Merge branch 'for-3.2' of git://linux-nfs.org/~bfields/linux
[linux-btrfs-devel.git] / mm / slub.c
blob7c54fe83a90c509b543dbee8a1a7c11f844b9899
1 /*
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
12 #include <linux/mm.h>
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>
19 #include <linux/proc_fs.h>
20 #include <linux/seq_file.h>
21 #include <linux/kmemcheck.h>
22 #include <linux/cpu.h>
23 #include <linux/cpuset.h>
24 #include <linux/mempolicy.h>
25 #include <linux/ctype.h>
26 #include <linux/debugobjects.h>
27 #include <linux/kallsyms.h>
28 #include <linux/memory.h>
29 #include <linux/math64.h>
30 #include <linux/fault-inject.h>
31 #include <linux/stacktrace.h>
33 #include <trace/events/kmem.h>
36 * Lock order:
37 * 1. slub_lock (Global Semaphore)
38 * 2. node->list_lock
39 * 3. slab_lock(page) (Only on some arches and for debugging)
41 * slub_lock
43 * The role of the slub_lock is to protect the list of all the slabs
44 * and to synchronize major metadata changes to slab cache structures.
46 * The slab_lock is only used for debugging and on arches that do not
47 * have the ability to do a cmpxchg_double. It only protects the second
48 * double word in the page struct. Meaning
49 * A. page->freelist -> List of object free in a page
50 * B. page->counters -> Counters of objects
51 * C. page->frozen -> frozen state
53 * If a slab is frozen then it is exempt from list management. It is not
54 * on any list. The processor that froze the slab is the one who can
55 * perform list operations on the page. Other processors may put objects
56 * onto the freelist but the processor that froze the slab is the only
57 * one that can retrieve the objects from the page's freelist.
59 * The list_lock protects the partial and full list on each node and
60 * the partial slab counter. If taken then no new slabs may be added or
61 * removed from the lists nor make the number of partial slabs be modified.
62 * (Note that the total number of slabs is an atomic value that may be
63 * modified without taking the list lock).
65 * The list_lock is a centralized lock and thus we avoid taking it as
66 * much as possible. As long as SLUB does not have to handle partial
67 * slabs, operations can continue without any centralized lock. F.e.
68 * allocating a long series of objects that fill up slabs does not require
69 * the list lock.
70 * Interrupts are disabled during allocation and deallocation in order to
71 * make the slab allocator safe to use in the context of an irq. In addition
72 * interrupts are disabled to ensure that the processor does not change
73 * while handling per_cpu slabs, due to kernel preemption.
75 * SLUB assigns one slab for allocation to each processor.
76 * Allocations only occur from these slabs called cpu slabs.
78 * Slabs with free elements are kept on a partial list and during regular
79 * operations no list for full slabs is used. If an object in a full slab is
80 * freed then the slab will show up again on the partial lists.
81 * We track full slabs for debugging purposes though because otherwise we
82 * cannot scan all objects.
84 * Slabs are freed when they become empty. Teardown and setup is
85 * minimal so we rely on the page allocators per cpu caches for
86 * fast frees and allocs.
88 * Overloading of page flags that are otherwise used for LRU management.
90 * PageActive The slab is frozen and exempt from list processing.
91 * This means that the slab is dedicated to a purpose
92 * such as satisfying allocations for a specific
93 * processor. Objects may be freed in the slab while
94 * it is frozen but slab_free will then skip the usual
95 * list operations. It is up to the processor holding
96 * the slab to integrate the slab into the slab lists
97 * when the slab is no longer needed.
99 * One use of this flag is to mark slabs that are
100 * used for allocations. Then such a slab becomes a cpu
101 * slab. The cpu slab may be equipped with an additional
102 * freelist that allows lockless access to
103 * free objects in addition to the regular freelist
104 * that requires the slab lock.
106 * PageError Slab requires special handling due to debug
107 * options set. This moves slab handling out of
108 * the fast path and disables lockless freelists.
111 #define SLAB_DEBUG_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
112 SLAB_TRACE | SLAB_DEBUG_FREE)
114 static inline int kmem_cache_debug(struct kmem_cache *s)
116 #ifdef CONFIG_SLUB_DEBUG
117 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
118 #else
119 return 0;
120 #endif
124 * Issues still to be resolved:
126 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
128 * - Variable sizing of the per node arrays
131 /* Enable to test recovery from slab corruption on boot */
132 #undef SLUB_RESILIENCY_TEST
134 /* Enable to log cmpxchg failures */
135 #undef SLUB_DEBUG_CMPXCHG
138 * Mininum number of partial slabs. These will be left on the partial
139 * lists even if they are empty. kmem_cache_shrink may reclaim them.
141 #define MIN_PARTIAL 5
144 * Maximum number of desirable partial slabs.
145 * The existence of more partial slabs makes kmem_cache_shrink
146 * sort the partial list by the number of objects in the.
148 #define MAX_PARTIAL 10
150 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
151 SLAB_POISON | SLAB_STORE_USER)
154 * Debugging flags that require metadata to be stored in the slab. These get
155 * disabled when slub_debug=O is used and a cache's min order increases with
156 * metadata.
158 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
161 * Set of flags that will prevent slab merging
163 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
164 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
165 SLAB_FAILSLAB)
167 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
168 SLAB_CACHE_DMA | SLAB_NOTRACK)
170 #define OO_SHIFT 16
171 #define OO_MASK ((1 << OO_SHIFT) - 1)
172 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
174 /* Internal SLUB flags */
175 #define __OBJECT_POISON 0x80000000UL /* Poison object */
176 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
178 static int kmem_size = sizeof(struct kmem_cache);
180 #ifdef CONFIG_SMP
181 static struct notifier_block slab_notifier;
182 #endif
184 static enum {
185 DOWN, /* No slab functionality available */
186 PARTIAL, /* Kmem_cache_node works */
187 UP, /* Everything works but does not show up in sysfs */
188 SYSFS /* Sysfs up */
189 } slab_state = DOWN;
191 /* A list of all slab caches on the system */
192 static DECLARE_RWSEM(slub_lock);
193 static LIST_HEAD(slab_caches);
196 * Tracking user of a slab.
198 #define TRACK_ADDRS_COUNT 16
199 struct track {
200 unsigned long addr; /* Called from address */
201 #ifdef CONFIG_STACKTRACE
202 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
203 #endif
204 int cpu; /* Was running on cpu */
205 int pid; /* Pid context */
206 unsigned long when; /* When did the operation occur */
209 enum track_item { TRACK_ALLOC, TRACK_FREE };
211 #ifdef CONFIG_SYSFS
212 static int sysfs_slab_add(struct kmem_cache *);
213 static int sysfs_slab_alias(struct kmem_cache *, const char *);
214 static void sysfs_slab_remove(struct kmem_cache *);
216 #else
217 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
218 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
219 { return 0; }
220 static inline void sysfs_slab_remove(struct kmem_cache *s)
222 kfree(s->name);
223 kfree(s);
226 #endif
228 static inline void stat(const struct kmem_cache *s, enum stat_item si)
230 #ifdef CONFIG_SLUB_STATS
231 __this_cpu_inc(s->cpu_slab->stat[si]);
232 #endif
235 /********************************************************************
236 * Core slab cache functions
237 *******************************************************************/
239 int slab_is_available(void)
241 return slab_state >= UP;
244 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
246 return s->node[node];
249 /* Verify that a pointer has an address that is valid within a slab page */
250 static inline int check_valid_pointer(struct kmem_cache *s,
251 struct page *page, const void *object)
253 void *base;
255 if (!object)
256 return 1;
258 base = page_address(page);
259 if (object < base || object >= base + page->objects * s->size ||
260 (object - base) % s->size) {
261 return 0;
264 return 1;
267 static inline void *get_freepointer(struct kmem_cache *s, void *object)
269 return *(void **)(object + s->offset);
272 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
274 void *p;
276 #ifdef CONFIG_DEBUG_PAGEALLOC
277 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
278 #else
279 p = get_freepointer(s, object);
280 #endif
281 return p;
284 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
286 *(void **)(object + s->offset) = fp;
289 /* Loop over all objects in a slab */
290 #define for_each_object(__p, __s, __addr, __objects) \
291 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
292 __p += (__s)->size)
294 /* Determine object index from a given position */
295 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
297 return (p - addr) / s->size;
300 static inline size_t slab_ksize(const struct kmem_cache *s)
302 #ifdef CONFIG_SLUB_DEBUG
304 * Debugging requires use of the padding between object
305 * and whatever may come after it.
307 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
308 return s->objsize;
310 #endif
312 * If we have the need to store the freelist pointer
313 * back there or track user information then we can
314 * only use the space before that information.
316 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
317 return s->inuse;
319 * Else we can use all the padding etc for the allocation
321 return s->size;
324 static inline int order_objects(int order, unsigned long size, int reserved)
326 return ((PAGE_SIZE << order) - reserved) / size;
329 static inline struct kmem_cache_order_objects oo_make(int order,
330 unsigned long size, int reserved)
332 struct kmem_cache_order_objects x = {
333 (order << OO_SHIFT) + order_objects(order, size, reserved)
336 return x;
339 static inline int oo_order(struct kmem_cache_order_objects x)
341 return x.x >> OO_SHIFT;
344 static inline int oo_objects(struct kmem_cache_order_objects x)
346 return x.x & OO_MASK;
350 * Per slab locking using the pagelock
352 static __always_inline void slab_lock(struct page *page)
354 bit_spin_lock(PG_locked, &page->flags);
357 static __always_inline void slab_unlock(struct page *page)
359 __bit_spin_unlock(PG_locked, &page->flags);
362 /* Interrupts must be disabled (for the fallback code to work right) */
363 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
364 void *freelist_old, unsigned long counters_old,
365 void *freelist_new, unsigned long counters_new,
366 const char *n)
368 VM_BUG_ON(!irqs_disabled());
369 #ifdef CONFIG_CMPXCHG_DOUBLE
370 if (s->flags & __CMPXCHG_DOUBLE) {
371 if (cmpxchg_double(&page->freelist,
372 freelist_old, counters_old,
373 freelist_new, counters_new))
374 return 1;
375 } else
376 #endif
378 slab_lock(page);
379 if (page->freelist == freelist_old && page->counters == counters_old) {
380 page->freelist = freelist_new;
381 page->counters = counters_new;
382 slab_unlock(page);
383 return 1;
385 slab_unlock(page);
388 cpu_relax();
389 stat(s, CMPXCHG_DOUBLE_FAIL);
391 #ifdef SLUB_DEBUG_CMPXCHG
392 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
393 #endif
395 return 0;
398 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
399 void *freelist_old, unsigned long counters_old,
400 void *freelist_new, unsigned long counters_new,
401 const char *n)
403 #ifdef CONFIG_CMPXCHG_DOUBLE
404 if (s->flags & __CMPXCHG_DOUBLE) {
405 if (cmpxchg_double(&page->freelist,
406 freelist_old, counters_old,
407 freelist_new, counters_new))
408 return 1;
409 } else
410 #endif
412 unsigned long flags;
414 local_irq_save(flags);
415 slab_lock(page);
416 if (page->freelist == freelist_old && page->counters == counters_old) {
417 page->freelist = freelist_new;
418 page->counters = counters_new;
419 slab_unlock(page);
420 local_irq_restore(flags);
421 return 1;
423 slab_unlock(page);
424 local_irq_restore(flags);
427 cpu_relax();
428 stat(s, CMPXCHG_DOUBLE_FAIL);
430 #ifdef SLUB_DEBUG_CMPXCHG
431 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
432 #endif
434 return 0;
437 #ifdef CONFIG_SLUB_DEBUG
439 * Determine a map of object in use on a page.
441 * Node listlock must be held to guarantee that the page does
442 * not vanish from under us.
444 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
446 void *p;
447 void *addr = page_address(page);
449 for (p = page->freelist; p; p = get_freepointer(s, p))
450 set_bit(slab_index(p, s, addr), map);
454 * Debug settings:
456 #ifdef CONFIG_SLUB_DEBUG_ON
457 static int slub_debug = DEBUG_DEFAULT_FLAGS;
458 #else
459 static int slub_debug;
460 #endif
462 static char *slub_debug_slabs;
463 static int disable_higher_order_debug;
466 * Object debugging
468 static void print_section(char *text, u8 *addr, unsigned int length)
470 int i, offset;
471 int newline = 1;
472 char ascii[17];
474 ascii[16] = 0;
476 for (i = 0; i < length; i++) {
477 if (newline) {
478 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
479 newline = 0;
481 printk(KERN_CONT " %02x", addr[i]);
482 offset = i % 16;
483 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
484 if (offset == 15) {
485 printk(KERN_CONT " %s\n", ascii);
486 newline = 1;
489 if (!newline) {
490 i %= 16;
491 while (i < 16) {
492 printk(KERN_CONT " ");
493 ascii[i] = ' ';
494 i++;
496 printk(KERN_CONT " %s\n", ascii);
500 static struct track *get_track(struct kmem_cache *s, void *object,
501 enum track_item alloc)
503 struct track *p;
505 if (s->offset)
506 p = object + s->offset + sizeof(void *);
507 else
508 p = object + s->inuse;
510 return p + alloc;
513 static void set_track(struct kmem_cache *s, void *object,
514 enum track_item alloc, unsigned long addr)
516 struct track *p = get_track(s, object, alloc);
518 if (addr) {
519 #ifdef CONFIG_STACKTRACE
520 struct stack_trace trace;
521 int i;
523 trace.nr_entries = 0;
524 trace.max_entries = TRACK_ADDRS_COUNT;
525 trace.entries = p->addrs;
526 trace.skip = 3;
527 save_stack_trace(&trace);
529 /* See rant in lockdep.c */
530 if (trace.nr_entries != 0 &&
531 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
532 trace.nr_entries--;
534 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
535 p->addrs[i] = 0;
536 #endif
537 p->addr = addr;
538 p->cpu = smp_processor_id();
539 p->pid = current->pid;
540 p->when = jiffies;
541 } else
542 memset(p, 0, sizeof(struct track));
545 static void init_tracking(struct kmem_cache *s, void *object)
547 if (!(s->flags & SLAB_STORE_USER))
548 return;
550 set_track(s, object, TRACK_FREE, 0UL);
551 set_track(s, object, TRACK_ALLOC, 0UL);
554 static void print_track(const char *s, struct track *t)
556 if (!t->addr)
557 return;
559 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
560 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
561 #ifdef CONFIG_STACKTRACE
563 int i;
564 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
565 if (t->addrs[i])
566 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
567 else
568 break;
570 #endif
573 static void print_tracking(struct kmem_cache *s, void *object)
575 if (!(s->flags & SLAB_STORE_USER))
576 return;
578 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
579 print_track("Freed", get_track(s, object, TRACK_FREE));
582 static void print_page_info(struct page *page)
584 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
585 page, page->objects, page->inuse, page->freelist, page->flags);
589 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
591 va_list args;
592 char buf[100];
594 va_start(args, fmt);
595 vsnprintf(buf, sizeof(buf), fmt, args);
596 va_end(args);
597 printk(KERN_ERR "========================================"
598 "=====================================\n");
599 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
600 printk(KERN_ERR "----------------------------------------"
601 "-------------------------------------\n\n");
604 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
606 va_list args;
607 char buf[100];
609 va_start(args, fmt);
610 vsnprintf(buf, sizeof(buf), fmt, args);
611 va_end(args);
612 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
615 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
617 unsigned int off; /* Offset of last byte */
618 u8 *addr = page_address(page);
620 print_tracking(s, p);
622 print_page_info(page);
624 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
625 p, p - addr, get_freepointer(s, p));
627 if (p > addr + 16)
628 print_section("Bytes b4", p - 16, 16);
630 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE));
632 if (s->flags & SLAB_RED_ZONE)
633 print_section("Redzone", p + s->objsize,
634 s->inuse - s->objsize);
636 if (s->offset)
637 off = s->offset + sizeof(void *);
638 else
639 off = s->inuse;
641 if (s->flags & SLAB_STORE_USER)
642 off += 2 * sizeof(struct track);
644 if (off != s->size)
645 /* Beginning of the filler is the free pointer */
646 print_section("Padding", p + off, s->size - off);
648 dump_stack();
651 static void object_err(struct kmem_cache *s, struct page *page,
652 u8 *object, char *reason)
654 slab_bug(s, "%s", reason);
655 print_trailer(s, page, object);
658 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
660 va_list args;
661 char buf[100];
663 va_start(args, fmt);
664 vsnprintf(buf, sizeof(buf), fmt, args);
665 va_end(args);
666 slab_bug(s, "%s", buf);
667 print_page_info(page);
668 dump_stack();
671 static void init_object(struct kmem_cache *s, void *object, u8 val)
673 u8 *p = object;
675 if (s->flags & __OBJECT_POISON) {
676 memset(p, POISON_FREE, s->objsize - 1);
677 p[s->objsize - 1] = POISON_END;
680 if (s->flags & SLAB_RED_ZONE)
681 memset(p + s->objsize, val, s->inuse - s->objsize);
684 static u8 *check_bytes8(u8 *start, u8 value, unsigned int bytes)
686 while (bytes) {
687 if (*start != value)
688 return start;
689 start++;
690 bytes--;
692 return NULL;
695 static u8 *check_bytes(u8 *start, u8 value, unsigned int bytes)
697 u64 value64;
698 unsigned int words, prefix;
700 if (bytes <= 16)
701 return check_bytes8(start, value, bytes);
703 value64 = value | value << 8 | value << 16 | value << 24;
704 value64 = (value64 & 0xffffffff) | value64 << 32;
705 prefix = 8 - ((unsigned long)start) % 8;
707 if (prefix) {
708 u8 *r = check_bytes8(start, value, prefix);
709 if (r)
710 return r;
711 start += prefix;
712 bytes -= prefix;
715 words = bytes / 8;
717 while (words) {
718 if (*(u64 *)start != value64)
719 return check_bytes8(start, value, 8);
720 start += 8;
721 words--;
724 return check_bytes8(start, value, bytes % 8);
727 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
728 void *from, void *to)
730 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
731 memset(from, data, to - from);
734 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
735 u8 *object, char *what,
736 u8 *start, unsigned int value, unsigned int bytes)
738 u8 *fault;
739 u8 *end;
741 fault = check_bytes(start, value, bytes);
742 if (!fault)
743 return 1;
745 end = start + bytes;
746 while (end > fault && end[-1] == value)
747 end--;
749 slab_bug(s, "%s overwritten", what);
750 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
751 fault, end - 1, fault[0], value);
752 print_trailer(s, page, object);
754 restore_bytes(s, what, value, fault, end);
755 return 0;
759 * Object layout:
761 * object address
762 * Bytes of the object to be managed.
763 * If the freepointer may overlay the object then the free
764 * pointer is the first word of the object.
766 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
767 * 0xa5 (POISON_END)
769 * object + s->objsize
770 * Padding to reach word boundary. This is also used for Redzoning.
771 * Padding is extended by another word if Redzoning is enabled and
772 * objsize == inuse.
774 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
775 * 0xcc (RED_ACTIVE) for objects in use.
777 * object + s->inuse
778 * Meta data starts here.
780 * A. Free pointer (if we cannot overwrite object on free)
781 * B. Tracking data for SLAB_STORE_USER
782 * C. Padding to reach required alignment boundary or at mininum
783 * one word if debugging is on to be able to detect writes
784 * before the word boundary.
786 * Padding is done using 0x5a (POISON_INUSE)
788 * object + s->size
789 * Nothing is used beyond s->size.
791 * If slabcaches are merged then the objsize and inuse boundaries are mostly
792 * ignored. And therefore no slab options that rely on these boundaries
793 * may be used with merged slabcaches.
796 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
798 unsigned long off = s->inuse; /* The end of info */
800 if (s->offset)
801 /* Freepointer is placed after the object. */
802 off += sizeof(void *);
804 if (s->flags & SLAB_STORE_USER)
805 /* We also have user information there */
806 off += 2 * sizeof(struct track);
808 if (s->size == off)
809 return 1;
811 return check_bytes_and_report(s, page, p, "Object padding",
812 p + off, POISON_INUSE, s->size - off);
815 /* Check the pad bytes at the end of a slab page */
816 static int slab_pad_check(struct kmem_cache *s, struct page *page)
818 u8 *start;
819 u8 *fault;
820 u8 *end;
821 int length;
822 int remainder;
824 if (!(s->flags & SLAB_POISON))
825 return 1;
827 start = page_address(page);
828 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
829 end = start + length;
830 remainder = length % s->size;
831 if (!remainder)
832 return 1;
834 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
835 if (!fault)
836 return 1;
837 while (end > fault && end[-1] == POISON_INUSE)
838 end--;
840 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
841 print_section("Padding", end - remainder, remainder);
843 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
844 return 0;
847 static int check_object(struct kmem_cache *s, struct page *page,
848 void *object, u8 val)
850 u8 *p = object;
851 u8 *endobject = object + s->objsize;
853 if (s->flags & SLAB_RED_ZONE) {
854 if (!check_bytes_and_report(s, page, object, "Redzone",
855 endobject, val, s->inuse - s->objsize))
856 return 0;
857 } else {
858 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
859 check_bytes_and_report(s, page, p, "Alignment padding",
860 endobject, POISON_INUSE, s->inuse - s->objsize);
864 if (s->flags & SLAB_POISON) {
865 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
866 (!check_bytes_and_report(s, page, p, "Poison", p,
867 POISON_FREE, s->objsize - 1) ||
868 !check_bytes_and_report(s, page, p, "Poison",
869 p + s->objsize - 1, POISON_END, 1)))
870 return 0;
872 * check_pad_bytes cleans up on its own.
874 check_pad_bytes(s, page, p);
877 if (!s->offset && val == SLUB_RED_ACTIVE)
879 * Object and freepointer overlap. Cannot check
880 * freepointer while object is allocated.
882 return 1;
884 /* Check free pointer validity */
885 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
886 object_err(s, page, p, "Freepointer corrupt");
888 * No choice but to zap it and thus lose the remainder
889 * of the free objects in this slab. May cause
890 * another error because the object count is now wrong.
892 set_freepointer(s, p, NULL);
893 return 0;
895 return 1;
898 static int check_slab(struct kmem_cache *s, struct page *page)
900 int maxobj;
902 VM_BUG_ON(!irqs_disabled());
904 if (!PageSlab(page)) {
905 slab_err(s, page, "Not a valid slab page");
906 return 0;
909 maxobj = order_objects(compound_order(page), s->size, s->reserved);
910 if (page->objects > maxobj) {
911 slab_err(s, page, "objects %u > max %u",
912 s->name, page->objects, maxobj);
913 return 0;
915 if (page->inuse > page->objects) {
916 slab_err(s, page, "inuse %u > max %u",
917 s->name, page->inuse, page->objects);
918 return 0;
920 /* Slab_pad_check fixes things up after itself */
921 slab_pad_check(s, page);
922 return 1;
926 * Determine if a certain object on a page is on the freelist. Must hold the
927 * slab lock to guarantee that the chains are in a consistent state.
929 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
931 int nr = 0;
932 void *fp;
933 void *object = NULL;
934 unsigned long max_objects;
936 fp = page->freelist;
937 while (fp && nr <= page->objects) {
938 if (fp == search)
939 return 1;
940 if (!check_valid_pointer(s, page, fp)) {
941 if (object) {
942 object_err(s, page, object,
943 "Freechain corrupt");
944 set_freepointer(s, object, NULL);
945 break;
946 } else {
947 slab_err(s, page, "Freepointer corrupt");
948 page->freelist = NULL;
949 page->inuse = page->objects;
950 slab_fix(s, "Freelist cleared");
951 return 0;
953 break;
955 object = fp;
956 fp = get_freepointer(s, object);
957 nr++;
960 max_objects = order_objects(compound_order(page), s->size, s->reserved);
961 if (max_objects > MAX_OBJS_PER_PAGE)
962 max_objects = MAX_OBJS_PER_PAGE;
964 if (page->objects != max_objects) {
965 slab_err(s, page, "Wrong number of objects. Found %d but "
966 "should be %d", page->objects, max_objects);
967 page->objects = max_objects;
968 slab_fix(s, "Number of objects adjusted.");
970 if (page->inuse != page->objects - nr) {
971 slab_err(s, page, "Wrong object count. Counter is %d but "
972 "counted were %d", page->inuse, page->objects - nr);
973 page->inuse = page->objects - nr;
974 slab_fix(s, "Object count adjusted.");
976 return search == NULL;
979 static void trace(struct kmem_cache *s, struct page *page, void *object,
980 int alloc)
982 if (s->flags & SLAB_TRACE) {
983 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
984 s->name,
985 alloc ? "alloc" : "free",
986 object, page->inuse,
987 page->freelist);
989 if (!alloc)
990 print_section("Object", (void *)object, s->objsize);
992 dump_stack();
997 * Hooks for other subsystems that check memory allocations. In a typical
998 * production configuration these hooks all should produce no code at all.
1000 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1002 flags &= gfp_allowed_mask;
1003 lockdep_trace_alloc(flags);
1004 might_sleep_if(flags & __GFP_WAIT);
1006 return should_failslab(s->objsize, flags, s->flags);
1009 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
1011 flags &= gfp_allowed_mask;
1012 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
1013 kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, flags);
1016 static inline void slab_free_hook(struct kmem_cache *s, void *x)
1018 kmemleak_free_recursive(x, s->flags);
1021 * Trouble is that we may no longer disable interupts in the fast path
1022 * So in order to make the debug calls that expect irqs to be
1023 * disabled we need to disable interrupts temporarily.
1025 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1027 unsigned long flags;
1029 local_irq_save(flags);
1030 kmemcheck_slab_free(s, x, s->objsize);
1031 debug_check_no_locks_freed(x, s->objsize);
1032 local_irq_restore(flags);
1034 #endif
1035 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1036 debug_check_no_obj_freed(x, s->objsize);
1040 * Tracking of fully allocated slabs for debugging purposes.
1042 * list_lock must be held.
1044 static void add_full(struct kmem_cache *s,
1045 struct kmem_cache_node *n, struct page *page)
1047 if (!(s->flags & SLAB_STORE_USER))
1048 return;
1050 list_add(&page->lru, &n->full);
1054 * list_lock must be held.
1056 static void remove_full(struct kmem_cache *s, struct page *page)
1058 if (!(s->flags & SLAB_STORE_USER))
1059 return;
1061 list_del(&page->lru);
1064 /* Tracking of the number of slabs for debugging purposes */
1065 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1067 struct kmem_cache_node *n = get_node(s, node);
1069 return atomic_long_read(&n->nr_slabs);
1072 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1074 return atomic_long_read(&n->nr_slabs);
1077 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1079 struct kmem_cache_node *n = get_node(s, node);
1082 * May be called early in order to allocate a slab for the
1083 * kmem_cache_node structure. Solve the chicken-egg
1084 * dilemma by deferring the increment of the count during
1085 * bootstrap (see early_kmem_cache_node_alloc).
1087 if (n) {
1088 atomic_long_inc(&n->nr_slabs);
1089 atomic_long_add(objects, &n->total_objects);
1092 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1094 struct kmem_cache_node *n = get_node(s, node);
1096 atomic_long_dec(&n->nr_slabs);
1097 atomic_long_sub(objects, &n->total_objects);
1100 /* Object debug checks for alloc/free paths */
1101 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1102 void *object)
1104 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1105 return;
1107 init_object(s, object, SLUB_RED_INACTIVE);
1108 init_tracking(s, object);
1111 static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
1112 void *object, unsigned long addr)
1114 if (!check_slab(s, page))
1115 goto bad;
1117 if (!check_valid_pointer(s, page, object)) {
1118 object_err(s, page, object, "Freelist Pointer check fails");
1119 goto bad;
1122 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1123 goto bad;
1125 /* Success perform special debug activities for allocs */
1126 if (s->flags & SLAB_STORE_USER)
1127 set_track(s, object, TRACK_ALLOC, addr);
1128 trace(s, page, object, 1);
1129 init_object(s, object, SLUB_RED_ACTIVE);
1130 return 1;
1132 bad:
1133 if (PageSlab(page)) {
1135 * If this is a slab page then lets do the best we can
1136 * to avoid issues in the future. Marking all objects
1137 * as used avoids touching the remaining objects.
1139 slab_fix(s, "Marking all objects used");
1140 page->inuse = page->objects;
1141 page->freelist = NULL;
1143 return 0;
1146 static noinline int free_debug_processing(struct kmem_cache *s,
1147 struct page *page, void *object, unsigned long addr)
1149 unsigned long flags;
1150 int rc = 0;
1152 local_irq_save(flags);
1153 slab_lock(page);
1155 if (!check_slab(s, page))
1156 goto fail;
1158 if (!check_valid_pointer(s, page, object)) {
1159 slab_err(s, page, "Invalid object pointer 0x%p", object);
1160 goto fail;
1163 if (on_freelist(s, page, object)) {
1164 object_err(s, page, object, "Object already free");
1165 goto fail;
1168 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1169 goto out;
1171 if (unlikely(s != page->slab)) {
1172 if (!PageSlab(page)) {
1173 slab_err(s, page, "Attempt to free object(0x%p) "
1174 "outside of slab", object);
1175 } else if (!page->slab) {
1176 printk(KERN_ERR
1177 "SLUB <none>: no slab for object 0x%p.\n",
1178 object);
1179 dump_stack();
1180 } else
1181 object_err(s, page, object,
1182 "page slab pointer corrupt.");
1183 goto fail;
1186 if (s->flags & SLAB_STORE_USER)
1187 set_track(s, object, TRACK_FREE, addr);
1188 trace(s, page, object, 0);
1189 init_object(s, object, SLUB_RED_INACTIVE);
1190 rc = 1;
1191 out:
1192 slab_unlock(page);
1193 local_irq_restore(flags);
1194 return rc;
1196 fail:
1197 slab_fix(s, "Object at 0x%p not freed", object);
1198 goto out;
1201 static int __init setup_slub_debug(char *str)
1203 slub_debug = DEBUG_DEFAULT_FLAGS;
1204 if (*str++ != '=' || !*str)
1206 * No options specified. Switch on full debugging.
1208 goto out;
1210 if (*str == ',')
1212 * No options but restriction on slabs. This means full
1213 * debugging for slabs matching a pattern.
1215 goto check_slabs;
1217 if (tolower(*str) == 'o') {
1219 * Avoid enabling debugging on caches if its minimum order
1220 * would increase as a result.
1222 disable_higher_order_debug = 1;
1223 goto out;
1226 slub_debug = 0;
1227 if (*str == '-')
1229 * Switch off all debugging measures.
1231 goto out;
1234 * Determine which debug features should be switched on
1236 for (; *str && *str != ','; str++) {
1237 switch (tolower(*str)) {
1238 case 'f':
1239 slub_debug |= SLAB_DEBUG_FREE;
1240 break;
1241 case 'z':
1242 slub_debug |= SLAB_RED_ZONE;
1243 break;
1244 case 'p':
1245 slub_debug |= SLAB_POISON;
1246 break;
1247 case 'u':
1248 slub_debug |= SLAB_STORE_USER;
1249 break;
1250 case 't':
1251 slub_debug |= SLAB_TRACE;
1252 break;
1253 case 'a':
1254 slub_debug |= SLAB_FAILSLAB;
1255 break;
1256 default:
1257 printk(KERN_ERR "slub_debug option '%c' "
1258 "unknown. skipped\n", *str);
1262 check_slabs:
1263 if (*str == ',')
1264 slub_debug_slabs = str + 1;
1265 out:
1266 return 1;
1269 __setup("slub_debug", setup_slub_debug);
1271 static unsigned long kmem_cache_flags(unsigned long objsize,
1272 unsigned long flags, const char *name,
1273 void (*ctor)(void *))
1276 * Enable debugging if selected on the kernel commandline.
1278 if (slub_debug && (!slub_debug_slabs ||
1279 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1280 flags |= slub_debug;
1282 return flags;
1284 #else
1285 static inline void setup_object_debug(struct kmem_cache *s,
1286 struct page *page, void *object) {}
1288 static inline int alloc_debug_processing(struct kmem_cache *s,
1289 struct page *page, void *object, unsigned long addr) { return 0; }
1291 static inline int free_debug_processing(struct kmem_cache *s,
1292 struct page *page, void *object, unsigned long addr) { return 0; }
1294 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1295 { return 1; }
1296 static inline int check_object(struct kmem_cache *s, struct page *page,
1297 void *object, u8 val) { return 1; }
1298 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1299 struct page *page) {}
1300 static inline void remove_full(struct kmem_cache *s, struct page *page) {}
1301 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1302 unsigned long flags, const char *name,
1303 void (*ctor)(void *))
1305 return flags;
1307 #define slub_debug 0
1309 #define disable_higher_order_debug 0
1311 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1312 { return 0; }
1313 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1314 { return 0; }
1315 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1316 int objects) {}
1317 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1318 int objects) {}
1320 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1321 { return 0; }
1323 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1324 void *object) {}
1326 static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1328 #endif /* CONFIG_SLUB_DEBUG */
1331 * Slab allocation and freeing
1333 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1334 struct kmem_cache_order_objects oo)
1336 int order = oo_order(oo);
1338 flags |= __GFP_NOTRACK;
1340 if (node == NUMA_NO_NODE)
1341 return alloc_pages(flags, order);
1342 else
1343 return alloc_pages_exact_node(node, flags, order);
1346 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1348 struct page *page;
1349 struct kmem_cache_order_objects oo = s->oo;
1350 gfp_t alloc_gfp;
1352 flags &= gfp_allowed_mask;
1354 if (flags & __GFP_WAIT)
1355 local_irq_enable();
1357 flags |= s->allocflags;
1360 * Let the initial higher-order allocation fail under memory pressure
1361 * so we fall-back to the minimum order allocation.
1363 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1365 page = alloc_slab_page(alloc_gfp, node, oo);
1366 if (unlikely(!page)) {
1367 oo = s->min;
1369 * Allocation may have failed due to fragmentation.
1370 * Try a lower order alloc if possible
1372 page = alloc_slab_page(flags, node, oo);
1374 if (page)
1375 stat(s, ORDER_FALLBACK);
1378 if (flags & __GFP_WAIT)
1379 local_irq_disable();
1381 if (!page)
1382 return NULL;
1384 if (kmemcheck_enabled
1385 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1386 int pages = 1 << oo_order(oo);
1388 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1391 * Objects from caches that have a constructor don't get
1392 * cleared when they're allocated, so we need to do it here.
1394 if (s->ctor)
1395 kmemcheck_mark_uninitialized_pages(page, pages);
1396 else
1397 kmemcheck_mark_unallocated_pages(page, pages);
1400 page->objects = oo_objects(oo);
1401 mod_zone_page_state(page_zone(page),
1402 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1403 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1404 1 << oo_order(oo));
1406 return page;
1409 static void setup_object(struct kmem_cache *s, struct page *page,
1410 void *object)
1412 setup_object_debug(s, page, object);
1413 if (unlikely(s->ctor))
1414 s->ctor(object);
1417 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1419 struct page *page;
1420 void *start;
1421 void *last;
1422 void *p;
1424 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1426 page = allocate_slab(s,
1427 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1428 if (!page)
1429 goto out;
1431 inc_slabs_node(s, page_to_nid(page), page->objects);
1432 page->slab = s;
1433 page->flags |= 1 << PG_slab;
1435 start = page_address(page);
1437 if (unlikely(s->flags & SLAB_POISON))
1438 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1440 last = start;
1441 for_each_object(p, s, start, page->objects) {
1442 setup_object(s, page, last);
1443 set_freepointer(s, last, p);
1444 last = p;
1446 setup_object(s, page, last);
1447 set_freepointer(s, last, NULL);
1449 page->freelist = start;
1450 page->inuse = 0;
1451 page->frozen = 1;
1452 out:
1453 return page;
1456 static void __free_slab(struct kmem_cache *s, struct page *page)
1458 int order = compound_order(page);
1459 int pages = 1 << order;
1461 if (kmem_cache_debug(s)) {
1462 void *p;
1464 slab_pad_check(s, page);
1465 for_each_object(p, s, page_address(page),
1466 page->objects)
1467 check_object(s, page, p, SLUB_RED_INACTIVE);
1470 kmemcheck_free_shadow(page, compound_order(page));
1472 mod_zone_page_state(page_zone(page),
1473 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1474 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1475 -pages);
1477 __ClearPageSlab(page);
1478 reset_page_mapcount(page);
1479 if (current->reclaim_state)
1480 current->reclaim_state->reclaimed_slab += pages;
1481 __free_pages(page, order);
1484 #define need_reserve_slab_rcu \
1485 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1487 static void rcu_free_slab(struct rcu_head *h)
1489 struct page *page;
1491 if (need_reserve_slab_rcu)
1492 page = virt_to_head_page(h);
1493 else
1494 page = container_of((struct list_head *)h, struct page, lru);
1496 __free_slab(page->slab, page);
1499 static void free_slab(struct kmem_cache *s, struct page *page)
1501 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1502 struct rcu_head *head;
1504 if (need_reserve_slab_rcu) {
1505 int order = compound_order(page);
1506 int offset = (PAGE_SIZE << order) - s->reserved;
1508 VM_BUG_ON(s->reserved != sizeof(*head));
1509 head = page_address(page) + offset;
1510 } else {
1512 * RCU free overloads the RCU head over the LRU
1514 head = (void *)&page->lru;
1517 call_rcu(head, rcu_free_slab);
1518 } else
1519 __free_slab(s, page);
1522 static void discard_slab(struct kmem_cache *s, struct page *page)
1524 dec_slabs_node(s, page_to_nid(page), page->objects);
1525 free_slab(s, page);
1529 * Management of partially allocated slabs.
1531 * list_lock must be held.
1533 static inline void add_partial(struct kmem_cache_node *n,
1534 struct page *page, int tail)
1536 n->nr_partial++;
1537 if (tail)
1538 list_add_tail(&page->lru, &n->partial);
1539 else
1540 list_add(&page->lru, &n->partial);
1544 * list_lock must be held.
1546 static inline void remove_partial(struct kmem_cache_node *n,
1547 struct page *page)
1549 list_del(&page->lru);
1550 n->nr_partial--;
1554 * Lock slab, remove from the partial list and put the object into the
1555 * per cpu freelist.
1557 * Must hold list_lock.
1559 static inline int acquire_slab(struct kmem_cache *s,
1560 struct kmem_cache_node *n, struct page *page)
1562 void *freelist;
1563 unsigned long counters;
1564 struct page new;
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 do {
1572 freelist = page->freelist;
1573 counters = page->counters;
1574 new.counters = counters;
1575 new.inuse = page->objects;
1577 VM_BUG_ON(new.frozen);
1578 new.frozen = 1;
1580 } while (!__cmpxchg_double_slab(s, page,
1581 freelist, counters,
1582 NULL, new.counters,
1583 "lock and freeze"));
1585 remove_partial(n, page);
1587 if (freelist) {
1588 /* Populate the per cpu freelist */
1589 this_cpu_write(s->cpu_slab->freelist, freelist);
1590 this_cpu_write(s->cpu_slab->page, page);
1591 this_cpu_write(s->cpu_slab->node, page_to_nid(page));
1592 return 1;
1593 } else {
1595 * Slab page came from the wrong list. No object to allocate
1596 * from. Put it onto the correct list and continue partial
1597 * scan.
1599 printk(KERN_ERR "SLUB: %s : Page without available objects on"
1600 " partial list\n", s->name);
1601 return 0;
1606 * Try to allocate a partial slab from a specific node.
1608 static struct page *get_partial_node(struct kmem_cache *s,
1609 struct kmem_cache_node *n)
1611 struct page *page;
1614 * Racy check. If we mistakenly see no partial slabs then we
1615 * just allocate an empty slab. If we mistakenly try to get a
1616 * partial slab and there is none available then get_partials()
1617 * will return NULL.
1619 if (!n || !n->nr_partial)
1620 return NULL;
1622 spin_lock(&n->list_lock);
1623 list_for_each_entry(page, &n->partial, lru)
1624 if (acquire_slab(s, n, page))
1625 goto out;
1626 page = NULL;
1627 out:
1628 spin_unlock(&n->list_lock);
1629 return page;
1633 * Get a page from somewhere. Search in increasing NUMA distances.
1635 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1637 #ifdef CONFIG_NUMA
1638 struct zonelist *zonelist;
1639 struct zoneref *z;
1640 struct zone *zone;
1641 enum zone_type high_zoneidx = gfp_zone(flags);
1642 struct page *page;
1645 * The defrag ratio allows a configuration of the tradeoffs between
1646 * inter node defragmentation and node local allocations. A lower
1647 * defrag_ratio increases the tendency to do local allocations
1648 * instead of attempting to obtain partial slabs from other nodes.
1650 * If the defrag_ratio is set to 0 then kmalloc() always
1651 * returns node local objects. If the ratio is higher then kmalloc()
1652 * may return off node objects because partial slabs are obtained
1653 * from other nodes and filled up.
1655 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1656 * defrag_ratio = 1000) then every (well almost) allocation will
1657 * first attempt to defrag slab caches on other nodes. This means
1658 * scanning over all nodes to look for partial slabs which may be
1659 * expensive if we do it every time we are trying to find a slab
1660 * with available objects.
1662 if (!s->remote_node_defrag_ratio ||
1663 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1664 return NULL;
1666 get_mems_allowed();
1667 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1668 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1669 struct kmem_cache_node *n;
1671 n = get_node(s, zone_to_nid(zone));
1673 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1674 n->nr_partial > s->min_partial) {
1675 page = get_partial_node(s, n);
1676 if (page) {
1677 put_mems_allowed();
1678 return page;
1682 put_mems_allowed();
1683 #endif
1684 return NULL;
1688 * Get a partial page, lock it and return it.
1690 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1692 struct page *page;
1693 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1695 page = get_partial_node(s, get_node(s, searchnode));
1696 if (page || node != NUMA_NO_NODE)
1697 return page;
1699 return get_any_partial(s, flags);
1702 #ifdef CONFIG_PREEMPT
1704 * Calculate the next globally unique transaction for disambiguiation
1705 * during cmpxchg. The transactions start with the cpu number and are then
1706 * incremented by CONFIG_NR_CPUS.
1708 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1709 #else
1711 * No preemption supported therefore also no need to check for
1712 * different cpus.
1714 #define TID_STEP 1
1715 #endif
1717 static inline unsigned long next_tid(unsigned long tid)
1719 return tid + TID_STEP;
1722 static inline unsigned int tid_to_cpu(unsigned long tid)
1724 return tid % TID_STEP;
1727 static inline unsigned long tid_to_event(unsigned long tid)
1729 return tid / TID_STEP;
1732 static inline unsigned int init_tid(int cpu)
1734 return cpu;
1737 static inline void note_cmpxchg_failure(const char *n,
1738 const struct kmem_cache *s, unsigned long tid)
1740 #ifdef SLUB_DEBUG_CMPXCHG
1741 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1743 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1745 #ifdef CONFIG_PREEMPT
1746 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1747 printk("due to cpu change %d -> %d\n",
1748 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1749 else
1750 #endif
1751 if (tid_to_event(tid) != tid_to_event(actual_tid))
1752 printk("due to cpu running other code. Event %ld->%ld\n",
1753 tid_to_event(tid), tid_to_event(actual_tid));
1754 else
1755 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1756 actual_tid, tid, next_tid(tid));
1757 #endif
1758 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1761 void init_kmem_cache_cpus(struct kmem_cache *s)
1763 int cpu;
1765 for_each_possible_cpu(cpu)
1766 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1769 * Remove the cpu slab
1773 * Remove the cpu slab
1775 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1777 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1778 struct page *page = c->page;
1779 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1780 int lock = 0;
1781 enum slab_modes l = M_NONE, m = M_NONE;
1782 void *freelist;
1783 void *nextfree;
1784 int tail = 0;
1785 struct page new;
1786 struct page old;
1788 if (page->freelist) {
1789 stat(s, DEACTIVATE_REMOTE_FREES);
1790 tail = 1;
1793 c->tid = next_tid(c->tid);
1794 c->page = NULL;
1795 freelist = c->freelist;
1796 c->freelist = NULL;
1799 * Stage one: Free all available per cpu objects back
1800 * to the page freelist while it is still frozen. Leave the
1801 * last one.
1803 * There is no need to take the list->lock because the page
1804 * is still frozen.
1806 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1807 void *prior;
1808 unsigned long counters;
1810 do {
1811 prior = page->freelist;
1812 counters = page->counters;
1813 set_freepointer(s, freelist, prior);
1814 new.counters = counters;
1815 new.inuse--;
1816 VM_BUG_ON(!new.frozen);
1818 } while (!__cmpxchg_double_slab(s, page,
1819 prior, counters,
1820 freelist, new.counters,
1821 "drain percpu freelist"));
1823 freelist = nextfree;
1827 * Stage two: Ensure that the page is unfrozen while the
1828 * list presence reflects the actual number of objects
1829 * during unfreeze.
1831 * We setup the list membership and then perform a cmpxchg
1832 * with the count. If there is a mismatch then the page
1833 * is not unfrozen but the page is on the wrong list.
1835 * Then we restart the process which may have to remove
1836 * the page from the list that we just put it on again
1837 * because the number of objects in the slab may have
1838 * changed.
1840 redo:
1842 old.freelist = page->freelist;
1843 old.counters = page->counters;
1844 VM_BUG_ON(!old.frozen);
1846 /* Determine target state of the slab */
1847 new.counters = old.counters;
1848 if (freelist) {
1849 new.inuse--;
1850 set_freepointer(s, freelist, old.freelist);
1851 new.freelist = freelist;
1852 } else
1853 new.freelist = old.freelist;
1855 new.frozen = 0;
1857 if (!new.inuse && n->nr_partial > s->min_partial)
1858 m = M_FREE;
1859 else if (new.freelist) {
1860 m = M_PARTIAL;
1861 if (!lock) {
1862 lock = 1;
1864 * Taking the spinlock removes the possiblity
1865 * that acquire_slab() will see a slab page that
1866 * is frozen
1868 spin_lock(&n->list_lock);
1870 } else {
1871 m = M_FULL;
1872 if (kmem_cache_debug(s) && !lock) {
1873 lock = 1;
1875 * This also ensures that the scanning of full
1876 * slabs from diagnostic functions will not see
1877 * any frozen slabs.
1879 spin_lock(&n->list_lock);
1883 if (l != m) {
1885 if (l == M_PARTIAL)
1887 remove_partial(n, page);
1889 else if (l == M_FULL)
1891 remove_full(s, page);
1893 if (m == M_PARTIAL) {
1895 add_partial(n, page, tail);
1896 stat(s, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1898 } else if (m == M_FULL) {
1900 stat(s, DEACTIVATE_FULL);
1901 add_full(s, n, page);
1906 l = m;
1907 if (!__cmpxchg_double_slab(s, page,
1908 old.freelist, old.counters,
1909 new.freelist, new.counters,
1910 "unfreezing slab"))
1911 goto redo;
1913 if (lock)
1914 spin_unlock(&n->list_lock);
1916 if (m == M_FREE) {
1917 stat(s, DEACTIVATE_EMPTY);
1918 discard_slab(s, page);
1919 stat(s, FREE_SLAB);
1923 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1925 stat(s, CPUSLAB_FLUSH);
1926 deactivate_slab(s, c);
1930 * Flush cpu slab.
1932 * Called from IPI handler with interrupts disabled.
1934 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1936 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
1938 if (likely(c && c->page))
1939 flush_slab(s, c);
1942 static void flush_cpu_slab(void *d)
1944 struct kmem_cache *s = d;
1946 __flush_cpu_slab(s, smp_processor_id());
1949 static void flush_all(struct kmem_cache *s)
1951 on_each_cpu(flush_cpu_slab, s, 1);
1955 * Check if the objects in a per cpu structure fit numa
1956 * locality expectations.
1958 static inline int node_match(struct kmem_cache_cpu *c, int node)
1960 #ifdef CONFIG_NUMA
1961 if (node != NUMA_NO_NODE && c->node != node)
1962 return 0;
1963 #endif
1964 return 1;
1967 static int count_free(struct page *page)
1969 return page->objects - page->inuse;
1972 static unsigned long count_partial(struct kmem_cache_node *n,
1973 int (*get_count)(struct page *))
1975 unsigned long flags;
1976 unsigned long x = 0;
1977 struct page *page;
1979 spin_lock_irqsave(&n->list_lock, flags);
1980 list_for_each_entry(page, &n->partial, lru)
1981 x += get_count(page);
1982 spin_unlock_irqrestore(&n->list_lock, flags);
1983 return x;
1986 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
1988 #ifdef CONFIG_SLUB_DEBUG
1989 return atomic_long_read(&n->total_objects);
1990 #else
1991 return 0;
1992 #endif
1995 static noinline void
1996 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
1998 int node;
2000 printk(KERN_WARNING
2001 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2002 nid, gfpflags);
2003 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
2004 "default order: %d, min order: %d\n", s->name, s->objsize,
2005 s->size, oo_order(s->oo), oo_order(s->min));
2007 if (oo_order(s->min) > get_order(s->objsize))
2008 printk(KERN_WARNING " %s debugging increased min order, use "
2009 "slub_debug=O to disable.\n", s->name);
2011 for_each_online_node(node) {
2012 struct kmem_cache_node *n = get_node(s, node);
2013 unsigned long nr_slabs;
2014 unsigned long nr_objs;
2015 unsigned long nr_free;
2017 if (!n)
2018 continue;
2020 nr_free = count_partial(n, count_free);
2021 nr_slabs = node_nr_slabs(n);
2022 nr_objs = node_nr_objs(n);
2024 printk(KERN_WARNING
2025 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2026 node, nr_slabs, nr_objs, nr_free);
2031 * Slow path. The lockless freelist is empty or we need to perform
2032 * debugging duties.
2034 * Interrupts are disabled.
2036 * Processing is still very fast if new objects have been freed to the
2037 * regular freelist. In that case we simply take over the regular freelist
2038 * as the lockless freelist and zap the regular freelist.
2040 * If that is not working then we fall back to the partial lists. We take the
2041 * first element of the freelist as the object to allocate now and move the
2042 * rest of the freelist to the lockless freelist.
2044 * And if we were unable to get a new slab from the partial slab lists then
2045 * we need to allocate a new slab. This is the slowest path since it involves
2046 * a call to the page allocator and the setup of a new slab.
2048 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2049 unsigned long addr, struct kmem_cache_cpu *c)
2051 void **object;
2052 struct page *page;
2053 unsigned long flags;
2054 struct page new;
2055 unsigned long counters;
2057 local_irq_save(flags);
2058 #ifdef CONFIG_PREEMPT
2060 * We may have been preempted and rescheduled on a different
2061 * cpu before disabling interrupts. Need to reload cpu area
2062 * pointer.
2064 c = this_cpu_ptr(s->cpu_slab);
2065 #endif
2067 /* We handle __GFP_ZERO in the caller */
2068 gfpflags &= ~__GFP_ZERO;
2070 page = c->page;
2071 if (!page)
2072 goto new_slab;
2074 if (unlikely(!node_match(c, node))) {
2075 stat(s, ALLOC_NODE_MISMATCH);
2076 deactivate_slab(s, c);
2077 goto new_slab;
2080 stat(s, ALLOC_SLOWPATH);
2082 do {
2083 object = page->freelist;
2084 counters = page->counters;
2085 new.counters = counters;
2086 VM_BUG_ON(!new.frozen);
2089 * If there is no object left then we use this loop to
2090 * deactivate the slab which is simple since no objects
2091 * are left in the slab and therefore we do not need to
2092 * put the page back onto the partial list.
2094 * If there are objects left then we retrieve them
2095 * and use them to refill the per cpu queue.
2098 new.inuse = page->objects;
2099 new.frozen = object != NULL;
2101 } while (!__cmpxchg_double_slab(s, page,
2102 object, counters,
2103 NULL, new.counters,
2104 "__slab_alloc"));
2106 if (unlikely(!object)) {
2107 c->page = NULL;
2108 stat(s, DEACTIVATE_BYPASS);
2109 goto new_slab;
2112 stat(s, ALLOC_REFILL);
2114 load_freelist:
2115 VM_BUG_ON(!page->frozen);
2116 c->freelist = get_freepointer(s, object);
2117 c->tid = next_tid(c->tid);
2118 local_irq_restore(flags);
2119 return object;
2121 new_slab:
2122 page = get_partial(s, gfpflags, node);
2123 if (page) {
2124 stat(s, ALLOC_FROM_PARTIAL);
2125 object = c->freelist;
2127 if (kmem_cache_debug(s))
2128 goto debug;
2129 goto load_freelist;
2132 page = new_slab(s, gfpflags, node);
2134 if (page) {
2135 c = __this_cpu_ptr(s->cpu_slab);
2136 if (c->page)
2137 flush_slab(s, c);
2140 * No other reference to the page yet so we can
2141 * muck around with it freely without cmpxchg
2143 object = page->freelist;
2144 page->freelist = NULL;
2145 page->inuse = page->objects;
2147 stat(s, ALLOC_SLAB);
2148 c->node = page_to_nid(page);
2149 c->page = page;
2151 if (kmem_cache_debug(s))
2152 goto debug;
2153 goto load_freelist;
2155 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2156 slab_out_of_memory(s, gfpflags, node);
2157 local_irq_restore(flags);
2158 return NULL;
2160 debug:
2161 if (!object || !alloc_debug_processing(s, page, object, addr))
2162 goto new_slab;
2164 c->freelist = get_freepointer(s, object);
2165 deactivate_slab(s, c);
2166 c->page = NULL;
2167 c->node = NUMA_NO_NODE;
2168 local_irq_restore(flags);
2169 return object;
2173 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2174 * have the fastpath folded into their functions. So no function call
2175 * overhead for requests that can be satisfied on the fastpath.
2177 * The fastpath works by first checking if the lockless freelist can be used.
2178 * If not then __slab_alloc is called for slow processing.
2180 * Otherwise we can simply pick the next object from the lockless free list.
2182 static __always_inline void *slab_alloc(struct kmem_cache *s,
2183 gfp_t gfpflags, int node, unsigned long addr)
2185 void **object;
2186 struct kmem_cache_cpu *c;
2187 unsigned long tid;
2189 if (slab_pre_alloc_hook(s, gfpflags))
2190 return NULL;
2192 redo:
2195 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2196 * enabled. We may switch back and forth between cpus while
2197 * reading from one cpu area. That does not matter as long
2198 * as we end up on the original cpu again when doing the cmpxchg.
2200 c = __this_cpu_ptr(s->cpu_slab);
2203 * The transaction ids are globally unique per cpu and per operation on
2204 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2205 * occurs on the right processor and that there was no operation on the
2206 * linked list in between.
2208 tid = c->tid;
2209 barrier();
2211 object = c->freelist;
2212 if (unlikely(!object || !node_match(c, node)))
2214 object = __slab_alloc(s, gfpflags, node, addr, c);
2216 else {
2218 * The cmpxchg will only match if there was no additional
2219 * operation and if we are on the right processor.
2221 * The cmpxchg does the following atomically (without lock semantics!)
2222 * 1. Relocate first pointer to the current per cpu area.
2223 * 2. Verify that tid and freelist have not been changed
2224 * 3. If they were not changed replace tid and freelist
2226 * Since this is without lock semantics the protection is only against
2227 * code executing on this cpu *not* from access by other cpus.
2229 if (unlikely(!irqsafe_cpu_cmpxchg_double(
2230 s->cpu_slab->freelist, s->cpu_slab->tid,
2231 object, tid,
2232 get_freepointer_safe(s, object), next_tid(tid)))) {
2234 note_cmpxchg_failure("slab_alloc", s, tid);
2235 goto redo;
2237 stat(s, ALLOC_FASTPATH);
2240 if (unlikely(gfpflags & __GFP_ZERO) && object)
2241 memset(object, 0, s->objsize);
2243 slab_post_alloc_hook(s, gfpflags, object);
2245 return object;
2248 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2250 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2252 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
2254 return ret;
2256 EXPORT_SYMBOL(kmem_cache_alloc);
2258 #ifdef CONFIG_TRACING
2259 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2261 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2262 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2263 return ret;
2265 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2267 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2269 void *ret = kmalloc_order(size, flags, order);
2270 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2271 return ret;
2273 EXPORT_SYMBOL(kmalloc_order_trace);
2274 #endif
2276 #ifdef CONFIG_NUMA
2277 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2279 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2281 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2282 s->objsize, s->size, gfpflags, node);
2284 return ret;
2286 EXPORT_SYMBOL(kmem_cache_alloc_node);
2288 #ifdef CONFIG_TRACING
2289 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2290 gfp_t gfpflags,
2291 int node, size_t size)
2293 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2295 trace_kmalloc_node(_RET_IP_, ret,
2296 size, s->size, gfpflags, node);
2297 return ret;
2299 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2300 #endif
2301 #endif
2304 * Slow patch handling. This may still be called frequently since objects
2305 * have a longer lifetime than the cpu slabs in most processing loads.
2307 * So we still attempt to reduce cache line usage. Just take the slab
2308 * lock and free the item. If there is no additional partial page
2309 * handling required then we can return immediately.
2311 static void __slab_free(struct kmem_cache *s, struct page *page,
2312 void *x, unsigned long addr)
2314 void *prior;
2315 void **object = (void *)x;
2316 int was_frozen;
2317 int inuse;
2318 struct page new;
2319 unsigned long counters;
2320 struct kmem_cache_node *n = NULL;
2321 unsigned long uninitialized_var(flags);
2323 stat(s, FREE_SLOWPATH);
2325 if (kmem_cache_debug(s) && !free_debug_processing(s, page, x, addr))
2326 return;
2328 do {
2329 prior = page->freelist;
2330 counters = page->counters;
2331 set_freepointer(s, object, prior);
2332 new.counters = counters;
2333 was_frozen = new.frozen;
2334 new.inuse--;
2335 if ((!new.inuse || !prior) && !was_frozen && !n) {
2336 n = get_node(s, page_to_nid(page));
2338 * Speculatively acquire the list_lock.
2339 * If the cmpxchg does not succeed then we may
2340 * drop the list_lock without any processing.
2342 * Otherwise the list_lock will synchronize with
2343 * other processors updating the list of slabs.
2345 spin_lock_irqsave(&n->list_lock, flags);
2347 inuse = new.inuse;
2349 } while (!cmpxchg_double_slab(s, page,
2350 prior, counters,
2351 object, new.counters,
2352 "__slab_free"));
2354 if (likely(!n)) {
2356 * The list lock was not taken therefore no list
2357 * activity can be necessary.
2359 if (was_frozen)
2360 stat(s, FREE_FROZEN);
2361 return;
2365 * was_frozen may have been set after we acquired the list_lock in
2366 * an earlier loop. So we need to check it here again.
2368 if (was_frozen)
2369 stat(s, FREE_FROZEN);
2370 else {
2371 if (unlikely(!inuse && n->nr_partial > s->min_partial))
2372 goto slab_empty;
2375 * Objects left in the slab. If it was not on the partial list before
2376 * then add it.
2378 if (unlikely(!prior)) {
2379 remove_full(s, page);
2380 add_partial(n, page, 1);
2381 stat(s, FREE_ADD_PARTIAL);
2384 spin_unlock_irqrestore(&n->list_lock, flags);
2385 return;
2387 slab_empty:
2388 if (prior) {
2390 * Slab on the partial list.
2392 remove_partial(n, page);
2393 stat(s, FREE_REMOVE_PARTIAL);
2394 } else
2395 /* Slab must be on the full list */
2396 remove_full(s, page);
2398 spin_unlock_irqrestore(&n->list_lock, flags);
2399 stat(s, FREE_SLAB);
2400 discard_slab(s, page);
2404 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2405 * can perform fastpath freeing without additional function calls.
2407 * The fastpath is only possible if we are freeing to the current cpu slab
2408 * of this processor. This typically the case if we have just allocated
2409 * the item before.
2411 * If fastpath is not possible then fall back to __slab_free where we deal
2412 * with all sorts of special processing.
2414 static __always_inline void slab_free(struct kmem_cache *s,
2415 struct page *page, void *x, unsigned long addr)
2417 void **object = (void *)x;
2418 struct kmem_cache_cpu *c;
2419 unsigned long tid;
2421 slab_free_hook(s, x);
2423 redo:
2426 * Determine the currently cpus per cpu slab.
2427 * The cpu may change afterward. However that does not matter since
2428 * data is retrieved via this pointer. If we are on the same cpu
2429 * during the cmpxchg then the free will succedd.
2431 c = __this_cpu_ptr(s->cpu_slab);
2433 tid = c->tid;
2434 barrier();
2436 if (likely(page == c->page)) {
2437 set_freepointer(s, object, c->freelist);
2439 if (unlikely(!irqsafe_cpu_cmpxchg_double(
2440 s->cpu_slab->freelist, s->cpu_slab->tid,
2441 c->freelist, tid,
2442 object, next_tid(tid)))) {
2444 note_cmpxchg_failure("slab_free", s, tid);
2445 goto redo;
2447 stat(s, FREE_FASTPATH);
2448 } else
2449 __slab_free(s, page, x, addr);
2453 void kmem_cache_free(struct kmem_cache *s, void *x)
2455 struct page *page;
2457 page = virt_to_head_page(x);
2459 slab_free(s, page, x, _RET_IP_);
2461 trace_kmem_cache_free(_RET_IP_, x);
2463 EXPORT_SYMBOL(kmem_cache_free);
2466 * Object placement in a slab is made very easy because we always start at
2467 * offset 0. If we tune the size of the object to the alignment then we can
2468 * get the required alignment by putting one properly sized object after
2469 * another.
2471 * Notice that the allocation order determines the sizes of the per cpu
2472 * caches. Each processor has always one slab available for allocations.
2473 * Increasing the allocation order reduces the number of times that slabs
2474 * must be moved on and off the partial lists and is therefore a factor in
2475 * locking overhead.
2479 * Mininum / Maximum order of slab pages. This influences locking overhead
2480 * and slab fragmentation. A higher order reduces the number of partial slabs
2481 * and increases the number of allocations possible without having to
2482 * take the list_lock.
2484 static int slub_min_order;
2485 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2486 static int slub_min_objects;
2489 * Merge control. If this is set then no merging of slab caches will occur.
2490 * (Could be removed. This was introduced to pacify the merge skeptics.)
2492 static int slub_nomerge;
2495 * Calculate the order of allocation given an slab object size.
2497 * The order of allocation has significant impact on performance and other
2498 * system components. Generally order 0 allocations should be preferred since
2499 * order 0 does not cause fragmentation in the page allocator. Larger objects
2500 * be problematic to put into order 0 slabs because there may be too much
2501 * unused space left. We go to a higher order if more than 1/16th of the slab
2502 * would be wasted.
2504 * In order to reach satisfactory performance we must ensure that a minimum
2505 * number of objects is in one slab. Otherwise we may generate too much
2506 * activity on the partial lists which requires taking the list_lock. This is
2507 * less a concern for large slabs though which are rarely used.
2509 * slub_max_order specifies the order where we begin to stop considering the
2510 * number of objects in a slab as critical. If we reach slub_max_order then
2511 * we try to keep the page order as low as possible. So we accept more waste
2512 * of space in favor of a small page order.
2514 * Higher order allocations also allow the placement of more objects in a
2515 * slab and thereby reduce object handling overhead. If the user has
2516 * requested a higher mininum order then we start with that one instead of
2517 * the smallest order which will fit the object.
2519 static inline int slab_order(int size, int min_objects,
2520 int max_order, int fract_leftover, int reserved)
2522 int order;
2523 int rem;
2524 int min_order = slub_min_order;
2526 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2527 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2529 for (order = max(min_order,
2530 fls(min_objects * size - 1) - PAGE_SHIFT);
2531 order <= max_order; order++) {
2533 unsigned long slab_size = PAGE_SIZE << order;
2535 if (slab_size < min_objects * size + reserved)
2536 continue;
2538 rem = (slab_size - reserved) % size;
2540 if (rem <= slab_size / fract_leftover)
2541 break;
2545 return order;
2548 static inline int calculate_order(int size, int reserved)
2550 int order;
2551 int min_objects;
2552 int fraction;
2553 int max_objects;
2556 * Attempt to find best configuration for a slab. This
2557 * works by first attempting to generate a layout with
2558 * the best configuration and backing off gradually.
2560 * First we reduce the acceptable waste in a slab. Then
2561 * we reduce the minimum objects required in a slab.
2563 min_objects = slub_min_objects;
2564 if (!min_objects)
2565 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2566 max_objects = order_objects(slub_max_order, size, reserved);
2567 min_objects = min(min_objects, max_objects);
2569 while (min_objects > 1) {
2570 fraction = 16;
2571 while (fraction >= 4) {
2572 order = slab_order(size, min_objects,
2573 slub_max_order, fraction, reserved);
2574 if (order <= slub_max_order)
2575 return order;
2576 fraction /= 2;
2578 min_objects--;
2582 * We were unable to place multiple objects in a slab. Now
2583 * lets see if we can place a single object there.
2585 order = slab_order(size, 1, slub_max_order, 1, reserved);
2586 if (order <= slub_max_order)
2587 return order;
2590 * Doh this slab cannot be placed using slub_max_order.
2592 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2593 if (order < MAX_ORDER)
2594 return order;
2595 return -ENOSYS;
2599 * Figure out what the alignment of the objects will be.
2601 static unsigned long calculate_alignment(unsigned long flags,
2602 unsigned long align, unsigned long size)
2605 * If the user wants hardware cache aligned objects then follow that
2606 * suggestion if the object is sufficiently large.
2608 * The hardware cache alignment cannot override the specified
2609 * alignment though. If that is greater then use it.
2611 if (flags & SLAB_HWCACHE_ALIGN) {
2612 unsigned long ralign = cache_line_size();
2613 while (size <= ralign / 2)
2614 ralign /= 2;
2615 align = max(align, ralign);
2618 if (align < ARCH_SLAB_MINALIGN)
2619 align = ARCH_SLAB_MINALIGN;
2621 return ALIGN(align, sizeof(void *));
2624 static void
2625 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
2627 n->nr_partial = 0;
2628 spin_lock_init(&n->list_lock);
2629 INIT_LIST_HEAD(&n->partial);
2630 #ifdef CONFIG_SLUB_DEBUG
2631 atomic_long_set(&n->nr_slabs, 0);
2632 atomic_long_set(&n->total_objects, 0);
2633 INIT_LIST_HEAD(&n->full);
2634 #endif
2637 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2639 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2640 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2643 * Must align to double word boundary for the double cmpxchg
2644 * instructions to work; see __pcpu_double_call_return_bool().
2646 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2647 2 * sizeof(void *));
2649 if (!s->cpu_slab)
2650 return 0;
2652 init_kmem_cache_cpus(s);
2654 return 1;
2657 static struct kmem_cache *kmem_cache_node;
2660 * No kmalloc_node yet so do it by hand. We know that this is the first
2661 * slab on the node for this slabcache. There are no concurrent accesses
2662 * possible.
2664 * Note that this function only works on the kmalloc_node_cache
2665 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2666 * memory on a fresh node that has no slab structures yet.
2668 static void early_kmem_cache_node_alloc(int node)
2670 struct page *page;
2671 struct kmem_cache_node *n;
2673 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2675 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2677 BUG_ON(!page);
2678 if (page_to_nid(page) != node) {
2679 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2680 "node %d\n", node);
2681 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2682 "in order to be able to continue\n");
2685 n = page->freelist;
2686 BUG_ON(!n);
2687 page->freelist = get_freepointer(kmem_cache_node, n);
2688 page->inuse++;
2689 page->frozen = 0;
2690 kmem_cache_node->node[node] = n;
2691 #ifdef CONFIG_SLUB_DEBUG
2692 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2693 init_tracking(kmem_cache_node, n);
2694 #endif
2695 init_kmem_cache_node(n, kmem_cache_node);
2696 inc_slabs_node(kmem_cache_node, node, page->objects);
2698 add_partial(n, page, 0);
2701 static void free_kmem_cache_nodes(struct kmem_cache *s)
2703 int node;
2705 for_each_node_state(node, N_NORMAL_MEMORY) {
2706 struct kmem_cache_node *n = s->node[node];
2708 if (n)
2709 kmem_cache_free(kmem_cache_node, n);
2711 s->node[node] = NULL;
2715 static int init_kmem_cache_nodes(struct kmem_cache *s)
2717 int node;
2719 for_each_node_state(node, N_NORMAL_MEMORY) {
2720 struct kmem_cache_node *n;
2722 if (slab_state == DOWN) {
2723 early_kmem_cache_node_alloc(node);
2724 continue;
2726 n = kmem_cache_alloc_node(kmem_cache_node,
2727 GFP_KERNEL, node);
2729 if (!n) {
2730 free_kmem_cache_nodes(s);
2731 return 0;
2734 s->node[node] = n;
2735 init_kmem_cache_node(n, s);
2737 return 1;
2740 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2742 if (min < MIN_PARTIAL)
2743 min = MIN_PARTIAL;
2744 else if (min > MAX_PARTIAL)
2745 min = MAX_PARTIAL;
2746 s->min_partial = min;
2750 * calculate_sizes() determines the order and the distribution of data within
2751 * a slab object.
2753 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2755 unsigned long flags = s->flags;
2756 unsigned long size = s->objsize;
2757 unsigned long align = s->align;
2758 int order;
2761 * Round up object size to the next word boundary. We can only
2762 * place the free pointer at word boundaries and this determines
2763 * the possible location of the free pointer.
2765 size = ALIGN(size, sizeof(void *));
2767 #ifdef CONFIG_SLUB_DEBUG
2769 * Determine if we can poison the object itself. If the user of
2770 * the slab may touch the object after free or before allocation
2771 * then we should never poison the object itself.
2773 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2774 !s->ctor)
2775 s->flags |= __OBJECT_POISON;
2776 else
2777 s->flags &= ~__OBJECT_POISON;
2781 * If we are Redzoning then check if there is some space between the
2782 * end of the object and the free pointer. If not then add an
2783 * additional word to have some bytes to store Redzone information.
2785 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2786 size += sizeof(void *);
2787 #endif
2790 * With that we have determined the number of bytes in actual use
2791 * by the object. This is the potential offset to the free pointer.
2793 s->inuse = size;
2795 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2796 s->ctor)) {
2798 * Relocate free pointer after the object if it is not
2799 * permitted to overwrite the first word of the object on
2800 * kmem_cache_free.
2802 * This is the case if we do RCU, have a constructor or
2803 * destructor or are poisoning the objects.
2805 s->offset = size;
2806 size += sizeof(void *);
2809 #ifdef CONFIG_SLUB_DEBUG
2810 if (flags & SLAB_STORE_USER)
2812 * Need to store information about allocs and frees after
2813 * the object.
2815 size += 2 * sizeof(struct track);
2817 if (flags & SLAB_RED_ZONE)
2819 * Add some empty padding so that we can catch
2820 * overwrites from earlier objects rather than let
2821 * tracking information or the free pointer be
2822 * corrupted if a user writes before the start
2823 * of the object.
2825 size += sizeof(void *);
2826 #endif
2829 * Determine the alignment based on various parameters that the
2830 * user specified and the dynamic determination of cache line size
2831 * on bootup.
2833 align = calculate_alignment(flags, align, s->objsize);
2834 s->align = align;
2837 * SLUB stores one object immediately after another beginning from
2838 * offset 0. In order to align the objects we have to simply size
2839 * each object to conform to the alignment.
2841 size = ALIGN(size, align);
2842 s->size = size;
2843 if (forced_order >= 0)
2844 order = forced_order;
2845 else
2846 order = calculate_order(size, s->reserved);
2848 if (order < 0)
2849 return 0;
2851 s->allocflags = 0;
2852 if (order)
2853 s->allocflags |= __GFP_COMP;
2855 if (s->flags & SLAB_CACHE_DMA)
2856 s->allocflags |= SLUB_DMA;
2858 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2859 s->allocflags |= __GFP_RECLAIMABLE;
2862 * Determine the number of objects per slab
2864 s->oo = oo_make(order, size, s->reserved);
2865 s->min = oo_make(get_order(size), size, s->reserved);
2866 if (oo_objects(s->oo) > oo_objects(s->max))
2867 s->max = s->oo;
2869 return !!oo_objects(s->oo);
2873 static int kmem_cache_open(struct kmem_cache *s,
2874 const char *name, size_t size,
2875 size_t align, unsigned long flags,
2876 void (*ctor)(void *))
2878 memset(s, 0, kmem_size);
2879 s->name = name;
2880 s->ctor = ctor;
2881 s->objsize = size;
2882 s->align = align;
2883 s->flags = kmem_cache_flags(size, flags, name, ctor);
2884 s->reserved = 0;
2886 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
2887 s->reserved = sizeof(struct rcu_head);
2889 if (!calculate_sizes(s, -1))
2890 goto error;
2891 if (disable_higher_order_debug) {
2893 * Disable debugging flags that store metadata if the min slab
2894 * order increased.
2896 if (get_order(s->size) > get_order(s->objsize)) {
2897 s->flags &= ~DEBUG_METADATA_FLAGS;
2898 s->offset = 0;
2899 if (!calculate_sizes(s, -1))
2900 goto error;
2904 #ifdef CONFIG_CMPXCHG_DOUBLE
2905 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
2906 /* Enable fast mode */
2907 s->flags |= __CMPXCHG_DOUBLE;
2908 #endif
2911 * The larger the object size is, the more pages we want on the partial
2912 * list to avoid pounding the page allocator excessively.
2914 set_min_partial(s, ilog2(s->size));
2915 s->refcount = 1;
2916 #ifdef CONFIG_NUMA
2917 s->remote_node_defrag_ratio = 1000;
2918 #endif
2919 if (!init_kmem_cache_nodes(s))
2920 goto error;
2922 if (alloc_kmem_cache_cpus(s))
2923 return 1;
2925 free_kmem_cache_nodes(s);
2926 error:
2927 if (flags & SLAB_PANIC)
2928 panic("Cannot create slab %s size=%lu realsize=%u "
2929 "order=%u offset=%u flags=%lx\n",
2930 s->name, (unsigned long)size, s->size, oo_order(s->oo),
2931 s->offset, flags);
2932 return 0;
2936 * Determine the size of a slab object
2938 unsigned int kmem_cache_size(struct kmem_cache *s)
2940 return s->objsize;
2942 EXPORT_SYMBOL(kmem_cache_size);
2944 static void list_slab_objects(struct kmem_cache *s, struct page *page,
2945 const char *text)
2947 #ifdef CONFIG_SLUB_DEBUG
2948 void *addr = page_address(page);
2949 void *p;
2950 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
2951 sizeof(long), GFP_ATOMIC);
2952 if (!map)
2953 return;
2954 slab_err(s, page, "%s", text);
2955 slab_lock(page);
2957 get_map(s, page, map);
2958 for_each_object(p, s, addr, page->objects) {
2960 if (!test_bit(slab_index(p, s, addr), map)) {
2961 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2962 p, p - addr);
2963 print_tracking(s, p);
2966 slab_unlock(page);
2967 kfree(map);
2968 #endif
2972 * Attempt to free all partial slabs on a node.
2974 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2976 unsigned long flags;
2977 struct page *page, *h;
2979 spin_lock_irqsave(&n->list_lock, flags);
2980 list_for_each_entry_safe(page, h, &n->partial, lru) {
2981 if (!page->inuse) {
2982 remove_partial(n, page);
2983 discard_slab(s, page);
2984 } else {
2985 list_slab_objects(s, page,
2986 "Objects remaining on kmem_cache_close()");
2989 spin_unlock_irqrestore(&n->list_lock, flags);
2993 * Release all resources used by a slab cache.
2995 static inline int kmem_cache_close(struct kmem_cache *s)
2997 int node;
2999 flush_all(s);
3000 free_percpu(s->cpu_slab);
3001 /* Attempt to free all objects */
3002 for_each_node_state(node, N_NORMAL_MEMORY) {
3003 struct kmem_cache_node *n = get_node(s, node);
3005 free_partial(s, n);
3006 if (n->nr_partial || slabs_node(s, node))
3007 return 1;
3009 free_kmem_cache_nodes(s);
3010 return 0;
3014 * Close a cache and release the kmem_cache structure
3015 * (must be used for caches created using kmem_cache_create)
3017 void kmem_cache_destroy(struct kmem_cache *s)
3019 down_write(&slub_lock);
3020 s->refcount--;
3021 if (!s->refcount) {
3022 list_del(&s->list);
3023 if (kmem_cache_close(s)) {
3024 printk(KERN_ERR "SLUB %s: %s called for cache that "
3025 "still has objects.\n", s->name, __func__);
3026 dump_stack();
3028 if (s->flags & SLAB_DESTROY_BY_RCU)
3029 rcu_barrier();
3030 sysfs_slab_remove(s);
3032 up_write(&slub_lock);
3034 EXPORT_SYMBOL(kmem_cache_destroy);
3036 /********************************************************************
3037 * Kmalloc subsystem
3038 *******************************************************************/
3040 struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
3041 EXPORT_SYMBOL(kmalloc_caches);
3043 static struct kmem_cache *kmem_cache;
3045 #ifdef CONFIG_ZONE_DMA
3046 static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
3047 #endif
3049 static int __init setup_slub_min_order(char *str)
3051 get_option(&str, &slub_min_order);
3053 return 1;
3056 __setup("slub_min_order=", setup_slub_min_order);
3058 static int __init setup_slub_max_order(char *str)
3060 get_option(&str, &slub_max_order);
3061 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3063 return 1;
3066 __setup("slub_max_order=", setup_slub_max_order);
3068 static int __init setup_slub_min_objects(char *str)
3070 get_option(&str, &slub_min_objects);
3072 return 1;
3075 __setup("slub_min_objects=", setup_slub_min_objects);
3077 static int __init setup_slub_nomerge(char *str)
3079 slub_nomerge = 1;
3080 return 1;
3083 __setup("slub_nomerge", setup_slub_nomerge);
3085 static struct kmem_cache *__init create_kmalloc_cache(const char *name,
3086 int size, unsigned int flags)
3088 struct kmem_cache *s;
3090 s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3093 * This function is called with IRQs disabled during early-boot on
3094 * single CPU so there's no need to take slub_lock here.
3096 if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN,
3097 flags, NULL))
3098 goto panic;
3100 list_add(&s->list, &slab_caches);
3101 return s;
3103 panic:
3104 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
3105 return NULL;
3109 * Conversion table for small slabs sizes / 8 to the index in the
3110 * kmalloc array. This is necessary for slabs < 192 since we have non power
3111 * of two cache sizes there. The size of larger slabs can be determined using
3112 * fls.
3114 static s8 size_index[24] = {
3115 3, /* 8 */
3116 4, /* 16 */
3117 5, /* 24 */
3118 5, /* 32 */
3119 6, /* 40 */
3120 6, /* 48 */
3121 6, /* 56 */
3122 6, /* 64 */
3123 1, /* 72 */
3124 1, /* 80 */
3125 1, /* 88 */
3126 1, /* 96 */
3127 7, /* 104 */
3128 7, /* 112 */
3129 7, /* 120 */
3130 7, /* 128 */
3131 2, /* 136 */
3132 2, /* 144 */
3133 2, /* 152 */
3134 2, /* 160 */
3135 2, /* 168 */
3136 2, /* 176 */
3137 2, /* 184 */
3138 2 /* 192 */
3141 static inline int size_index_elem(size_t bytes)
3143 return (bytes - 1) / 8;
3146 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
3148 int index;
3150 if (size <= 192) {
3151 if (!size)
3152 return ZERO_SIZE_PTR;
3154 index = size_index[size_index_elem(size)];
3155 } else
3156 index = fls(size - 1);
3158 #ifdef CONFIG_ZONE_DMA
3159 if (unlikely((flags & SLUB_DMA)))
3160 return kmalloc_dma_caches[index];
3162 #endif
3163 return kmalloc_caches[index];
3166 void *__kmalloc(size_t size, gfp_t flags)
3168 struct kmem_cache *s;
3169 void *ret;
3171 if (unlikely(size > SLUB_MAX_SIZE))
3172 return kmalloc_large(size, flags);
3174 s = get_slab(size, flags);
3176 if (unlikely(ZERO_OR_NULL_PTR(s)))
3177 return s;
3179 ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_);
3181 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3183 return ret;
3185 EXPORT_SYMBOL(__kmalloc);
3187 #ifdef CONFIG_NUMA
3188 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3190 struct page *page;
3191 void *ptr = NULL;
3193 flags |= __GFP_COMP | __GFP_NOTRACK;
3194 page = alloc_pages_node(node, flags, get_order(size));
3195 if (page)
3196 ptr = page_address(page);
3198 kmemleak_alloc(ptr, size, 1, flags);
3199 return ptr;
3202 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3204 struct kmem_cache *s;
3205 void *ret;
3207 if (unlikely(size > SLUB_MAX_SIZE)) {
3208 ret = kmalloc_large_node(size, flags, node);
3210 trace_kmalloc_node(_RET_IP_, ret,
3211 size, PAGE_SIZE << get_order(size),
3212 flags, node);
3214 return ret;
3217 s = get_slab(size, flags);
3219 if (unlikely(ZERO_OR_NULL_PTR(s)))
3220 return s;
3222 ret = slab_alloc(s, flags, node, _RET_IP_);
3224 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3226 return ret;
3228 EXPORT_SYMBOL(__kmalloc_node);
3229 #endif
3231 size_t ksize(const void *object)
3233 struct page *page;
3235 if (unlikely(object == ZERO_SIZE_PTR))
3236 return 0;
3238 page = virt_to_head_page(object);
3240 if (unlikely(!PageSlab(page))) {
3241 WARN_ON(!PageCompound(page));
3242 return PAGE_SIZE << compound_order(page);
3245 return slab_ksize(page->slab);
3247 EXPORT_SYMBOL(ksize);
3249 #ifdef CONFIG_SLUB_DEBUG
3250 bool verify_mem_not_deleted(const void *x)
3252 struct page *page;
3253 void *object = (void *)x;
3254 unsigned long flags;
3255 bool rv;
3257 if (unlikely(ZERO_OR_NULL_PTR(x)))
3258 return false;
3260 local_irq_save(flags);
3262 page = virt_to_head_page(x);
3263 if (unlikely(!PageSlab(page))) {
3264 /* maybe it was from stack? */
3265 rv = true;
3266 goto out_unlock;
3269 slab_lock(page);
3270 if (on_freelist(page->slab, page, object)) {
3271 object_err(page->slab, page, object, "Object is on free-list");
3272 rv = false;
3273 } else {
3274 rv = true;
3276 slab_unlock(page);
3278 out_unlock:
3279 local_irq_restore(flags);
3280 return rv;
3282 EXPORT_SYMBOL(verify_mem_not_deleted);
3283 #endif
3285 void kfree(const void *x)
3287 struct page *page;
3288 void *object = (void *)x;
3290 trace_kfree(_RET_IP_, x);
3292 if (unlikely(ZERO_OR_NULL_PTR(x)))
3293 return;
3295 page = virt_to_head_page(x);
3296 if (unlikely(!PageSlab(page))) {
3297 BUG_ON(!PageCompound(page));
3298 kmemleak_free(x);
3299 put_page(page);
3300 return;
3302 slab_free(page->slab, page, object, _RET_IP_);
3304 EXPORT_SYMBOL(kfree);
3307 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3308 * the remaining slabs by the number of items in use. The slabs with the
3309 * most items in use come first. New allocations will then fill those up
3310 * and thus they can be removed from the partial lists.
3312 * The slabs with the least items are placed last. This results in them
3313 * being allocated from last increasing the chance that the last objects
3314 * are freed in them.
3316 int kmem_cache_shrink(struct kmem_cache *s)
3318 int node;
3319 int i;
3320 struct kmem_cache_node *n;
3321 struct page *page;
3322 struct page *t;
3323 int objects = oo_objects(s->max);
3324 struct list_head *slabs_by_inuse =
3325 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3326 unsigned long flags;
3328 if (!slabs_by_inuse)
3329 return -ENOMEM;
3331 flush_all(s);
3332 for_each_node_state(node, N_NORMAL_MEMORY) {
3333 n = get_node(s, node);
3335 if (!n->nr_partial)
3336 continue;
3338 for (i = 0; i < objects; i++)
3339 INIT_LIST_HEAD(slabs_by_inuse + i);
3341 spin_lock_irqsave(&n->list_lock, flags);
3344 * Build lists indexed by the items in use in each slab.
3346 * Note that concurrent frees may occur while we hold the
3347 * list_lock. page->inuse here is the upper limit.
3349 list_for_each_entry_safe(page, t, &n->partial, lru) {
3350 if (!page->inuse) {
3351 remove_partial(n, page);
3352 discard_slab(s, page);
3353 } else {
3354 list_move(&page->lru,
3355 slabs_by_inuse + page->inuse);
3360 * Rebuild the partial list with the slabs filled up most
3361 * first and the least used slabs at the end.
3363 for (i = objects - 1; i >= 0; i--)
3364 list_splice(slabs_by_inuse + i, n->partial.prev);
3366 spin_unlock_irqrestore(&n->list_lock, flags);
3369 kfree(slabs_by_inuse);
3370 return 0;
3372 EXPORT_SYMBOL(kmem_cache_shrink);
3374 #if defined(CONFIG_MEMORY_HOTPLUG)
3375 static int slab_mem_going_offline_callback(void *arg)
3377 struct kmem_cache *s;
3379 down_read(&slub_lock);
3380 list_for_each_entry(s, &slab_caches, list)
3381 kmem_cache_shrink(s);
3382 up_read(&slub_lock);
3384 return 0;
3387 static void slab_mem_offline_callback(void *arg)
3389 struct kmem_cache_node *n;
3390 struct kmem_cache *s;
3391 struct memory_notify *marg = arg;
3392 int offline_node;
3394 offline_node = marg->status_change_nid;
3397 * If the node still has available memory. we need kmem_cache_node
3398 * for it yet.
3400 if (offline_node < 0)
3401 return;
3403 down_read(&slub_lock);
3404 list_for_each_entry(s, &slab_caches, list) {
3405 n = get_node(s, offline_node);
3406 if (n) {
3408 * if n->nr_slabs > 0, slabs still exist on the node
3409 * that is going down. We were unable to free them,
3410 * and offline_pages() function shouldn't call this
3411 * callback. So, we must fail.
3413 BUG_ON(slabs_node(s, offline_node));
3415 s->node[offline_node] = NULL;
3416 kmem_cache_free(kmem_cache_node, n);
3419 up_read(&slub_lock);
3422 static int slab_mem_going_online_callback(void *arg)
3424 struct kmem_cache_node *n;
3425 struct kmem_cache *s;
3426 struct memory_notify *marg = arg;
3427 int nid = marg->status_change_nid;
3428 int ret = 0;
3431 * If the node's memory is already available, then kmem_cache_node is
3432 * already created. Nothing to do.
3434 if (nid < 0)
3435 return 0;
3438 * We are bringing a node online. No memory is available yet. We must
3439 * allocate a kmem_cache_node structure in order to bring the node
3440 * online.
3442 down_read(&slub_lock);
3443 list_for_each_entry(s, &slab_caches, list) {
3445 * XXX: kmem_cache_alloc_node will fallback to other nodes
3446 * since memory is not yet available from the node that
3447 * is brought up.
3449 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3450 if (!n) {
3451 ret = -ENOMEM;
3452 goto out;
3454 init_kmem_cache_node(n, s);
3455 s->node[nid] = n;
3457 out:
3458 up_read(&slub_lock);
3459 return ret;
3462 static int slab_memory_callback(struct notifier_block *self,
3463 unsigned long action, void *arg)
3465 int ret = 0;
3467 switch (action) {
3468 case MEM_GOING_ONLINE:
3469 ret = slab_mem_going_online_callback(arg);
3470 break;
3471 case MEM_GOING_OFFLINE:
3472 ret = slab_mem_going_offline_callback(arg);
3473 break;
3474 case MEM_OFFLINE:
3475 case MEM_CANCEL_ONLINE:
3476 slab_mem_offline_callback(arg);
3477 break;
3478 case MEM_ONLINE:
3479 case MEM_CANCEL_OFFLINE:
3480 break;
3482 if (ret)
3483 ret = notifier_from_errno(ret);
3484 else
3485 ret = NOTIFY_OK;
3486 return ret;
3489 #endif /* CONFIG_MEMORY_HOTPLUG */
3491 /********************************************************************
3492 * Basic setup of slabs
3493 *******************************************************************/
3496 * Used for early kmem_cache structures that were allocated using
3497 * the page allocator
3500 static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s)
3502 int node;
3504 list_add(&s->list, &slab_caches);
3505 s->refcount = -1;
3507 for_each_node_state(node, N_NORMAL_MEMORY) {
3508 struct kmem_cache_node *n = get_node(s, node);
3509 struct page *p;
3511 if (n) {
3512 list_for_each_entry(p, &n->partial, lru)
3513 p->slab = s;
3515 #ifdef CONFIG_SLUB_DEBUG
3516 list_for_each_entry(p, &n->full, lru)
3517 p->slab = s;
3518 #endif
3523 void __init kmem_cache_init(void)
3525 int i;
3526 int caches = 0;
3527 struct kmem_cache *temp_kmem_cache;
3528 int order;
3529 struct kmem_cache *temp_kmem_cache_node;
3530 unsigned long kmalloc_size;
3532 kmem_size = offsetof(struct kmem_cache, node) +
3533 nr_node_ids * sizeof(struct kmem_cache_node *);
3535 /* Allocate two kmem_caches from the page allocator */
3536 kmalloc_size = ALIGN(kmem_size, cache_line_size());
3537 order = get_order(2 * kmalloc_size);
3538 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order);
3541 * Must first have the slab cache available for the allocations of the
3542 * struct kmem_cache_node's. There is special bootstrap code in
3543 * kmem_cache_open for slab_state == DOWN.
3545 kmem_cache_node = (void *)kmem_cache + kmalloc_size;
3547 kmem_cache_open(kmem_cache_node, "kmem_cache_node",
3548 sizeof(struct kmem_cache_node),
3549 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3551 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3553 /* Able to allocate the per node structures */
3554 slab_state = PARTIAL;
3556 temp_kmem_cache = kmem_cache;
3557 kmem_cache_open(kmem_cache, "kmem_cache", kmem_size,
3558 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3559 kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3560 memcpy(kmem_cache, temp_kmem_cache, kmem_size);
3563 * Allocate kmem_cache_node properly from the kmem_cache slab.
3564 * kmem_cache_node is separately allocated so no need to
3565 * update any list pointers.
3567 temp_kmem_cache_node = kmem_cache_node;
3569 kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3570 memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size);
3572 kmem_cache_bootstrap_fixup(kmem_cache_node);
3574 caches++;
3575 kmem_cache_bootstrap_fixup(kmem_cache);
3576 caches++;
3577 /* Free temporary boot structure */
3578 free_pages((unsigned long)temp_kmem_cache, order);
3580 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3583 * Patch up the size_index table if we have strange large alignment
3584 * requirements for the kmalloc array. This is only the case for
3585 * MIPS it seems. The standard arches will not generate any code here.
3587 * Largest permitted alignment is 256 bytes due to the way we
3588 * handle the index determination for the smaller caches.
3590 * Make sure that nothing crazy happens if someone starts tinkering
3591 * around with ARCH_KMALLOC_MINALIGN
3593 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3594 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3596 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3597 int elem = size_index_elem(i);
3598 if (elem >= ARRAY_SIZE(size_index))
3599 break;
3600 size_index[elem] = KMALLOC_SHIFT_LOW;
3603 if (KMALLOC_MIN_SIZE == 64) {
3605 * The 96 byte size cache is not used if the alignment
3606 * is 64 byte.
3608 for (i = 64 + 8; i <= 96; i += 8)
3609 size_index[size_index_elem(i)] = 7;
3610 } else if (KMALLOC_MIN_SIZE == 128) {
3612 * The 192 byte sized cache is not used if the alignment
3613 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3614 * instead.
3616 for (i = 128 + 8; i <= 192; i += 8)
3617 size_index[size_index_elem(i)] = 8;
3620 /* Caches that are not of the two-to-the-power-of size */
3621 if (KMALLOC_MIN_SIZE <= 32) {
3622 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3623 caches++;
3626 if (KMALLOC_MIN_SIZE <= 64) {
3627 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3628 caches++;
3631 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3632 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3633 caches++;
3636 slab_state = UP;
3638 /* Provide the correct kmalloc names now that the caches are up */
3639 if (KMALLOC_MIN_SIZE <= 32) {
3640 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3641 BUG_ON(!kmalloc_caches[1]->name);
3644 if (KMALLOC_MIN_SIZE <= 64) {
3645 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3646 BUG_ON(!kmalloc_caches[2]->name);
3649 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3650 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3652 BUG_ON(!s);
3653 kmalloc_caches[i]->name = s;
3656 #ifdef CONFIG_SMP
3657 register_cpu_notifier(&slab_notifier);
3658 #endif
3660 #ifdef CONFIG_ZONE_DMA
3661 for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3662 struct kmem_cache *s = kmalloc_caches[i];
3664 if (s && s->size) {
3665 char *name = kasprintf(GFP_NOWAIT,
3666 "dma-kmalloc-%d", s->objsize);
3668 BUG_ON(!name);
3669 kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3670 s->objsize, SLAB_CACHE_DMA);
3673 #endif
3674 printk(KERN_INFO
3675 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3676 " CPUs=%d, Nodes=%d\n",
3677 caches, cache_line_size(),
3678 slub_min_order, slub_max_order, slub_min_objects,
3679 nr_cpu_ids, nr_node_ids);
3682 void __init kmem_cache_init_late(void)
3687 * Find a mergeable slab cache
3689 static int slab_unmergeable(struct kmem_cache *s)
3691 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3692 return 1;
3694 if (s->ctor)
3695 return 1;
3698 * We may have set a slab to be unmergeable during bootstrap.
3700 if (s->refcount < 0)
3701 return 1;
3703 return 0;
3706 static struct kmem_cache *find_mergeable(size_t size,
3707 size_t align, unsigned long flags, const char *name,
3708 void (*ctor)(void *))
3710 struct kmem_cache *s;
3712 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3713 return NULL;
3715 if (ctor)
3716 return NULL;
3718 size = ALIGN(size, sizeof(void *));
3719 align = calculate_alignment(flags, align, size);
3720 size = ALIGN(size, align);
3721 flags = kmem_cache_flags(size, flags, name, NULL);
3723 list_for_each_entry(s, &slab_caches, list) {
3724 if (slab_unmergeable(s))
3725 continue;
3727 if (size > s->size)
3728 continue;
3730 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3731 continue;
3733 * Check if alignment is compatible.
3734 * Courtesy of Adrian Drzewiecki
3736 if ((s->size & ~(align - 1)) != s->size)
3737 continue;
3739 if (s->size - size >= sizeof(void *))
3740 continue;
3742 return s;
3744 return NULL;
3747 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3748 size_t align, unsigned long flags, void (*ctor)(void *))
3750 struct kmem_cache *s;
3751 char *n;
3753 if (WARN_ON(!name))
3754 return NULL;
3756 down_write(&slub_lock);
3757 s = find_mergeable(size, align, flags, name, ctor);
3758 if (s) {
3759 s->refcount++;
3761 * Adjust the object sizes so that we clear
3762 * the complete object on kzalloc.
3764 s->objsize = max(s->objsize, (int)size);
3765 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3767 if (sysfs_slab_alias(s, name)) {
3768 s->refcount--;
3769 goto err;
3771 up_write(&slub_lock);
3772 return s;
3775 n = kstrdup(name, GFP_KERNEL);
3776 if (!n)
3777 goto err;
3779 s = kmalloc(kmem_size, GFP_KERNEL);
3780 if (s) {
3781 if (kmem_cache_open(s, n,
3782 size, align, flags, ctor)) {
3783 list_add(&s->list, &slab_caches);
3784 if (sysfs_slab_add(s)) {
3785 list_del(&s->list);
3786 kfree(n);
3787 kfree(s);
3788 goto err;
3790 up_write(&slub_lock);
3791 return s;
3793 kfree(n);
3794 kfree(s);
3796 err:
3797 up_write(&slub_lock);
3799 if (flags & SLAB_PANIC)
3800 panic("Cannot create slabcache %s\n", name);
3801 else
3802 s = NULL;
3803 return s;
3805 EXPORT_SYMBOL(kmem_cache_create);
3807 #ifdef CONFIG_SMP
3809 * Use the cpu notifier to insure that the cpu slabs are flushed when
3810 * necessary.
3812 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3813 unsigned long action, void *hcpu)
3815 long cpu = (long)hcpu;
3816 struct kmem_cache *s;
3817 unsigned long flags;
3819 switch (action) {
3820 case CPU_UP_CANCELED:
3821 case CPU_UP_CANCELED_FROZEN:
3822 case CPU_DEAD:
3823 case CPU_DEAD_FROZEN:
3824 down_read(&slub_lock);
3825 list_for_each_entry(s, &slab_caches, list) {
3826 local_irq_save(flags);
3827 __flush_cpu_slab(s, cpu);
3828 local_irq_restore(flags);
3830 up_read(&slub_lock);
3831 break;
3832 default:
3833 break;
3835 return NOTIFY_OK;
3838 static struct notifier_block __cpuinitdata slab_notifier = {
3839 .notifier_call = slab_cpuup_callback
3842 #endif
3844 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3846 struct kmem_cache *s;
3847 void *ret;
3849 if (unlikely(size > SLUB_MAX_SIZE))
3850 return kmalloc_large(size, gfpflags);
3852 s = get_slab(size, gfpflags);
3854 if (unlikely(ZERO_OR_NULL_PTR(s)))
3855 return s;
3857 ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller);
3859 /* Honor the call site pointer we received. */
3860 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3862 return ret;
3865 #ifdef CONFIG_NUMA
3866 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3867 int node, unsigned long caller)
3869 struct kmem_cache *s;
3870 void *ret;
3872 if (unlikely(size > SLUB_MAX_SIZE)) {
3873 ret = kmalloc_large_node(size, gfpflags, node);
3875 trace_kmalloc_node(caller, ret,
3876 size, PAGE_SIZE << get_order(size),
3877 gfpflags, node);
3879 return ret;
3882 s = get_slab(size, gfpflags);
3884 if (unlikely(ZERO_OR_NULL_PTR(s)))
3885 return s;
3887 ret = slab_alloc(s, gfpflags, node, caller);
3889 /* Honor the call site pointer we received. */
3890 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3892 return ret;
3894 #endif
3896 #ifdef CONFIG_SYSFS
3897 static int count_inuse(struct page *page)
3899 return page->inuse;
3902 static int count_total(struct page *page)
3904 return page->objects;
3906 #endif
3908 #ifdef CONFIG_SLUB_DEBUG
3909 static int validate_slab(struct kmem_cache *s, struct page *page,
3910 unsigned long *map)
3912 void *p;
3913 void *addr = page_address(page);
3915 if (!check_slab(s, page) ||
3916 !on_freelist(s, page, NULL))
3917 return 0;
3919 /* Now we know that a valid freelist exists */
3920 bitmap_zero(map, page->objects);
3922 get_map(s, page, map);
3923 for_each_object(p, s, addr, page->objects) {
3924 if (test_bit(slab_index(p, s, addr), map))
3925 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3926 return 0;
3929 for_each_object(p, s, addr, page->objects)
3930 if (!test_bit(slab_index(p, s, addr), map))
3931 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3932 return 0;
3933 return 1;
3936 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3937 unsigned long *map)
3939 slab_lock(page);
3940 validate_slab(s, page, map);
3941 slab_unlock(page);
3944 static int validate_slab_node(struct kmem_cache *s,
3945 struct kmem_cache_node *n, unsigned long *map)
3947 unsigned long count = 0;
3948 struct page *page;
3949 unsigned long flags;
3951 spin_lock_irqsave(&n->list_lock, flags);
3953 list_for_each_entry(page, &n->partial, lru) {
3954 validate_slab_slab(s, page, map);
3955 count++;
3957 if (count != n->nr_partial)
3958 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3959 "counter=%ld\n", s->name, count, n->nr_partial);
3961 if (!(s->flags & SLAB_STORE_USER))
3962 goto out;
3964 list_for_each_entry(page, &n->full, lru) {
3965 validate_slab_slab(s, page, map);
3966 count++;
3968 if (count != atomic_long_read(&n->nr_slabs))
3969 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3970 "counter=%ld\n", s->name, count,
3971 atomic_long_read(&n->nr_slabs));
3973 out:
3974 spin_unlock_irqrestore(&n->list_lock, flags);
3975 return count;
3978 static long validate_slab_cache(struct kmem_cache *s)
3980 int node;
3981 unsigned long count = 0;
3982 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3983 sizeof(unsigned long), GFP_KERNEL);
3985 if (!map)
3986 return -ENOMEM;
3988 flush_all(s);
3989 for_each_node_state(node, N_NORMAL_MEMORY) {
3990 struct kmem_cache_node *n = get_node(s, node);
3992 count += validate_slab_node(s, n, map);
3994 kfree(map);
3995 return count;
3998 * Generate lists of code addresses where slabcache objects are allocated
3999 * and freed.
4002 struct location {
4003 unsigned long count;
4004 unsigned long addr;
4005 long long sum_time;
4006 long min_time;
4007 long max_time;
4008 long min_pid;
4009 long max_pid;
4010 DECLARE_BITMAP(cpus, NR_CPUS);
4011 nodemask_t nodes;
4014 struct loc_track {
4015 unsigned long max;
4016 unsigned long count;
4017 struct location *loc;
4020 static void free_loc_track(struct loc_track *t)
4022 if (t->max)
4023 free_pages((unsigned long)t->loc,
4024 get_order(sizeof(struct location) * t->max));
4027 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4029 struct location *l;
4030 int order;
4032 order = get_order(sizeof(struct location) * max);
4034 l = (void *)__get_free_pages(flags, order);
4035 if (!l)
4036 return 0;
4038 if (t->count) {
4039 memcpy(l, t->loc, sizeof(struct location) * t->count);
4040 free_loc_track(t);
4042 t->max = max;
4043 t->loc = l;
4044 return 1;
4047 static int add_location(struct loc_track *t, struct kmem_cache *s,
4048 const struct track *track)
4050 long start, end, pos;
4051 struct location *l;
4052 unsigned long caddr;
4053 unsigned long age = jiffies - track->when;
4055 start = -1;
4056 end = t->count;
4058 for ( ; ; ) {
4059 pos = start + (end - start + 1) / 2;
4062 * There is nothing at "end". If we end up there
4063 * we need to add something to before end.
4065 if (pos == end)
4066 break;
4068 caddr = t->loc[pos].addr;
4069 if (track->addr == caddr) {
4071 l = &t->loc[pos];
4072 l->count++;
4073 if (track->when) {
4074 l->sum_time += age;
4075 if (age < l->min_time)
4076 l->min_time = age;
4077 if (age > l->max_time)
4078 l->max_time = age;
4080 if (track->pid < l->min_pid)
4081 l->min_pid = track->pid;
4082 if (track->pid > l->max_pid)
4083 l->max_pid = track->pid;
4085 cpumask_set_cpu(track->cpu,
4086 to_cpumask(l->cpus));
4088 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4089 return 1;
4092 if (track->addr < caddr)
4093 end = pos;
4094 else
4095 start = pos;
4099 * Not found. Insert new tracking element.
4101 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4102 return 0;
4104 l = t->loc + pos;
4105 if (pos < t->count)
4106 memmove(l + 1, l,
4107 (t->count - pos) * sizeof(struct location));
4108 t->count++;
4109 l->count = 1;
4110 l->addr = track->addr;
4111 l->sum_time = age;
4112 l->min_time = age;
4113 l->max_time = age;
4114 l->min_pid = track->pid;
4115 l->max_pid = track->pid;
4116 cpumask_clear(to_cpumask(l->cpus));
4117 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4118 nodes_clear(l->nodes);
4119 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4120 return 1;
4123 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4124 struct page *page, enum track_item alloc,
4125 unsigned long *map)
4127 void *addr = page_address(page);
4128 void *p;
4130 bitmap_zero(map, page->objects);
4131 get_map(s, page, map);
4133 for_each_object(p, s, addr, page->objects)
4134 if (!test_bit(slab_index(p, s, addr), map))
4135 add_location(t, s, get_track(s, p, alloc));
4138 static int list_locations(struct kmem_cache *s, char *buf,
4139 enum track_item alloc)
4141 int len = 0;
4142 unsigned long i;
4143 struct loc_track t = { 0, 0, NULL };
4144 int node;
4145 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4146 sizeof(unsigned long), GFP_KERNEL);
4148 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4149 GFP_TEMPORARY)) {
4150 kfree(map);
4151 return sprintf(buf, "Out of memory\n");
4153 /* Push back cpu slabs */
4154 flush_all(s);
4156 for_each_node_state(node, N_NORMAL_MEMORY) {
4157 struct kmem_cache_node *n = get_node(s, node);
4158 unsigned long flags;
4159 struct page *page;
4161 if (!atomic_long_read(&n->nr_slabs))
4162 continue;
4164 spin_lock_irqsave(&n->list_lock, flags);
4165 list_for_each_entry(page, &n->partial, lru)
4166 process_slab(&t, s, page, alloc, map);
4167 list_for_each_entry(page, &n->full, lru)
4168 process_slab(&t, s, page, alloc, map);
4169 spin_unlock_irqrestore(&n->list_lock, flags);
4172 for (i = 0; i < t.count; i++) {
4173 struct location *l = &t.loc[i];
4175 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4176 break;
4177 len += sprintf(buf + len, "%7ld ", l->count);
4179 if (l->addr)
4180 len += sprintf(buf + len, "%pS", (void *)l->addr);
4181 else
4182 len += sprintf(buf + len, "<not-available>");
4184 if (l->sum_time != l->min_time) {
4185 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4186 l->min_time,
4187 (long)div_u64(l->sum_time, l->count),
4188 l->max_time);
4189 } else
4190 len += sprintf(buf + len, " age=%ld",
4191 l->min_time);
4193 if (l->min_pid != l->max_pid)
4194 len += sprintf(buf + len, " pid=%ld-%ld",
4195 l->min_pid, l->max_pid);
4196 else
4197 len += sprintf(buf + len, " pid=%ld",
4198 l->min_pid);
4200 if (num_online_cpus() > 1 &&
4201 !cpumask_empty(to_cpumask(l->cpus)) &&
4202 len < PAGE_SIZE - 60) {
4203 len += sprintf(buf + len, " cpus=");
4204 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4205 to_cpumask(l->cpus));
4208 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4209 len < PAGE_SIZE - 60) {
4210 len += sprintf(buf + len, " nodes=");
4211 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4212 l->nodes);
4215 len += sprintf(buf + len, "\n");
4218 free_loc_track(&t);
4219 kfree(map);
4220 if (!t.count)
4221 len += sprintf(buf, "No data\n");
4222 return len;
4224 #endif
4226 #ifdef SLUB_RESILIENCY_TEST
4227 static void resiliency_test(void)
4229 u8 *p;
4231 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
4233 printk(KERN_ERR "SLUB resiliency testing\n");
4234 printk(KERN_ERR "-----------------------\n");
4235 printk(KERN_ERR "A. Corruption after allocation\n");
4237 p = kzalloc(16, GFP_KERNEL);
4238 p[16] = 0x12;
4239 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4240 " 0x12->0x%p\n\n", p + 16);
4242 validate_slab_cache(kmalloc_caches[4]);
4244 /* Hmmm... The next two are dangerous */
4245 p = kzalloc(32, GFP_KERNEL);
4246 p[32 + sizeof(void *)] = 0x34;
4247 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4248 " 0x34 -> -0x%p\n", p);
4249 printk(KERN_ERR
4250 "If allocated object is overwritten then not detectable\n\n");
4252 validate_slab_cache(kmalloc_caches[5]);
4253 p = kzalloc(64, GFP_KERNEL);
4254 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4255 *p = 0x56;
4256 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4258 printk(KERN_ERR
4259 "If allocated object is overwritten then not detectable\n\n");
4260 validate_slab_cache(kmalloc_caches[6]);
4262 printk(KERN_ERR "\nB. Corruption after free\n");
4263 p = kzalloc(128, GFP_KERNEL);
4264 kfree(p);
4265 *p = 0x78;
4266 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4267 validate_slab_cache(kmalloc_caches[7]);
4269 p = kzalloc(256, GFP_KERNEL);
4270 kfree(p);
4271 p[50] = 0x9a;
4272 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4274 validate_slab_cache(kmalloc_caches[8]);
4276 p = kzalloc(512, GFP_KERNEL);
4277 kfree(p);
4278 p[512] = 0xab;
4279 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4280 validate_slab_cache(kmalloc_caches[9]);
4282 #else
4283 #ifdef CONFIG_SYSFS
4284 static void resiliency_test(void) {};
4285 #endif
4286 #endif
4288 #ifdef CONFIG_SYSFS
4289 enum slab_stat_type {
4290 SL_ALL, /* All slabs */
4291 SL_PARTIAL, /* Only partially allocated slabs */
4292 SL_CPU, /* Only slabs used for cpu caches */
4293 SL_OBJECTS, /* Determine allocated objects not slabs */
4294 SL_TOTAL /* Determine object capacity not slabs */
4297 #define SO_ALL (1 << SL_ALL)
4298 #define SO_PARTIAL (1 << SL_PARTIAL)
4299 #define SO_CPU (1 << SL_CPU)
4300 #define SO_OBJECTS (1 << SL_OBJECTS)
4301 #define SO_TOTAL (1 << SL_TOTAL)
4303 static ssize_t show_slab_objects(struct kmem_cache *s,
4304 char *buf, unsigned long flags)
4306 unsigned long total = 0;
4307 int node;
4308 int x;
4309 unsigned long *nodes;
4310 unsigned long *per_cpu;
4312 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4313 if (!nodes)
4314 return -ENOMEM;
4315 per_cpu = nodes + nr_node_ids;
4317 if (flags & SO_CPU) {
4318 int cpu;
4320 for_each_possible_cpu(cpu) {
4321 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4323 if (!c || c->node < 0)
4324 continue;
4326 if (c->page) {
4327 if (flags & SO_TOTAL)
4328 x = c->page->objects;
4329 else if (flags & SO_OBJECTS)
4330 x = c->page->inuse;
4331 else
4332 x = 1;
4334 total += x;
4335 nodes[c->node] += x;
4337 per_cpu[c->node]++;
4341 lock_memory_hotplug();
4342 #ifdef CONFIG_SLUB_DEBUG
4343 if (flags & SO_ALL) {
4344 for_each_node_state(node, N_NORMAL_MEMORY) {
4345 struct kmem_cache_node *n = get_node(s, node);
4347 if (flags & SO_TOTAL)
4348 x = atomic_long_read(&n->total_objects);
4349 else if (flags & SO_OBJECTS)
4350 x = atomic_long_read(&n->total_objects) -
4351 count_partial(n, count_free);
4353 else
4354 x = atomic_long_read(&n->nr_slabs);
4355 total += x;
4356 nodes[node] += x;
4359 } else
4360 #endif
4361 if (flags & SO_PARTIAL) {
4362 for_each_node_state(node, N_NORMAL_MEMORY) {
4363 struct kmem_cache_node *n = get_node(s, node);
4365 if (flags & SO_TOTAL)
4366 x = count_partial(n, count_total);
4367 else if (flags & SO_OBJECTS)
4368 x = count_partial(n, count_inuse);
4369 else
4370 x = n->nr_partial;
4371 total += x;
4372 nodes[node] += x;
4375 x = sprintf(buf, "%lu", total);
4376 #ifdef CONFIG_NUMA
4377 for_each_node_state(node, N_NORMAL_MEMORY)
4378 if (nodes[node])
4379 x += sprintf(buf + x, " N%d=%lu",
4380 node, nodes[node]);
4381 #endif
4382 unlock_memory_hotplug();
4383 kfree(nodes);
4384 return x + sprintf(buf + x, "\n");
4387 #ifdef CONFIG_SLUB_DEBUG
4388 static int any_slab_objects(struct kmem_cache *s)
4390 int node;
4392 for_each_online_node(node) {
4393 struct kmem_cache_node *n = get_node(s, node);
4395 if (!n)
4396 continue;
4398 if (atomic_long_read(&n->total_objects))
4399 return 1;
4401 return 0;
4403 #endif
4405 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4406 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4408 struct slab_attribute {
4409 struct attribute attr;
4410 ssize_t (*show)(struct kmem_cache *s, char *buf);
4411 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4414 #define SLAB_ATTR_RO(_name) \
4415 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
4417 #define SLAB_ATTR(_name) \
4418 static struct slab_attribute _name##_attr = \
4419 __ATTR(_name, 0644, _name##_show, _name##_store)
4421 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4423 return sprintf(buf, "%d\n", s->size);
4425 SLAB_ATTR_RO(slab_size);
4427 static ssize_t align_show(struct kmem_cache *s, char *buf)
4429 return sprintf(buf, "%d\n", s->align);
4431 SLAB_ATTR_RO(align);
4433 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4435 return sprintf(buf, "%d\n", s->objsize);
4437 SLAB_ATTR_RO(object_size);
4439 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4441 return sprintf(buf, "%d\n", oo_objects(s->oo));
4443 SLAB_ATTR_RO(objs_per_slab);
4445 static ssize_t order_store(struct kmem_cache *s,
4446 const char *buf, size_t length)
4448 unsigned long order;
4449 int err;
4451 err = strict_strtoul(buf, 10, &order);
4452 if (err)
4453 return err;
4455 if (order > slub_max_order || order < slub_min_order)
4456 return -EINVAL;
4458 calculate_sizes(s, order);
4459 return length;
4462 static ssize_t order_show(struct kmem_cache *s, char *buf)
4464 return sprintf(buf, "%d\n", oo_order(s->oo));
4466 SLAB_ATTR(order);
4468 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4470 return sprintf(buf, "%lu\n", s->min_partial);
4473 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4474 size_t length)
4476 unsigned long min;
4477 int err;
4479 err = strict_strtoul(buf, 10, &min);
4480 if (err)
4481 return err;
4483 set_min_partial(s, min);
4484 return length;
4486 SLAB_ATTR(min_partial);
4488 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4490 if (!s->ctor)
4491 return 0;
4492 return sprintf(buf, "%pS\n", s->ctor);
4494 SLAB_ATTR_RO(ctor);
4496 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4498 return sprintf(buf, "%d\n", s->refcount - 1);
4500 SLAB_ATTR_RO(aliases);
4502 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4504 return show_slab_objects(s, buf, SO_PARTIAL);
4506 SLAB_ATTR_RO(partial);
4508 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4510 return show_slab_objects(s, buf, SO_CPU);
4512 SLAB_ATTR_RO(cpu_slabs);
4514 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4516 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4518 SLAB_ATTR_RO(objects);
4520 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4522 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4524 SLAB_ATTR_RO(objects_partial);
4526 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4528 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4531 static ssize_t reclaim_account_store(struct kmem_cache *s,
4532 const char *buf, size_t length)
4534 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4535 if (buf[0] == '1')
4536 s->flags |= SLAB_RECLAIM_ACCOUNT;
4537 return length;
4539 SLAB_ATTR(reclaim_account);
4541 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4543 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4545 SLAB_ATTR_RO(hwcache_align);
4547 #ifdef CONFIG_ZONE_DMA
4548 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4550 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4552 SLAB_ATTR_RO(cache_dma);
4553 #endif
4555 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4557 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4559 SLAB_ATTR_RO(destroy_by_rcu);
4561 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4563 return sprintf(buf, "%d\n", s->reserved);
4565 SLAB_ATTR_RO(reserved);
4567 #ifdef CONFIG_SLUB_DEBUG
4568 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4570 return show_slab_objects(s, buf, SO_ALL);
4572 SLAB_ATTR_RO(slabs);
4574 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4576 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4578 SLAB_ATTR_RO(total_objects);
4580 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4582 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4585 static ssize_t sanity_checks_store(struct kmem_cache *s,
4586 const char *buf, size_t length)
4588 s->flags &= ~SLAB_DEBUG_FREE;
4589 if (buf[0] == '1') {
4590 s->flags &= ~__CMPXCHG_DOUBLE;
4591 s->flags |= SLAB_DEBUG_FREE;
4593 return length;
4595 SLAB_ATTR(sanity_checks);
4597 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4599 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4602 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4603 size_t length)
4605 s->flags &= ~SLAB_TRACE;
4606 if (buf[0] == '1') {
4607 s->flags &= ~__CMPXCHG_DOUBLE;
4608 s->flags |= SLAB_TRACE;
4610 return length;
4612 SLAB_ATTR(trace);
4614 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4616 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4619 static ssize_t red_zone_store(struct kmem_cache *s,
4620 const char *buf, size_t length)
4622 if (any_slab_objects(s))
4623 return -EBUSY;
4625 s->flags &= ~SLAB_RED_ZONE;
4626 if (buf[0] == '1') {
4627 s->flags &= ~__CMPXCHG_DOUBLE;
4628 s->flags |= SLAB_RED_ZONE;
4630 calculate_sizes(s, -1);
4631 return length;
4633 SLAB_ATTR(red_zone);
4635 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4637 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4640 static ssize_t poison_store(struct kmem_cache *s,
4641 const char *buf, size_t length)
4643 if (any_slab_objects(s))
4644 return -EBUSY;
4646 s->flags &= ~SLAB_POISON;
4647 if (buf[0] == '1') {
4648 s->flags &= ~__CMPXCHG_DOUBLE;
4649 s->flags |= SLAB_POISON;
4651 calculate_sizes(s, -1);
4652 return length;
4654 SLAB_ATTR(poison);
4656 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4658 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4661 static ssize_t store_user_store(struct kmem_cache *s,
4662 const char *buf, size_t length)
4664 if (any_slab_objects(s))
4665 return -EBUSY;
4667 s->flags &= ~SLAB_STORE_USER;
4668 if (buf[0] == '1') {
4669 s->flags &= ~__CMPXCHG_DOUBLE;
4670 s->flags |= SLAB_STORE_USER;
4672 calculate_sizes(s, -1);
4673 return length;
4675 SLAB_ATTR(store_user);
4677 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4679 return 0;
4682 static ssize_t validate_store(struct kmem_cache *s,
4683 const char *buf, size_t length)
4685 int ret = -EINVAL;
4687 if (buf[0] == '1') {
4688 ret = validate_slab_cache(s);
4689 if (ret >= 0)
4690 ret = length;
4692 return ret;
4694 SLAB_ATTR(validate);
4696 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4698 if (!(s->flags & SLAB_STORE_USER))
4699 return -ENOSYS;
4700 return list_locations(s, buf, TRACK_ALLOC);
4702 SLAB_ATTR_RO(alloc_calls);
4704 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4706 if (!(s->flags & SLAB_STORE_USER))
4707 return -ENOSYS;
4708 return list_locations(s, buf, TRACK_FREE);
4710 SLAB_ATTR_RO(free_calls);
4711 #endif /* CONFIG_SLUB_DEBUG */
4713 #ifdef CONFIG_FAILSLAB
4714 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4716 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4719 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4720 size_t length)
4722 s->flags &= ~SLAB_FAILSLAB;
4723 if (buf[0] == '1')
4724 s->flags |= SLAB_FAILSLAB;
4725 return length;
4727 SLAB_ATTR(failslab);
4728 #endif
4730 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4732 return 0;
4735 static ssize_t shrink_store(struct kmem_cache *s,
4736 const char *buf, size_t length)
4738 if (buf[0] == '1') {
4739 int rc = kmem_cache_shrink(s);
4741 if (rc)
4742 return rc;
4743 } else
4744 return -EINVAL;
4745 return length;
4747 SLAB_ATTR(shrink);
4749 #ifdef CONFIG_NUMA
4750 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4752 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4755 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4756 const char *buf, size_t length)
4758 unsigned long ratio;
4759 int err;
4761 err = strict_strtoul(buf, 10, &ratio);
4762 if (err)
4763 return err;
4765 if (ratio <= 100)
4766 s->remote_node_defrag_ratio = ratio * 10;
4768 return length;
4770 SLAB_ATTR(remote_node_defrag_ratio);
4771 #endif
4773 #ifdef CONFIG_SLUB_STATS
4774 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4776 unsigned long sum = 0;
4777 int cpu;
4778 int len;
4779 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4781 if (!data)
4782 return -ENOMEM;
4784 for_each_online_cpu(cpu) {
4785 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4787 data[cpu] = x;
4788 sum += x;
4791 len = sprintf(buf, "%lu", sum);
4793 #ifdef CONFIG_SMP
4794 for_each_online_cpu(cpu) {
4795 if (data[cpu] && len < PAGE_SIZE - 20)
4796 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4798 #endif
4799 kfree(data);
4800 return len + sprintf(buf + len, "\n");
4803 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4805 int cpu;
4807 for_each_online_cpu(cpu)
4808 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4811 #define STAT_ATTR(si, text) \
4812 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4814 return show_stat(s, buf, si); \
4816 static ssize_t text##_store(struct kmem_cache *s, \
4817 const char *buf, size_t length) \
4819 if (buf[0] != '0') \
4820 return -EINVAL; \
4821 clear_stat(s, si); \
4822 return length; \
4824 SLAB_ATTR(text); \
4826 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4827 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4828 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4829 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4830 STAT_ATTR(FREE_FROZEN, free_frozen);
4831 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4832 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4833 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4834 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4835 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4836 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
4837 STAT_ATTR(FREE_SLAB, free_slab);
4838 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4839 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4840 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4841 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4842 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4843 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4844 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
4845 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4846 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4847 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4848 #endif
4850 static struct attribute *slab_attrs[] = {
4851 &slab_size_attr.attr,
4852 &object_size_attr.attr,
4853 &objs_per_slab_attr.attr,
4854 &order_attr.attr,
4855 &min_partial_attr.attr,
4856 &objects_attr.attr,
4857 &objects_partial_attr.attr,
4858 &partial_attr.attr,
4859 &cpu_slabs_attr.attr,
4860 &ctor_attr.attr,
4861 &aliases_attr.attr,
4862 &align_attr.attr,
4863 &hwcache_align_attr.attr,
4864 &reclaim_account_attr.attr,
4865 &destroy_by_rcu_attr.attr,
4866 &shrink_attr.attr,
4867 &reserved_attr.attr,
4868 #ifdef CONFIG_SLUB_DEBUG
4869 &total_objects_attr.attr,
4870 &slabs_attr.attr,
4871 &sanity_checks_attr.attr,
4872 &trace_attr.attr,
4873 &red_zone_attr.attr,
4874 &poison_attr.attr,
4875 &store_user_attr.attr,
4876 &validate_attr.attr,
4877 &alloc_calls_attr.attr,
4878 &free_calls_attr.attr,
4879 #endif
4880 #ifdef CONFIG_ZONE_DMA
4881 &cache_dma_attr.attr,
4882 #endif
4883 #ifdef CONFIG_NUMA
4884 &remote_node_defrag_ratio_attr.attr,
4885 #endif
4886 #ifdef CONFIG_SLUB_STATS
4887 &alloc_fastpath_attr.attr,
4888 &alloc_slowpath_attr.attr,
4889 &free_fastpath_attr.attr,
4890 &free_slowpath_attr.attr,
4891 &free_frozen_attr.attr,
4892 &free_add_partial_attr.attr,
4893 &free_remove_partial_attr.attr,
4894 &alloc_from_partial_attr.attr,
4895 &alloc_slab_attr.attr,
4896 &alloc_refill_attr.attr,
4897 &alloc_node_mismatch_attr.attr,
4898 &free_slab_attr.attr,
4899 &cpuslab_flush_attr.attr,
4900 &deactivate_full_attr.attr,
4901 &deactivate_empty_attr.attr,
4902 &deactivate_to_head_attr.attr,
4903 &deactivate_to_tail_attr.attr,
4904 &deactivate_remote_frees_attr.attr,
4905 &deactivate_bypass_attr.attr,
4906 &order_fallback_attr.attr,
4907 &cmpxchg_double_fail_attr.attr,
4908 &cmpxchg_double_cpu_fail_attr.attr,
4909 #endif
4910 #ifdef CONFIG_FAILSLAB
4911 &failslab_attr.attr,
4912 #endif
4914 NULL
4917 static struct attribute_group slab_attr_group = {
4918 .attrs = slab_attrs,
4921 static ssize_t slab_attr_show(struct kobject *kobj,
4922 struct attribute *attr,
4923 char *buf)
4925 struct slab_attribute *attribute;
4926 struct kmem_cache *s;
4927 int err;
4929 attribute = to_slab_attr(attr);
4930 s = to_slab(kobj);
4932 if (!attribute->show)
4933 return -EIO;
4935 err = attribute->show(s, buf);
4937 return err;
4940 static ssize_t slab_attr_store(struct kobject *kobj,
4941 struct attribute *attr,
4942 const char *buf, size_t len)
4944 struct slab_attribute *attribute;
4945 struct kmem_cache *s;
4946 int err;
4948 attribute = to_slab_attr(attr);
4949 s = to_slab(kobj);
4951 if (!attribute->store)
4952 return -EIO;
4954 err = attribute->store(s, buf, len);
4956 return err;
4959 static void kmem_cache_release(struct kobject *kobj)
4961 struct kmem_cache *s = to_slab(kobj);
4963 kfree(s->name);
4964 kfree(s);
4967 static const struct sysfs_ops slab_sysfs_ops = {
4968 .show = slab_attr_show,
4969 .store = slab_attr_store,
4972 static struct kobj_type slab_ktype = {
4973 .sysfs_ops = &slab_sysfs_ops,
4974 .release = kmem_cache_release
4977 static int uevent_filter(struct kset *kset, struct kobject *kobj)
4979 struct kobj_type *ktype = get_ktype(kobj);
4981 if (ktype == &slab_ktype)
4982 return 1;
4983 return 0;
4986 static const struct kset_uevent_ops slab_uevent_ops = {
4987 .filter = uevent_filter,
4990 static struct kset *slab_kset;
4992 #define ID_STR_LENGTH 64
4994 /* Create a unique string id for a slab cache:
4996 * Format :[flags-]size
4998 static char *create_unique_id(struct kmem_cache *s)
5000 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5001 char *p = name;
5003 BUG_ON(!name);
5005 *p++ = ':';
5007 * First flags affecting slabcache operations. We will only
5008 * get here for aliasable slabs so we do not need to support
5009 * too many flags. The flags here must cover all flags that
5010 * are matched during merging to guarantee that the id is
5011 * unique.
5013 if (s->flags & SLAB_CACHE_DMA)
5014 *p++ = 'd';
5015 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5016 *p++ = 'a';
5017 if (s->flags & SLAB_DEBUG_FREE)
5018 *p++ = 'F';
5019 if (!(s->flags & SLAB_NOTRACK))
5020 *p++ = 't';
5021 if (p != name + 1)
5022 *p++ = '-';
5023 p += sprintf(p, "%07d", s->size);
5024 BUG_ON(p > name + ID_STR_LENGTH - 1);
5025 return name;
5028 static int sysfs_slab_add(struct kmem_cache *s)
5030 int err;
5031 const char *name;
5032 int unmergeable;
5034 if (slab_state < SYSFS)
5035 /* Defer until later */
5036 return 0;
5038 unmergeable = slab_unmergeable(s);
5039 if (unmergeable) {
5041 * Slabcache can never be merged so we can use the name proper.
5042 * This is typically the case for debug situations. In that
5043 * case we can catch duplicate names easily.
5045 sysfs_remove_link(&slab_kset->kobj, s->name);
5046 name = s->name;
5047 } else {
5049 * Create a unique name for the slab as a target
5050 * for the symlinks.
5052 name = create_unique_id(s);
5055 s->kobj.kset = slab_kset;
5056 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5057 if (err) {
5058 kobject_put(&s->kobj);
5059 return err;
5062 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5063 if (err) {
5064 kobject_del(&s->kobj);
5065 kobject_put(&s->kobj);
5066 return err;
5068 kobject_uevent(&s->kobj, KOBJ_ADD);
5069 if (!unmergeable) {
5070 /* Setup first alias */
5071 sysfs_slab_alias(s, s->name);
5072 kfree(name);
5074 return 0;
5077 static void sysfs_slab_remove(struct kmem_cache *s)
5079 if (slab_state < SYSFS)
5081 * Sysfs has not been setup yet so no need to remove the
5082 * cache from sysfs.
5084 return;
5086 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5087 kobject_del(&s->kobj);
5088 kobject_put(&s->kobj);
5092 * Need to buffer aliases during bootup until sysfs becomes
5093 * available lest we lose that information.
5095 struct saved_alias {
5096 struct kmem_cache *s;
5097 const char *name;
5098 struct saved_alias *next;
5101 static struct saved_alias *alias_list;
5103 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5105 struct saved_alias *al;
5107 if (slab_state == SYSFS) {
5109 * If we have a leftover link then remove it.
5111 sysfs_remove_link(&slab_kset->kobj, name);
5112 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5115 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5116 if (!al)
5117 return -ENOMEM;
5119 al->s = s;
5120 al->name = name;
5121 al->next = alias_list;
5122 alias_list = al;
5123 return 0;
5126 static int __init slab_sysfs_init(void)
5128 struct kmem_cache *s;
5129 int err;
5131 down_write(&slub_lock);
5133 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5134 if (!slab_kset) {
5135 up_write(&slub_lock);
5136 printk(KERN_ERR "Cannot register slab subsystem.\n");
5137 return -ENOSYS;
5140 slab_state = SYSFS;
5142 list_for_each_entry(s, &slab_caches, list) {
5143 err = sysfs_slab_add(s);
5144 if (err)
5145 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5146 " to sysfs\n", s->name);
5149 while (alias_list) {
5150 struct saved_alias *al = alias_list;
5152 alias_list = alias_list->next;
5153 err = sysfs_slab_alias(al->s, al->name);
5154 if (err)
5155 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5156 " %s to sysfs\n", s->name);
5157 kfree(al);
5160 up_write(&slub_lock);
5161 resiliency_test();
5162 return 0;
5165 __initcall(slab_sysfs_init);
5166 #endif /* CONFIG_SYSFS */
5169 * The /proc/slabinfo ABI
5171 #ifdef CONFIG_SLABINFO
5172 static void print_slabinfo_header(struct seq_file *m)
5174 seq_puts(m, "slabinfo - version: 2.1\n");
5175 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
5176 "<objperslab> <pagesperslab>");
5177 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
5178 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
5179 seq_putc(m, '\n');
5182 static void *s_start(struct seq_file *m, loff_t *pos)
5184 loff_t n = *pos;
5186 down_read(&slub_lock);
5187 if (!n)
5188 print_slabinfo_header(m);
5190 return seq_list_start(&slab_caches, *pos);
5193 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
5195 return seq_list_next(p, &slab_caches, pos);
5198 static void s_stop(struct seq_file *m, void *p)
5200 up_read(&slub_lock);
5203 static int s_show(struct seq_file *m, void *p)
5205 unsigned long nr_partials = 0;
5206 unsigned long nr_slabs = 0;
5207 unsigned long nr_inuse = 0;
5208 unsigned long nr_objs = 0;
5209 unsigned long nr_free = 0;
5210 struct kmem_cache *s;
5211 int node;
5213 s = list_entry(p, struct kmem_cache, list);
5215 for_each_online_node(node) {
5216 struct kmem_cache_node *n = get_node(s, node);
5218 if (!n)
5219 continue;
5221 nr_partials += n->nr_partial;
5222 nr_slabs += atomic_long_read(&n->nr_slabs);
5223 nr_objs += atomic_long_read(&n->total_objects);
5224 nr_free += count_partial(n, count_free);
5227 nr_inuse = nr_objs - nr_free;
5229 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
5230 nr_objs, s->size, oo_objects(s->oo),
5231 (1 << oo_order(s->oo)));
5232 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
5233 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
5234 0UL);
5235 seq_putc(m, '\n');
5236 return 0;
5239 static const struct seq_operations slabinfo_op = {
5240 .start = s_start,
5241 .next = s_next,
5242 .stop = s_stop,
5243 .show = s_show,
5246 static int slabinfo_open(struct inode *inode, struct file *file)
5248 return seq_open(file, &slabinfo_op);
5251 static const struct file_operations proc_slabinfo_operations = {
5252 .open = slabinfo_open,
5253 .read = seq_read,
5254 .llseek = seq_lseek,
5255 .release = seq_release,
5258 static int __init slab_proc_init(void)
5260 proc_create("slabinfo", S_IRUGO, NULL, &proc_slabinfo_operations);
5261 return 0;
5263 module_init(slab_proc_init);
5264 #endif /* CONFIG_SLABINFO */