &free_frozen_attr.attr,
4931 &free_add_partial_attr.attr,
4932 &free_remove_partial_attr.attr,
4933 &alloc_from_partial_attr.attr,
4934 &alloc_slab_attr.attr,
4935 &alloc_refill_attr.attr,
4936 &alloc_node_mismatch_attr.attr,
4937 &free_slab_attr.attr,
4938 &cpuslab_flush_attr.attr,
4939 &deactivate_full_attr.attr,
4940 &deactivate_empty_attr.attr,
4941 &deactivate_to_head_attr.attr,
4942 &deactivate_to_tail_attr.attr,
4943 &deactivate_remote_frees_attr.attr,
4944 &deactivate_bypass_attr.attr,
4945 &order_fallback_attr.attr,
4946 &cmpxchg_double_fail_attr.attr,
4947 &cmpxchg_double_cpu_fail_attr.attr,
4948 &cpu_partial_alloc_attr.attr,
4949 &cpu_partial_free_attr.attr,
4950 &cpu_partial_node_attr.attr,
4951 &cpu_partial_drain_attr.attr,
4953 #ifdef CONFIG_FAILSLAB
4954 &failslab_attr.attr,
4960 static struct attribute_group slab_attr_group = {
4961 .attrs = slab_attrs,
4964 static ssize_t slab_attr_show(struct kobject *kobj,
4965 struct attribute *attr,
4968 struct slab_attribute *attribute;
4969 struct kmem_cache *s;
4972 attribute = to_slab_attr(attr);
4975 if (!attribute->show)
4978 err = attribute->show(s, buf);
4983 static ssize_t slab_attr_store(struct kobject *kobj,
4984 struct attribute *attr,
4985 const char *buf, size_t len)
4987 struct slab_attribute *attribute;
4988 struct kmem_cache *s;
4991 attribute = to_slab_attr(attr);
4994 if (!attribute->store)
4997 err = attribute->store(s, buf, len);
4998 #ifdef CONFIG_MEMCG_KMEM
4999 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5002 mutex_lock(&slab_mutex);
5003 if (s->max_attr_size < len)
5004 s->max_attr_size = len;
5007 * This is a best effort propagation, so this function's return
5008 * value will be determined by the parent cache only. This is
5009 * basically because not all attributes will have a well
5010 * defined semantics for rollbacks - most of the actions will
5011 * have permanent effects.
5013 * Returning the error value of any of the children that fail
5014 * is not 100 % defined, in the sense that users seeing the
5015 * error code won't be able to know anything about the state of
5018 * Only returning the error code for the parent cache at least
5019 * has well defined semantics. The cache being written to
5020 * directly either failed or succeeded, in which case we loop
5021 * through the descendants with best-effort propagation.
5023 for_each_memcg_cache_index(i) {
5024 struct kmem_cache *c = cache_from_memcg_idx(s, i);
5026 attribute->store(c, buf, len);
5028 mutex_unlock(&slab_mutex);
5034 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5036 #ifdef CONFIG_MEMCG_KMEM
5038 char *buffer = NULL;
5039 struct kmem_cache *root_cache;
5041 if (is_root_cache(s))
5044 root_cache = s->memcg_params->root_cache;
5047 * This mean this cache had no attribute written. Therefore, no point
5048 * in copying default values around
5050 if (!root_cache->max_attr_size)
5053 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5056 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5058 if (!attr || !attr->store || !attr->show)
5062 * It is really bad that we have to allocate here, so we will
5063 * do it only as a fallback. If we actually allocate, though,
5064 * we can just use the allocated buffer until the end.
5066 * Most of the slub attributes will tend to be very small in
5067 * size, but sysfs allows buffers up to a page, so they can
5068 * theoretically happen.
5072 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5075 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5076 if (WARN_ON(!buffer))
5081 attr->show(root_cache, buf);
5082 attr->store(s, buf, strlen(buf));
5086 free_page((unsigned long)buffer);
5090 static void kmem_cache_release(struct kobject *k)
5092 slab_kmem_cache_release(to_slab(k));
5095 static const struct sysfs_ops slab_sysfs_ops = {
5096 .show = slab_attr_show,
5097 .store = slab_attr_store,
5100 static struct kobj_type slab_ktype = {
5101 .sysfs_ops = &slab_sysfs_ops,
5102 .release = kmem_cache_release,
5105 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5107 struct kobj_type *ktype = get_ktype(kobj);
5109 if (ktype == &slab_ktype)
5114 static const struct kset_uevent_ops slab_uevent_ops = {
5115 .filter = uevent_filter,
5118 static struct kset *slab_kset;
5120 static inline struct kset *cache_kset(struct kmem_cache *s)
5122 #ifdef CONFIG_MEMCG_KMEM
5123 if (!is_root_cache(s))
5124 return s->memcg_params->root_cache->memcg_kset;
5129 #define ID_STR_LENGTH 64
5131 /* Create a unique string id for a slab cache:
5133 * Format :[flags-]size
5135 static char *create_unique_id(struct kmem_cache *s)
5137 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5144 * First flags affecting slabcache operations. We will only
5145 * get here for aliasable slabs so we do not need to support
5146 * too many flags. The flags here must cover all flags that
5147 * are matched during merging to guarantee that the id is
5150 if (s->flags & SLAB_CACHE_DMA)
5152 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5154 if (s->flags & SLAB_DEBUG_FREE)
5156 if (!(s->flags & SLAB_NOTRACK))
5160 p += sprintf(p, "%07d", s->size);
5162 #ifdef CONFIG_MEMCG_KMEM
5163 if (!is_root_cache(s))
5164 p += sprintf(p, "-%08d",
5165 memcg_cache_id(s->memcg_params->memcg));
5168 BUG_ON(p > name + ID_STR_LENGTH - 1);
5172 static int sysfs_slab_add(struct kmem_cache *s)
5176 int unmergeable = slab_unmergeable(s);
5180 * Slabcache can never be merged so we can use the name proper.
5181 * This is typically the case for debug situations. In that
5182 * case we can catch duplicate names easily.
5184 sysfs_remove_link(&slab_kset->kobj, s->name);
5188 * Create a unique name for the slab as a target
5191 name = create_unique_id(s);
5194 s->kobj.kset = cache_kset(s);
5195 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5199 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5203 #ifdef CONFIG_MEMCG_KMEM
5204 if (is_root_cache(s)) {
5205 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5206 if (!s->memcg_kset) {
5213 kobject_uevent(&s->kobj, KOBJ_ADD);
5215 /* Setup first alias */
5216 sysfs_slab_alias(s, s->name);
5223 kobject_del(&s->kobj);
5225 kobject_put(&s->kobj);
5229 void sysfs_slab_remove(struct kmem_cache *s)
5231 if (slab_state < FULL)
5233 * Sysfs has not been setup yet so no need to remove the
5238 #ifdef CONFIG_MEMCG_KMEM
5239 kset_unregister(s->memcg_kset);
5241 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5242 kobject_del(&s->kobj);
5243 kobject_put(&s->kobj);
5247 * Need to buffer aliases during bootup until sysfs becomes
5248 * available lest we lose that information.
5250 struct saved_alias {
5251 struct kmem_cache *s;
5253 struct saved_alias *next;
5256 static struct saved_alias *alias_list;
5258 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5260 struct saved_alias *al;
5262 if (slab_state == FULL) {
5264 * If we have a leftover link then remove it.
5266 sysfs_remove_link(&slab_kset->kobj, name);
5267 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5270 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5276 al->next = alias_list;
5281 static int __init slab_sysfs_init(void)
5283 struct kmem_cache *s;
5286 mutex_lock(&slab_mutex);
5288 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5290 mutex_unlock(&slab_mutex);
5291 pr_err("Cannot register slab subsystem.\n");
5297 list_for_each_entry(s, &slab_caches, list) {
5298 err = sysfs_slab_add(s);
5300 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5304 while (alias_list) {
5305 struct saved_alias *al = alias_list;
5307 alias_list = alias_list->next;
5308 err = sysfs_slab_alias(al->s, al->name);
5310 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5315 mutex_unlock(&slab_mutex);
5320 __initcall(slab_sysfs_init);
5321 #endif /* CONFIG_SYSFS */
5324 * The /proc/slabinfo ABI
5326 #ifdef CONFIG_SLABINFO
5327 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5329 unsigned long nr_slabs = 0;
5330 unsigned long nr_objs = 0;
5331 unsigned long nr_free = 0;
5334 for_each_online_node(node) {
5335 struct kmem_cache_node *n = get_node(s, node);
5340 nr_slabs += node_nr_slabs(n);
5341 nr_objs += node_nr_objs(n);
5342 nr_free += count_partial(n, count_free);
5345 sinfo->active_objs = nr_objs - nr_free;
5346 sinfo->num_objs = nr_objs;
5347 sinfo->active_slabs = nr_slabs;
5348 sinfo->num_slabs = nr_slabs;
5349 sinfo->objects_per_slab = oo_objects(s->oo);
5350 sinfo->cache_order = oo_order(s->oo);
5353 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5357 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5358 size_t count, loff_t *ppos)
5362 #endif /* CONFIG_SLABINFO */