[NET]: Avoid pointless allocation casts in BSD compression module
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / mm / slub.c
blob04151da399c6ce8c344409c2ce7fbf289ac645e5
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 and only
6 * uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter <clameter@sgi.com>
9 */
11 #include <linux/mm.h>
12 #include <linux/module.h>
13 #include <linux/bit_spinlock.h>
14 #include <linux/interrupt.h>
15 #include <linux/bitops.h>
16 #include <linux/slab.h>
17 #include <linux/seq_file.h>
18 #include <linux/cpu.h>
19 #include <linux/cpuset.h>
20 #include <linux/mempolicy.h>
21 #include <linux/ctype.h>
22 #include <linux/kallsyms.h>
25 * Lock order:
26 * 1. slab_lock(page)
27 * 2. slab->list_lock
29 * The slab_lock protects operations on the object of a particular
30 * slab and its metadata in the page struct. If the slab lock
31 * has been taken then no allocations nor frees can be performed
32 * on the objects in the slab nor can the slab be added or removed
33 * from the partial or full lists since this would mean modifying
34 * the page_struct of the slab.
36 * The list_lock protects the partial and full list on each node and
37 * the partial slab counter. If taken then no new slabs may be added or
38 * removed from the lists nor make the number of partial slabs be modified.
39 * (Note that the total number of slabs is an atomic value that may be
40 * modified without taking the list lock).
42 * The list_lock is a centralized lock and thus we avoid taking it as
43 * much as possible. As long as SLUB does not have to handle partial
44 * slabs, operations can continue without any centralized lock. F.e.
45 * allocating a long series of objects that fill up slabs does not require
46 * the list lock.
48 * The lock order is sometimes inverted when we are trying to get a slab
49 * off a list. We take the list_lock and then look for a page on the list
50 * to use. While we do that objects in the slabs may be freed. We can
51 * only operate on the slab if we have also taken the slab_lock. So we use
52 * a slab_trylock() on the slab. If trylock was successful then no frees
53 * can occur anymore and we can use the slab for allocations etc. If the
54 * slab_trylock() does not succeed then frees are in progress in the slab and
55 * we must stay away from it for a while since we may cause a bouncing
56 * cacheline if we try to acquire the lock. So go onto the next slab.
57 * If all pages are busy then we may allocate a new slab instead of reusing
58 * a partial slab. A new slab has noone operating on it and thus there is
59 * no danger of cacheline contention.
61 * Interrupts are disabled during allocation and deallocation in order to
62 * make the slab allocator safe to use in the context of an irq. In addition
63 * interrupts are disabled to ensure that the processor does not change
64 * while handling per_cpu slabs, due to kernel preemption.
66 * SLUB assigns one slab for allocation to each processor.
67 * Allocations only occur from these slabs called cpu slabs.
69 * Slabs with free elements are kept on a partial list and during regular
70 * operations no list for full slabs is used. If an object in a full slab is
71 * freed then the slab will show up again on the partial lists.
72 * We track full slabs for debugging purposes though because otherwise we
73 * cannot scan all objects.
75 * Slabs are freed when they become empty. Teardown and setup is
76 * minimal so we rely on the page allocators per cpu caches for
77 * fast frees and allocs.
79 * Overloading of page flags that are otherwise used for LRU management.
81 * PageActive The slab is frozen and exempt from list processing.
82 * This means that the slab is dedicated to a purpose
83 * such as satisfying allocations for a specific
84 * processor. Objects may be freed in the slab while
85 * it is frozen but slab_free will then skip the usual
86 * list operations. It is up to the processor holding
87 * the slab to integrate the slab into the slab lists
88 * when the slab is no longer needed.
90 * One use of this flag is to mark slabs that are
91 * used for allocations. Then such a slab becomes a cpu
92 * slab. The cpu slab may be equipped with an additional
93 * lockless_freelist that allows lockless access to
94 * free objects in addition to the regular freelist
95 * that requires the slab lock.
97 * PageError Slab requires special handling due to debug
98 * options set. This moves slab handling out of
99 * the fast path and disables lockless freelists.
102 #define FROZEN (1 << PG_active)
104 #ifdef CONFIG_SLUB_DEBUG
105 #define SLABDEBUG (1 << PG_error)
106 #else
107 #define SLABDEBUG 0
108 #endif
110 static inline int SlabFrozen(struct page *page)
112 return page->flags & FROZEN;
115 static inline void SetSlabFrozen(struct page *page)
117 page->flags |= FROZEN;
120 static inline void ClearSlabFrozen(struct page *page)
122 page->flags &= ~FROZEN;
125 static inline int SlabDebug(struct page *page)
127 return page->flags & SLABDEBUG;
130 static inline void SetSlabDebug(struct page *page)
132 page->flags |= SLABDEBUG;
135 static inline void ClearSlabDebug(struct page *page)
137 page->flags &= ~SLABDEBUG;
141 * Issues still to be resolved:
143 * - The per cpu array is updated for each new slab and and is a remote
144 * cacheline for most nodes. This could become a bouncing cacheline given
145 * enough frequent updates. There are 16 pointers in a cacheline, so at
146 * max 16 cpus could compete for the cacheline which may be okay.
148 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
150 * - Variable sizing of the per node arrays
153 /* Enable to test recovery from slab corruption on boot */
154 #undef SLUB_RESILIENCY_TEST
156 #if PAGE_SHIFT <= 12
159 * Small page size. Make sure that we do not fragment memory
161 #define DEFAULT_MAX_ORDER 1
162 #define DEFAULT_MIN_OBJECTS 4
164 #else
167 * Large page machines are customarily able to handle larger
168 * page orders.
170 #define DEFAULT_MAX_ORDER 2
171 #define DEFAULT_MIN_OBJECTS 8
173 #endif
176 * Mininum number of partial slabs. These will be left on the partial
177 * lists even if they are empty. kmem_cache_shrink may reclaim them.
179 #define MIN_PARTIAL 2
182 * Maximum number of desirable partial slabs.
183 * The existence of more partial slabs makes kmem_cache_shrink
184 * sort the partial list by the number of objects in the.
186 #define MAX_PARTIAL 10
188 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
189 SLAB_POISON | SLAB_STORE_USER)
192 * Set of flags that will prevent slab merging
194 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
195 SLAB_TRACE | SLAB_DESTROY_BY_RCU)
197 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
198 SLAB_CACHE_DMA)
200 #ifndef ARCH_KMALLOC_MINALIGN
201 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
202 #endif
204 #ifndef ARCH_SLAB_MINALIGN
205 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
206 #endif
209 * The page->inuse field is 16 bit thus we have this limitation
211 #define MAX_OBJECTS_PER_SLAB 65535
213 /* Internal SLUB flags */
214 #define __OBJECT_POISON 0x80000000 /* Poison object */
215 #define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */
217 /* Not all arches define cache_line_size */
218 #ifndef cache_line_size
219 #define cache_line_size() L1_CACHE_BYTES
220 #endif
222 static int kmem_size = sizeof(struct kmem_cache);
224 #ifdef CONFIG_SMP
225 static struct notifier_block slab_notifier;
226 #endif
228 static enum {
229 DOWN, /* No slab functionality available */
230 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
231 UP, /* Everything works but does not show up in sysfs */
232 SYSFS /* Sysfs up */
233 } slab_state = DOWN;
235 /* A list of all slab caches on the system */
236 static DECLARE_RWSEM(slub_lock);
237 static LIST_HEAD(slab_caches);
240 * Tracking user of a slab.
242 struct track {
243 void *addr; /* Called from address */
244 int cpu; /* Was running on cpu */
245 int pid; /* Pid context */
246 unsigned long when; /* When did the operation occur */
249 enum track_item { TRACK_ALLOC, TRACK_FREE };
251 #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
252 static int sysfs_slab_add(struct kmem_cache *);
253 static int sysfs_slab_alias(struct kmem_cache *, const char *);
254 static void sysfs_slab_remove(struct kmem_cache *);
255 #else
256 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
257 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
258 { return 0; }
259 static inline void sysfs_slab_remove(struct kmem_cache *s) {}
260 #endif
262 /********************************************************************
263 * Core slab cache functions
264 *******************************************************************/
266 int slab_is_available(void)
268 return slab_state >= UP;
271 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
273 #ifdef CONFIG_NUMA
274 return s->node[node];
275 #else
276 return &s->local_node;
277 #endif
280 static inline int check_valid_pointer(struct kmem_cache *s,
281 struct page *page, const void *object)
283 void *base;
285 if (!object)
286 return 1;
288 base = page_address(page);
289 if (object < base || object >= base + s->objects * s->size ||
290 (object - base) % s->size) {
291 return 0;
294 return 1;
298 * Slow version of get and set free pointer.
300 * This version requires touching the cache lines of kmem_cache which
301 * we avoid to do in the fast alloc free paths. There we obtain the offset
302 * from the page struct.
304 static inline void *get_freepointer(struct kmem_cache *s, void *object)
306 return *(void **)(object + s->offset);
309 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
311 *(void **)(object + s->offset) = fp;
314 /* Loop over all objects in a slab */
315 #define for_each_object(__p, __s, __addr) \
316 for (__p = (__addr); __p < (__addr) + (__s)->objects * (__s)->size;\
317 __p += (__s)->size)
319 /* Scan freelist */
320 #define for_each_free_object(__p, __s, __free) \
321 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
323 /* Determine object index from a given position */
324 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
326 return (p - addr) / s->size;
329 #ifdef CONFIG_SLUB_DEBUG
331 * Debug settings:
333 #ifdef CONFIG_SLUB_DEBUG_ON
334 static int slub_debug = DEBUG_DEFAULT_FLAGS;
335 #else
336 static int slub_debug;
337 #endif
339 static char *slub_debug_slabs;
342 * Object debugging
344 static void print_section(char *text, u8 *addr, unsigned int length)
346 int i, offset;
347 int newline = 1;
348 char ascii[17];
350 ascii[16] = 0;
352 for (i = 0; i < length; i++) {
353 if (newline) {
354 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
355 newline = 0;
357 printk(" %02x", addr[i]);
358 offset = i % 16;
359 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
360 if (offset == 15) {
361 printk(" %s\n",ascii);
362 newline = 1;
365 if (!newline) {
366 i %= 16;
367 while (i < 16) {
368 printk(" ");
369 ascii[i] = ' ';
370 i++;
372 printk(" %s\n", ascii);
376 static struct track *get_track(struct kmem_cache *s, void *object,
377 enum track_item alloc)
379 struct track *p;
381 if (s->offset)
382 p = object + s->offset + sizeof(void *);
383 else
384 p = object + s->inuse;
386 return p + alloc;
389 static void set_track(struct kmem_cache *s, void *object,
390 enum track_item alloc, void *addr)
392 struct track *p;
394 if (s->offset)
395 p = object + s->offset + sizeof(void *);
396 else
397 p = object + s->inuse;
399 p += alloc;
400 if (addr) {
401 p->addr = addr;
402 p->cpu = smp_processor_id();
403 p->pid = current ? current->pid : -1;
404 p->when = jiffies;
405 } else
406 memset(p, 0, sizeof(struct track));
409 static void init_tracking(struct kmem_cache *s, void *object)
411 if (!(s->flags & SLAB_STORE_USER))
412 return;
414 set_track(s, object, TRACK_FREE, NULL);
415 set_track(s, object, TRACK_ALLOC, NULL);
418 static void print_track(const char *s, struct track *t)
420 if (!t->addr)
421 return;
423 printk(KERN_ERR "INFO: %s in ", s);
424 __print_symbol("%s", (unsigned long)t->addr);
425 printk(" age=%lu cpu=%u pid=%d\n", jiffies - t->when, t->cpu, t->pid);
428 static void print_tracking(struct kmem_cache *s, void *object)
430 if (!(s->flags & SLAB_STORE_USER))
431 return;
433 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
434 print_track("Freed", get_track(s, object, TRACK_FREE));
437 static void print_page_info(struct page *page)
439 printk(KERN_ERR "INFO: Slab 0x%p used=%u fp=0x%p flags=0x%04lx\n",
440 page, page->inuse, page->freelist, page->flags);
444 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
446 va_list args;
447 char buf[100];
449 va_start(args, fmt);
450 vsnprintf(buf, sizeof(buf), fmt, args);
451 va_end(args);
452 printk(KERN_ERR "========================================"
453 "=====================================\n");
454 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
455 printk(KERN_ERR "----------------------------------------"
456 "-------------------------------------\n\n");
459 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
461 va_list args;
462 char buf[100];
464 va_start(args, fmt);
465 vsnprintf(buf, sizeof(buf), fmt, args);
466 va_end(args);
467 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
470 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
472 unsigned int off; /* Offset of last byte */
473 u8 *addr = page_address(page);
475 print_tracking(s, p);
477 print_page_info(page);
479 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
480 p, p - addr, get_freepointer(s, p));
482 if (p > addr + 16)
483 print_section("Bytes b4", p - 16, 16);
485 print_section("Object", p, min(s->objsize, 128));
487 if (s->flags & SLAB_RED_ZONE)
488 print_section("Redzone", p + s->objsize,
489 s->inuse - s->objsize);
491 if (s->offset)
492 off = s->offset + sizeof(void *);
493 else
494 off = s->inuse;
496 if (s->flags & SLAB_STORE_USER)
497 off += 2 * sizeof(struct track);
499 if (off != s->size)
500 /* Beginning of the filler is the free pointer */
501 print_section("Padding", p + off, s->size - off);
503 dump_stack();
506 static void object_err(struct kmem_cache *s, struct page *page,
507 u8 *object, char *reason)
509 slab_bug(s, reason);
510 print_trailer(s, page, object);
513 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
515 va_list args;
516 char buf[100];
518 va_start(args, fmt);
519 vsnprintf(buf, sizeof(buf), fmt, args);
520 va_end(args);
521 slab_bug(s, fmt);
522 print_page_info(page);
523 dump_stack();
526 static void init_object(struct kmem_cache *s, void *object, int active)
528 u8 *p = object;
530 if (s->flags & __OBJECT_POISON) {
531 memset(p, POISON_FREE, s->objsize - 1);
532 p[s->objsize -1] = POISON_END;
535 if (s->flags & SLAB_RED_ZONE)
536 memset(p + s->objsize,
537 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
538 s->inuse - s->objsize);
541 static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
543 while (bytes) {
544 if (*start != (u8)value)
545 return start;
546 start++;
547 bytes--;
549 return NULL;
552 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
553 void *from, void *to)
555 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
556 memset(from, data, to - from);
559 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
560 u8 *object, char *what,
561 u8* start, unsigned int value, unsigned int bytes)
563 u8 *fault;
564 u8 *end;
566 fault = check_bytes(start, value, bytes);
567 if (!fault)
568 return 1;
570 end = start + bytes;
571 while (end > fault && end[-1] == value)
572 end--;
574 slab_bug(s, "%s overwritten", what);
575 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
576 fault, end - 1, fault[0], value);
577 print_trailer(s, page, object);
579 restore_bytes(s, what, value, fault, end);
580 return 0;
584 * Object layout:
586 * object address
587 * Bytes of the object to be managed.
588 * If the freepointer may overlay the object then the free
589 * pointer is the first word of the object.
591 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
592 * 0xa5 (POISON_END)
594 * object + s->objsize
595 * Padding to reach word boundary. This is also used for Redzoning.
596 * Padding is extended by another word if Redzoning is enabled and
597 * objsize == inuse.
599 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
600 * 0xcc (RED_ACTIVE) for objects in use.
602 * object + s->inuse
603 * Meta data starts here.
605 * A. Free pointer (if we cannot overwrite object on free)
606 * B. Tracking data for SLAB_STORE_USER
607 * C. Padding to reach required alignment boundary or at mininum
608 * one word if debuggin is on to be able to detect writes
609 * before the word boundary.
611 * Padding is done using 0x5a (POISON_INUSE)
613 * object + s->size
614 * Nothing is used beyond s->size.
616 * If slabcaches are merged then the objsize and inuse boundaries are mostly
617 * ignored. And therefore no slab options that rely on these boundaries
618 * may be used with merged slabcaches.
621 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
623 unsigned long off = s->inuse; /* The end of info */
625 if (s->offset)
626 /* Freepointer is placed after the object. */
627 off += sizeof(void *);
629 if (s->flags & SLAB_STORE_USER)
630 /* We also have user information there */
631 off += 2 * sizeof(struct track);
633 if (s->size == off)
634 return 1;
636 return check_bytes_and_report(s, page, p, "Object padding",
637 p + off, POISON_INUSE, s->size - off);
640 static int slab_pad_check(struct kmem_cache *s, struct page *page)
642 u8 *start;
643 u8 *fault;
644 u8 *end;
645 int length;
646 int remainder;
648 if (!(s->flags & SLAB_POISON))
649 return 1;
651 start = page_address(page);
652 end = start + (PAGE_SIZE << s->order);
653 length = s->objects * s->size;
654 remainder = end - (start + length);
655 if (!remainder)
656 return 1;
658 fault = check_bytes(start + length, POISON_INUSE, remainder);
659 if (!fault)
660 return 1;
661 while (end > fault && end[-1] == POISON_INUSE)
662 end--;
664 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
665 print_section("Padding", start, length);
667 restore_bytes(s, "slab padding", POISON_INUSE, start, end);
668 return 0;
671 static int check_object(struct kmem_cache *s, struct page *page,
672 void *object, int active)
674 u8 *p = object;
675 u8 *endobject = object + s->objsize;
677 if (s->flags & SLAB_RED_ZONE) {
678 unsigned int red =
679 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
681 if (!check_bytes_and_report(s, page, object, "Redzone",
682 endobject, red, s->inuse - s->objsize))
683 return 0;
684 } else {
685 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse)
686 check_bytes_and_report(s, page, p, "Alignment padding", endobject,
687 POISON_INUSE, s->inuse - s->objsize);
690 if (s->flags & SLAB_POISON) {
691 if (!active && (s->flags & __OBJECT_POISON) &&
692 (!check_bytes_and_report(s, page, p, "Poison", p,
693 POISON_FREE, s->objsize - 1) ||
694 !check_bytes_and_report(s, page, p, "Poison",
695 p + s->objsize -1, POISON_END, 1)))
696 return 0;
698 * check_pad_bytes cleans up on its own.
700 check_pad_bytes(s, page, p);
703 if (!s->offset && active)
705 * Object and freepointer overlap. Cannot check
706 * freepointer while object is allocated.
708 return 1;
710 /* Check free pointer validity */
711 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
712 object_err(s, page, p, "Freepointer corrupt");
714 * No choice but to zap it and thus loose the remainder
715 * of the free objects in this slab. May cause
716 * another error because the object count is now wrong.
718 set_freepointer(s, p, NULL);
719 return 0;
721 return 1;
724 static int check_slab(struct kmem_cache *s, struct page *page)
726 VM_BUG_ON(!irqs_disabled());
728 if (!PageSlab(page)) {
729 slab_err(s, page, "Not a valid slab page");
730 return 0;
732 if (page->offset * sizeof(void *) != s->offset) {
733 slab_err(s, page, "Corrupted offset %lu",
734 (unsigned long)(page->offset * sizeof(void *)));
735 return 0;
737 if (page->inuse > s->objects) {
738 slab_err(s, page, "inuse %u > max %u",
739 s->name, page->inuse, s->objects);
740 return 0;
742 /* Slab_pad_check fixes things up after itself */
743 slab_pad_check(s, page);
744 return 1;
748 * Determine if a certain object on a page is on the freelist. Must hold the
749 * slab lock to guarantee that the chains are in a consistent state.
751 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
753 int nr = 0;
754 void *fp = page->freelist;
755 void *object = NULL;
757 while (fp && nr <= s->objects) {
758 if (fp == search)
759 return 1;
760 if (!check_valid_pointer(s, page, fp)) {
761 if (object) {
762 object_err(s, page, object,
763 "Freechain corrupt");
764 set_freepointer(s, object, NULL);
765 break;
766 } else {
767 slab_err(s, page, "Freepointer corrupt");
768 page->freelist = NULL;
769 page->inuse = s->objects;
770 slab_fix(s, "Freelist cleared");
771 return 0;
773 break;
775 object = fp;
776 fp = get_freepointer(s, object);
777 nr++;
780 if (page->inuse != s->objects - nr) {
781 slab_err(s, page, "Wrong object count. Counter is %d but "
782 "counted were %d", page->inuse, s->objects - nr);
783 page->inuse = s->objects - nr;
784 slab_fix(s, "Object count adjusted.");
786 return search == NULL;
789 static void trace(struct kmem_cache *s, struct page *page, void *object, int alloc)
791 if (s->flags & SLAB_TRACE) {
792 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
793 s->name,
794 alloc ? "alloc" : "free",
795 object, page->inuse,
796 page->freelist);
798 if (!alloc)
799 print_section("Object", (void *)object, s->objsize);
801 dump_stack();
806 * Tracking of fully allocated slabs for debugging purposes.
808 static void add_full(struct kmem_cache_node *n, struct page *page)
810 spin_lock(&n->list_lock);
811 list_add(&page->lru, &n->full);
812 spin_unlock(&n->list_lock);
815 static void remove_full(struct kmem_cache *s, struct page *page)
817 struct kmem_cache_node *n;
819 if (!(s->flags & SLAB_STORE_USER))
820 return;
822 n = get_node(s, page_to_nid(page));
824 spin_lock(&n->list_lock);
825 list_del(&page->lru);
826 spin_unlock(&n->list_lock);
829 static void setup_object_debug(struct kmem_cache *s, struct page *page,
830 void *object)
832 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
833 return;
835 init_object(s, object, 0);
836 init_tracking(s, object);
839 static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
840 void *object, void *addr)
842 if (!check_slab(s, page))
843 goto bad;
845 if (object && !on_freelist(s, page, object)) {
846 object_err(s, page, object, "Object already allocated");
847 goto bad;
850 if (!check_valid_pointer(s, page, object)) {
851 object_err(s, page, object, "Freelist Pointer check fails");
852 goto bad;
855 if (object && !check_object(s, page, object, 0))
856 goto bad;
858 /* Success perform special debug activities for allocs */
859 if (s->flags & SLAB_STORE_USER)
860 set_track(s, object, TRACK_ALLOC, addr);
861 trace(s, page, object, 1);
862 init_object(s, object, 1);
863 return 1;
865 bad:
866 if (PageSlab(page)) {
868 * If this is a slab page then lets do the best we can
869 * to avoid issues in the future. Marking all objects
870 * as used avoids touching the remaining objects.
872 slab_fix(s, "Marking all objects used");
873 page->inuse = s->objects;
874 page->freelist = NULL;
875 /* Fix up fields that may be corrupted */
876 page->offset = s->offset / sizeof(void *);
878 return 0;
881 static int free_debug_processing(struct kmem_cache *s, struct page *page,
882 void *object, void *addr)
884 if (!check_slab(s, page))
885 goto fail;
887 if (!check_valid_pointer(s, page, object)) {
888 slab_err(s, page, "Invalid object pointer 0x%p", object);
889 goto fail;
892 if (on_freelist(s, page, object)) {
893 object_err(s, page, object, "Object already free");
894 goto fail;
897 if (!check_object(s, page, object, 1))
898 return 0;
900 if (unlikely(s != page->slab)) {
901 if (!PageSlab(page))
902 slab_err(s, page, "Attempt to free object(0x%p) "
903 "outside of slab", object);
904 else
905 if (!page->slab) {
906 printk(KERN_ERR
907 "SLUB <none>: no slab for object 0x%p.\n",
908 object);
909 dump_stack();
911 else
912 object_err(s, page, object,
913 "page slab pointer corrupt.");
914 goto fail;
917 /* Special debug activities for freeing objects */
918 if (!SlabFrozen(page) && !page->freelist)
919 remove_full(s, page);
920 if (s->flags & SLAB_STORE_USER)
921 set_track(s, object, TRACK_FREE, addr);
922 trace(s, page, object, 0);
923 init_object(s, object, 0);
924 return 1;
926 fail:
927 slab_fix(s, "Object at 0x%p not freed", object);
928 return 0;
931 static int __init setup_slub_debug(char *str)
933 slub_debug = DEBUG_DEFAULT_FLAGS;
934 if (*str++ != '=' || !*str)
936 * No options specified. Switch on full debugging.
938 goto out;
940 if (*str == ',')
942 * No options but restriction on slabs. This means full
943 * debugging for slabs matching a pattern.
945 goto check_slabs;
947 slub_debug = 0;
948 if (*str == '-')
950 * Switch off all debugging measures.
952 goto out;
955 * Determine which debug features should be switched on
957 for ( ;*str && *str != ','; str++) {
958 switch (tolower(*str)) {
959 case 'f':
960 slub_debug |= SLAB_DEBUG_FREE;
961 break;
962 case 'z':
963 slub_debug |= SLAB_RED_ZONE;
964 break;
965 case 'p':
966 slub_debug |= SLAB_POISON;
967 break;
968 case 'u':
969 slub_debug |= SLAB_STORE_USER;
970 break;
971 case 't':
972 slub_debug |= SLAB_TRACE;
973 break;
974 default:
975 printk(KERN_ERR "slub_debug option '%c' "
976 "unknown. skipped\n",*str);
980 check_slabs:
981 if (*str == ',')
982 slub_debug_slabs = str + 1;
983 out:
984 return 1;
987 __setup("slub_debug", setup_slub_debug);
989 static void kmem_cache_open_debug_check(struct kmem_cache *s)
992 * The page->offset field is only 16 bit wide. This is an offset
993 * in units of words from the beginning of an object. If the slab
994 * size is bigger then we cannot move the free pointer behind the
995 * object anymore.
997 * On 32 bit platforms the limit is 256k. On 64bit platforms
998 * the limit is 512k.
1000 * Debugging or ctor may create a need to move the free
1001 * pointer. Fail if this happens.
1003 if (s->objsize >= 65535 * sizeof(void *)) {
1004 BUG_ON(s->flags & (SLAB_RED_ZONE | SLAB_POISON |
1005 SLAB_STORE_USER | SLAB_DESTROY_BY_RCU));
1006 BUG_ON(s->ctor);
1008 else
1010 * Enable debugging if selected on the kernel commandline.
1012 if (slub_debug && (!slub_debug_slabs ||
1013 strncmp(slub_debug_slabs, s->name,
1014 strlen(slub_debug_slabs)) == 0))
1015 s->flags |= slub_debug;
1017 #else
1018 static inline void setup_object_debug(struct kmem_cache *s,
1019 struct page *page, void *object) {}
1021 static inline int alloc_debug_processing(struct kmem_cache *s,
1022 struct page *page, void *object, void *addr) { return 0; }
1024 static inline int free_debug_processing(struct kmem_cache *s,
1025 struct page *page, void *object, void *addr) { return 0; }
1027 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1028 { return 1; }
1029 static inline int check_object(struct kmem_cache *s, struct page *page,
1030 void *object, int active) { return 1; }
1031 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
1032 static inline void kmem_cache_open_debug_check(struct kmem_cache *s) {}
1033 #define slub_debug 0
1034 #endif
1036 * Slab allocation and freeing
1038 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1040 struct page * page;
1041 int pages = 1 << s->order;
1043 if (s->order)
1044 flags |= __GFP_COMP;
1046 if (s->flags & SLAB_CACHE_DMA)
1047 flags |= SLUB_DMA;
1049 if (node == -1)
1050 page = alloc_pages(flags, s->order);
1051 else
1052 page = alloc_pages_node(node, flags, s->order);
1054 if (!page)
1055 return NULL;
1057 mod_zone_page_state(page_zone(page),
1058 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1059 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1060 pages);
1062 return page;
1065 static void setup_object(struct kmem_cache *s, struct page *page,
1066 void *object)
1068 setup_object_debug(s, page, object);
1069 if (unlikely(s->ctor))
1070 s->ctor(object, s, 0);
1073 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1075 struct page *page;
1076 struct kmem_cache_node *n;
1077 void *start;
1078 void *end;
1079 void *last;
1080 void *p;
1082 BUG_ON(flags & ~(GFP_DMA | __GFP_ZERO | GFP_LEVEL_MASK));
1084 if (flags & __GFP_WAIT)
1085 local_irq_enable();
1087 page = allocate_slab(s, flags & GFP_LEVEL_MASK, node);
1088 if (!page)
1089 goto out;
1091 n = get_node(s, page_to_nid(page));
1092 if (n)
1093 atomic_long_inc(&n->nr_slabs);
1094 page->offset = s->offset / sizeof(void *);
1095 page->slab = s;
1096 page->flags |= 1 << PG_slab;
1097 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1098 SLAB_STORE_USER | SLAB_TRACE))
1099 SetSlabDebug(page);
1101 start = page_address(page);
1102 end = start + s->objects * s->size;
1104 if (unlikely(s->flags & SLAB_POISON))
1105 memset(start, POISON_INUSE, PAGE_SIZE << s->order);
1107 last = start;
1108 for_each_object(p, s, start) {
1109 setup_object(s, page, last);
1110 set_freepointer(s, last, p);
1111 last = p;
1113 setup_object(s, page, last);
1114 set_freepointer(s, last, NULL);
1116 page->freelist = start;
1117 page->lockless_freelist = NULL;
1118 page->inuse = 0;
1119 out:
1120 if (flags & __GFP_WAIT)
1121 local_irq_disable();
1122 return page;
1125 static void __free_slab(struct kmem_cache *s, struct page *page)
1127 int pages = 1 << s->order;
1129 if (unlikely(SlabDebug(page))) {
1130 void *p;
1132 slab_pad_check(s, page);
1133 for_each_object(p, s, page_address(page))
1134 check_object(s, page, p, 0);
1135 ClearSlabDebug(page);
1138 mod_zone_page_state(page_zone(page),
1139 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1140 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1141 - pages);
1143 page->mapping = NULL;
1144 __free_pages(page, s->order);
1147 static void rcu_free_slab(struct rcu_head *h)
1149 struct page *page;
1151 page = container_of((struct list_head *)h, struct page, lru);
1152 __free_slab(page->slab, page);
1155 static void free_slab(struct kmem_cache *s, struct page *page)
1157 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1159 * RCU free overloads the RCU head over the LRU
1161 struct rcu_head *head = (void *)&page->lru;
1163 call_rcu(head, rcu_free_slab);
1164 } else
1165 __free_slab(s, page);
1168 static void discard_slab(struct kmem_cache *s, struct page *page)
1170 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1172 atomic_long_dec(&n->nr_slabs);
1173 reset_page_mapcount(page);
1174 __ClearPageSlab(page);
1175 free_slab(s, page);
1179 * Per slab locking using the pagelock
1181 static __always_inline void slab_lock(struct page *page)
1183 bit_spin_lock(PG_locked, &page->flags);
1186 static __always_inline void slab_unlock(struct page *page)
1188 bit_spin_unlock(PG_locked, &page->flags);
1191 static __always_inline int slab_trylock(struct page *page)
1193 int rc = 1;
1195 rc = bit_spin_trylock(PG_locked, &page->flags);
1196 return rc;
1200 * Management of partially allocated slabs
1202 static void add_partial_tail(struct kmem_cache_node *n, struct page *page)
1204 spin_lock(&n->list_lock);
1205 n->nr_partial++;
1206 list_add_tail(&page->lru, &n->partial);
1207 spin_unlock(&n->list_lock);
1210 static void add_partial(struct kmem_cache_node *n, struct page *page)
1212 spin_lock(&n->list_lock);
1213 n->nr_partial++;
1214 list_add(&page->lru, &n->partial);
1215 spin_unlock(&n->list_lock);
1218 static void remove_partial(struct kmem_cache *s,
1219 struct page *page)
1221 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1223 spin_lock(&n->list_lock);
1224 list_del(&page->lru);
1225 n->nr_partial--;
1226 spin_unlock(&n->list_lock);
1230 * Lock slab and remove from the partial list.
1232 * Must hold list_lock.
1234 static inline int lock_and_freeze_slab(struct kmem_cache_node *n, struct page *page)
1236 if (slab_trylock(page)) {
1237 list_del(&page->lru);
1238 n->nr_partial--;
1239 SetSlabFrozen(page);
1240 return 1;
1242 return 0;
1246 * Try to allocate a partial slab from a specific node.
1248 static struct page *get_partial_node(struct kmem_cache_node *n)
1250 struct page *page;
1253 * Racy check. If we mistakenly see no partial slabs then we
1254 * just allocate an empty slab. If we mistakenly try to get a
1255 * partial slab and there is none available then get_partials()
1256 * will return NULL.
1258 if (!n || !n->nr_partial)
1259 return NULL;
1261 spin_lock(&n->list_lock);
1262 list_for_each_entry(page, &n->partial, lru)
1263 if (lock_and_freeze_slab(n, page))
1264 goto out;
1265 page = NULL;
1266 out:
1267 spin_unlock(&n->list_lock);
1268 return page;
1272 * Get a page from somewhere. Search in increasing NUMA distances.
1274 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1276 #ifdef CONFIG_NUMA
1277 struct zonelist *zonelist;
1278 struct zone **z;
1279 struct page *page;
1282 * The defrag ratio allows a configuration of the tradeoffs between
1283 * inter node defragmentation and node local allocations. A lower
1284 * defrag_ratio increases the tendency to do local allocations
1285 * instead of attempting to obtain partial slabs from other nodes.
1287 * If the defrag_ratio is set to 0 then kmalloc() always
1288 * returns node local objects. If the ratio is higher then kmalloc()
1289 * may return off node objects because partial slabs are obtained
1290 * from other nodes and filled up.
1292 * If /sys/slab/xx/defrag_ratio is set to 100 (which makes
1293 * defrag_ratio = 1000) then every (well almost) allocation will
1294 * first attempt to defrag slab caches on other nodes. This means
1295 * scanning over all nodes to look for partial slabs which may be
1296 * expensive if we do it every time we are trying to find a slab
1297 * with available objects.
1299 if (!s->defrag_ratio || get_cycles() % 1024 > s->defrag_ratio)
1300 return NULL;
1302 zonelist = &NODE_DATA(slab_node(current->mempolicy))
1303 ->node_zonelists[gfp_zone(flags)];
1304 for (z = zonelist->zones; *z; z++) {
1305 struct kmem_cache_node *n;
1307 n = get_node(s, zone_to_nid(*z));
1309 if (n && cpuset_zone_allowed_hardwall(*z, flags) &&
1310 n->nr_partial > MIN_PARTIAL) {
1311 page = get_partial_node(n);
1312 if (page)
1313 return page;
1316 #endif
1317 return NULL;
1321 * Get a partial page, lock it and return it.
1323 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1325 struct page *page;
1326 int searchnode = (node == -1) ? numa_node_id() : node;
1328 page = get_partial_node(get_node(s, searchnode));
1329 if (page || (flags & __GFP_THISNODE))
1330 return page;
1332 return get_any_partial(s, flags);
1336 * Move a page back to the lists.
1338 * Must be called with the slab lock held.
1340 * On exit the slab lock will have been dropped.
1342 static void unfreeze_slab(struct kmem_cache *s, struct page *page)
1344 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1346 ClearSlabFrozen(page);
1347 if (page->inuse) {
1349 if (page->freelist)
1350 add_partial(n, page);
1351 else if (SlabDebug(page) && (s->flags & SLAB_STORE_USER))
1352 add_full(n, page);
1353 slab_unlock(page);
1355 } else {
1356 if (n->nr_partial < MIN_PARTIAL) {
1358 * Adding an empty slab to the partial slabs in order
1359 * to avoid page allocator overhead. This slab needs
1360 * to come after the other slabs with objects in
1361 * order to fill them up. That way the size of the
1362 * partial list stays small. kmem_cache_shrink can
1363 * reclaim empty slabs from the partial list.
1365 add_partial_tail(n, page);
1366 slab_unlock(page);
1367 } else {
1368 slab_unlock(page);
1369 discard_slab(s, page);
1375 * Remove the cpu slab
1377 static void deactivate_slab(struct kmem_cache *s, struct page *page, int cpu)
1380 * Merge cpu freelist into freelist. Typically we get here
1381 * because both freelists are empty. So this is unlikely
1382 * to occur.
1384 while (unlikely(page->lockless_freelist)) {
1385 void **object;
1387 /* Retrieve object from cpu_freelist */
1388 object = page->lockless_freelist;
1389 page->lockless_freelist = page->lockless_freelist[page->offset];
1391 /* And put onto the regular freelist */
1392 object[page->offset] = page->freelist;
1393 page->freelist = object;
1394 page->inuse--;
1396 s->cpu_slab[cpu] = NULL;
1397 unfreeze_slab(s, page);
1400 static inline void flush_slab(struct kmem_cache *s, struct page *page, int cpu)
1402 slab_lock(page);
1403 deactivate_slab(s, page, cpu);
1407 * Flush cpu slab.
1408 * Called from IPI handler with interrupts disabled.
1410 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1412 struct page *page = s->cpu_slab[cpu];
1414 if (likely(page))
1415 flush_slab(s, page, cpu);
1418 static void flush_cpu_slab(void *d)
1420 struct kmem_cache *s = d;
1421 int cpu = smp_processor_id();
1423 __flush_cpu_slab(s, cpu);
1426 static void flush_all(struct kmem_cache *s)
1428 #ifdef CONFIG_SMP
1429 on_each_cpu(flush_cpu_slab, s, 1, 1);
1430 #else
1431 unsigned long flags;
1433 local_irq_save(flags);
1434 flush_cpu_slab(s);
1435 local_irq_restore(flags);
1436 #endif
1440 * Slow path. The lockless freelist is empty or we need to perform
1441 * debugging duties.
1443 * Interrupts are disabled.
1445 * Processing is still very fast if new objects have been freed to the
1446 * regular freelist. In that case we simply take over the regular freelist
1447 * as the lockless freelist and zap the regular freelist.
1449 * If that is not working then we fall back to the partial lists. We take the
1450 * first element of the freelist as the object to allocate now and move the
1451 * rest of the freelist to the lockless freelist.
1453 * And if we were unable to get a new slab from the partial slab lists then
1454 * we need to allocate a new slab. This is slowest path since we may sleep.
1456 static void *__slab_alloc(struct kmem_cache *s,
1457 gfp_t gfpflags, int node, void *addr, struct page *page)
1459 void **object;
1460 int cpu = smp_processor_id();
1462 if (!page)
1463 goto new_slab;
1465 slab_lock(page);
1466 if (unlikely(node != -1 && page_to_nid(page) != node))
1467 goto another_slab;
1468 load_freelist:
1469 object = page->freelist;
1470 if (unlikely(!object))
1471 goto another_slab;
1472 if (unlikely(SlabDebug(page)))
1473 goto debug;
1475 object = page->freelist;
1476 page->lockless_freelist = object[page->offset];
1477 page->inuse = s->objects;
1478 page->freelist = NULL;
1479 slab_unlock(page);
1480 return object;
1482 another_slab:
1483 deactivate_slab(s, page, cpu);
1485 new_slab:
1486 page = get_partial(s, gfpflags, node);
1487 if (page) {
1488 s->cpu_slab[cpu] = page;
1489 goto load_freelist;
1492 page = new_slab(s, gfpflags, node);
1493 if (page) {
1494 cpu = smp_processor_id();
1495 if (s->cpu_slab[cpu]) {
1497 * Someone else populated the cpu_slab while we
1498 * enabled interrupts, or we have gotten scheduled
1499 * on another cpu. The page may not be on the
1500 * requested node even if __GFP_THISNODE was
1501 * specified. So we need to recheck.
1503 if (node == -1 ||
1504 page_to_nid(s->cpu_slab[cpu]) == node) {
1506 * Current cpuslab is acceptable and we
1507 * want the current one since its cache hot
1509 discard_slab(s, page);
1510 page = s->cpu_slab[cpu];
1511 slab_lock(page);
1512 goto load_freelist;
1514 /* New slab does not fit our expectations */
1515 flush_slab(s, s->cpu_slab[cpu], cpu);
1517 slab_lock(page);
1518 SetSlabFrozen(page);
1519 s->cpu_slab[cpu] = page;
1520 goto load_freelist;
1522 return NULL;
1523 debug:
1524 object = page->freelist;
1525 if (!alloc_debug_processing(s, page, object, addr))
1526 goto another_slab;
1528 page->inuse++;
1529 page->freelist = object[page->offset];
1530 slab_unlock(page);
1531 return object;
1535 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1536 * have the fastpath folded into their functions. So no function call
1537 * overhead for requests that can be satisfied on the fastpath.
1539 * The fastpath works by first checking if the lockless freelist can be used.
1540 * If not then __slab_alloc is called for slow processing.
1542 * Otherwise we can simply pick the next object from the lockless free list.
1544 static void __always_inline *slab_alloc(struct kmem_cache *s,
1545 gfp_t gfpflags, int node, void *addr)
1547 struct page *page;
1548 void **object;
1549 unsigned long flags;
1551 local_irq_save(flags);
1552 page = s->cpu_slab[smp_processor_id()];
1553 if (unlikely(!page || !page->lockless_freelist ||
1554 (node != -1 && page_to_nid(page) != node)))
1556 object = __slab_alloc(s, gfpflags, node, addr, page);
1558 else {
1559 object = page->lockless_freelist;
1560 page->lockless_freelist = object[page->offset];
1562 local_irq_restore(flags);
1564 if (unlikely((gfpflags & __GFP_ZERO) && object))
1565 memset(object, 0, s->objsize);
1567 return object;
1570 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1572 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0));
1574 EXPORT_SYMBOL(kmem_cache_alloc);
1576 #ifdef CONFIG_NUMA
1577 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1579 return slab_alloc(s, gfpflags, node, __builtin_return_address(0));
1581 EXPORT_SYMBOL(kmem_cache_alloc_node);
1582 #endif
1585 * Slow patch handling. This may still be called frequently since objects
1586 * have a longer lifetime than the cpu slabs in most processing loads.
1588 * So we still attempt to reduce cache line usage. Just take the slab
1589 * lock and free the item. If there is no additional partial page
1590 * handling required then we can return immediately.
1592 static void __slab_free(struct kmem_cache *s, struct page *page,
1593 void *x, void *addr)
1595 void *prior;
1596 void **object = (void *)x;
1598 slab_lock(page);
1600 if (unlikely(SlabDebug(page)))
1601 goto debug;
1602 checks_ok:
1603 prior = object[page->offset] = page->freelist;
1604 page->freelist = object;
1605 page->inuse--;
1607 if (unlikely(SlabFrozen(page)))
1608 goto out_unlock;
1610 if (unlikely(!page->inuse))
1611 goto slab_empty;
1614 * Objects left in the slab. If it
1615 * was not on the partial list before
1616 * then add it.
1618 if (unlikely(!prior))
1619 add_partial(get_node(s, page_to_nid(page)), page);
1621 out_unlock:
1622 slab_unlock(page);
1623 return;
1625 slab_empty:
1626 if (prior)
1628 * Slab still on the partial list.
1630 remove_partial(s, page);
1632 slab_unlock(page);
1633 discard_slab(s, page);
1634 return;
1636 debug:
1637 if (!free_debug_processing(s, page, x, addr))
1638 goto out_unlock;
1639 goto checks_ok;
1643 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1644 * can perform fastpath freeing without additional function calls.
1646 * The fastpath is only possible if we are freeing to the current cpu slab
1647 * of this processor. This typically the case if we have just allocated
1648 * the item before.
1650 * If fastpath is not possible then fall back to __slab_free where we deal
1651 * with all sorts of special processing.
1653 static void __always_inline slab_free(struct kmem_cache *s,
1654 struct page *page, void *x, void *addr)
1656 void **object = (void *)x;
1657 unsigned long flags;
1659 local_irq_save(flags);
1660 debug_check_no_locks_freed(object, s->objsize);
1661 if (likely(page == s->cpu_slab[smp_processor_id()] &&
1662 !SlabDebug(page))) {
1663 object[page->offset] = page->lockless_freelist;
1664 page->lockless_freelist = object;
1665 } else
1666 __slab_free(s, page, x, addr);
1668 local_irq_restore(flags);
1671 void kmem_cache_free(struct kmem_cache *s, void *x)
1673 struct page *page;
1675 page = virt_to_head_page(x);
1677 slab_free(s, page, x, __builtin_return_address(0));
1679 EXPORT_SYMBOL(kmem_cache_free);
1681 /* Figure out on which slab object the object resides */
1682 static struct page *get_object_page(const void *x)
1684 struct page *page = virt_to_head_page(x);
1686 if (!PageSlab(page))
1687 return NULL;
1689 return page;
1693 * Object placement in a slab is made very easy because we always start at
1694 * offset 0. If we tune the size of the object to the alignment then we can
1695 * get the required alignment by putting one properly sized object after
1696 * another.
1698 * Notice that the allocation order determines the sizes of the per cpu
1699 * caches. Each processor has always one slab available for allocations.
1700 * Increasing the allocation order reduces the number of times that slabs
1701 * must be moved on and off the partial lists and is therefore a factor in
1702 * locking overhead.
1706 * Mininum / Maximum order of slab pages. This influences locking overhead
1707 * and slab fragmentation. A higher order reduces the number of partial slabs
1708 * and increases the number of allocations possible without having to
1709 * take the list_lock.
1711 static int slub_min_order;
1712 static int slub_max_order = DEFAULT_MAX_ORDER;
1713 static int slub_min_objects = DEFAULT_MIN_OBJECTS;
1716 * Merge control. If this is set then no merging of slab caches will occur.
1717 * (Could be removed. This was introduced to pacify the merge skeptics.)
1719 static int slub_nomerge;
1722 * Calculate the order of allocation given an slab object size.
1724 * The order of allocation has significant impact on performance and other
1725 * system components. Generally order 0 allocations should be preferred since
1726 * order 0 does not cause fragmentation in the page allocator. Larger objects
1727 * be problematic to put into order 0 slabs because there may be too much
1728 * unused space left. We go to a higher order if more than 1/8th of the slab
1729 * would be wasted.
1731 * In order to reach satisfactory performance we must ensure that a minimum
1732 * number of objects is in one slab. Otherwise we may generate too much
1733 * activity on the partial lists which requires taking the list_lock. This is
1734 * less a concern for large slabs though which are rarely used.
1736 * slub_max_order specifies the order where we begin to stop considering the
1737 * number of objects in a slab as critical. If we reach slub_max_order then
1738 * we try to keep the page order as low as possible. So we accept more waste
1739 * of space in favor of a small page order.
1741 * Higher order allocations also allow the placement of more objects in a
1742 * slab and thereby reduce object handling overhead. If the user has
1743 * requested a higher mininum order then we start with that one instead of
1744 * the smallest order which will fit the object.
1746 static inline int slab_order(int size, int min_objects,
1747 int max_order, int fract_leftover)
1749 int order;
1750 int rem;
1751 int min_order = slub_min_order;
1754 * If we would create too many object per slab then reduce
1755 * the slab order even if it goes below slub_min_order.
1757 while (min_order > 0 &&
1758 (PAGE_SIZE << min_order) >= MAX_OBJECTS_PER_SLAB * size)
1759 min_order--;
1761 for (order = max(min_order,
1762 fls(min_objects * size - 1) - PAGE_SHIFT);
1763 order <= max_order; order++) {
1765 unsigned long slab_size = PAGE_SIZE << order;
1767 if (slab_size < min_objects * size)
1768 continue;
1770 rem = slab_size % size;
1772 if (rem <= slab_size / fract_leftover)
1773 break;
1775 /* If the next size is too high then exit now */
1776 if (slab_size * 2 >= MAX_OBJECTS_PER_SLAB * size)
1777 break;
1780 return order;
1783 static inline int calculate_order(int size)
1785 int order;
1786 int min_objects;
1787 int fraction;
1790 * Attempt to find best configuration for a slab. This
1791 * works by first attempting to generate a layout with
1792 * the best configuration and backing off gradually.
1794 * First we reduce the acceptable waste in a slab. Then
1795 * we reduce the minimum objects required in a slab.
1797 min_objects = slub_min_objects;
1798 while (min_objects > 1) {
1799 fraction = 8;
1800 while (fraction >= 4) {
1801 order = slab_order(size, min_objects,
1802 slub_max_order, fraction);
1803 if (order <= slub_max_order)
1804 return order;
1805 fraction /= 2;
1807 min_objects /= 2;
1811 * We were unable to place multiple objects in a slab. Now
1812 * lets see if we can place a single object there.
1814 order = slab_order(size, 1, slub_max_order, 1);
1815 if (order <= slub_max_order)
1816 return order;
1819 * Doh this slab cannot be placed using slub_max_order.
1821 order = slab_order(size, 1, MAX_ORDER, 1);
1822 if (order <= MAX_ORDER)
1823 return order;
1824 return -ENOSYS;
1828 * Figure out what the alignment of the objects will be.
1830 static unsigned long calculate_alignment(unsigned long flags,
1831 unsigned long align, unsigned long size)
1834 * If the user wants hardware cache aligned objects then
1835 * follow that suggestion if the object is sufficiently
1836 * large.
1838 * The hardware cache alignment cannot override the
1839 * specified alignment though. If that is greater
1840 * then use it.
1842 if ((flags & SLAB_HWCACHE_ALIGN) &&
1843 size > cache_line_size() / 2)
1844 return max_t(unsigned long, align, cache_line_size());
1846 if (align < ARCH_SLAB_MINALIGN)
1847 return ARCH_SLAB_MINALIGN;
1849 return ALIGN(align, sizeof(void *));
1852 static void init_kmem_cache_node(struct kmem_cache_node *n)
1854 n->nr_partial = 0;
1855 atomic_long_set(&n->nr_slabs, 0);
1856 spin_lock_init(&n->list_lock);
1857 INIT_LIST_HEAD(&n->partial);
1858 #ifdef CONFIG_SLUB_DEBUG
1859 INIT_LIST_HEAD(&n->full);
1860 #endif
1863 #ifdef CONFIG_NUMA
1865 * No kmalloc_node yet so do it by hand. We know that this is the first
1866 * slab on the node for this slabcache. There are no concurrent accesses
1867 * possible.
1869 * Note that this function only works on the kmalloc_node_cache
1870 * when allocating for the kmalloc_node_cache.
1872 static struct kmem_cache_node * __init early_kmem_cache_node_alloc(gfp_t gfpflags,
1873 int node)
1875 struct page *page;
1876 struct kmem_cache_node *n;
1878 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
1880 page = new_slab(kmalloc_caches, gfpflags, node);
1882 BUG_ON(!page);
1883 if (page_to_nid(page) != node) {
1884 printk(KERN_ERR "SLUB: Unable to allocate memory from "
1885 "node %d\n", node);
1886 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
1887 "in order to be able to continue\n");
1890 n = page->freelist;
1891 BUG_ON(!n);
1892 page->freelist = get_freepointer(kmalloc_caches, n);
1893 page->inuse++;
1894 kmalloc_caches->node[node] = n;
1895 #ifdef CONFIG_SLUB_DEBUG
1896 init_object(kmalloc_caches, n, 1);
1897 init_tracking(kmalloc_caches, n);
1898 #endif
1899 init_kmem_cache_node(n);
1900 atomic_long_inc(&n->nr_slabs);
1901 add_partial(n, page);
1904 * new_slab() disables interupts. If we do not reenable interrupts here
1905 * then bootup would continue with interrupts disabled.
1907 local_irq_enable();
1908 return n;
1911 static void free_kmem_cache_nodes(struct kmem_cache *s)
1913 int node;
1915 for_each_online_node(node) {
1916 struct kmem_cache_node *n = s->node[node];
1917 if (n && n != &s->local_node)
1918 kmem_cache_free(kmalloc_caches, n);
1919 s->node[node] = NULL;
1923 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
1925 int node;
1926 int local_node;
1928 if (slab_state >= UP)
1929 local_node = page_to_nid(virt_to_page(s));
1930 else
1931 local_node = 0;
1933 for_each_online_node(node) {
1934 struct kmem_cache_node *n;
1936 if (local_node == node)
1937 n = &s->local_node;
1938 else {
1939 if (slab_state == DOWN) {
1940 n = early_kmem_cache_node_alloc(gfpflags,
1941 node);
1942 continue;
1944 n = kmem_cache_alloc_node(kmalloc_caches,
1945 gfpflags, node);
1947 if (!n) {
1948 free_kmem_cache_nodes(s);
1949 return 0;
1953 s->node[node] = n;
1954 init_kmem_cache_node(n);
1956 return 1;
1958 #else
1959 static void free_kmem_cache_nodes(struct kmem_cache *s)
1963 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
1965 init_kmem_cache_node(&s->local_node);
1966 return 1;
1968 #endif
1971 * calculate_sizes() determines the order and the distribution of data within
1972 * a slab object.
1974 static int calculate_sizes(struct kmem_cache *s)
1976 unsigned long flags = s->flags;
1977 unsigned long size = s->objsize;
1978 unsigned long align = s->align;
1981 * Determine if we can poison the object itself. If the user of
1982 * the slab may touch the object after free or before allocation
1983 * then we should never poison the object itself.
1985 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
1986 !s->ctor)
1987 s->flags |= __OBJECT_POISON;
1988 else
1989 s->flags &= ~__OBJECT_POISON;
1992 * Round up object size to the next word boundary. We can only
1993 * place the free pointer at word boundaries and this determines
1994 * the possible location of the free pointer.
1996 size = ALIGN(size, sizeof(void *));
1998 #ifdef CONFIG_SLUB_DEBUG
2000 * If we are Redzoning then check if there is some space between the
2001 * end of the object and the free pointer. If not then add an
2002 * additional word to have some bytes to store Redzone information.
2004 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2005 size += sizeof(void *);
2006 #endif
2009 * With that we have determined the number of bytes in actual use
2010 * by the object. This is the potential offset to the free pointer.
2012 s->inuse = size;
2014 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2015 s->ctor)) {
2017 * Relocate free pointer after the object if it is not
2018 * permitted to overwrite the first word of the object on
2019 * kmem_cache_free.
2021 * This is the case if we do RCU, have a constructor or
2022 * destructor or are poisoning the objects.
2024 s->offset = size;
2025 size += sizeof(void *);
2028 #ifdef CONFIG_SLUB_DEBUG
2029 if (flags & SLAB_STORE_USER)
2031 * Need to store information about allocs and frees after
2032 * the object.
2034 size += 2 * sizeof(struct track);
2036 if (flags & SLAB_RED_ZONE)
2038 * Add some empty padding so that we can catch
2039 * overwrites from earlier objects rather than let
2040 * tracking information or the free pointer be
2041 * corrupted if an user writes before the start
2042 * of the object.
2044 size += sizeof(void *);
2045 #endif
2048 * Determine the alignment based on various parameters that the
2049 * user specified and the dynamic determination of cache line size
2050 * on bootup.
2052 align = calculate_alignment(flags, align, s->objsize);
2055 * SLUB stores one object immediately after another beginning from
2056 * offset 0. In order to align the objects we have to simply size
2057 * each object to conform to the alignment.
2059 size = ALIGN(size, align);
2060 s->size = size;
2062 s->order = calculate_order(size);
2063 if (s->order < 0)
2064 return 0;
2067 * Determine the number of objects per slab
2069 s->objects = (PAGE_SIZE << s->order) / size;
2072 * Verify that the number of objects is within permitted limits.
2073 * The page->inuse field is only 16 bit wide! So we cannot have
2074 * more than 64k objects per slab.
2076 if (!s->objects || s->objects > MAX_OBJECTS_PER_SLAB)
2077 return 0;
2078 return 1;
2082 static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2083 const char *name, size_t size,
2084 size_t align, unsigned long flags,
2085 void (*ctor)(void *, struct kmem_cache *, unsigned long))
2087 memset(s, 0, kmem_size);
2088 s->name = name;
2089 s->ctor = ctor;
2090 s->objsize = size;
2091 s->flags = flags;
2092 s->align = align;
2093 kmem_cache_open_debug_check(s);
2095 if (!calculate_sizes(s))
2096 goto error;
2098 s->refcount = 1;
2099 #ifdef CONFIG_NUMA
2100 s->defrag_ratio = 100;
2101 #endif
2103 if (init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2104 return 1;
2105 error:
2106 if (flags & SLAB_PANIC)
2107 panic("Cannot create slab %s size=%lu realsize=%u "
2108 "order=%u offset=%u flags=%lx\n",
2109 s->name, (unsigned long)size, s->size, s->order,
2110 s->offset, flags);
2111 return 0;
2115 * Check if a given pointer is valid
2117 int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2119 struct page * page;
2121 page = get_object_page(object);
2123 if (!page || s != page->slab)
2124 /* No slab or wrong slab */
2125 return 0;
2127 if (!check_valid_pointer(s, page, object))
2128 return 0;
2131 * We could also check if the object is on the slabs freelist.
2132 * But this would be too expensive and it seems that the main
2133 * purpose of kmem_ptr_valid is to check if the object belongs
2134 * to a certain slab.
2136 return 1;
2138 EXPORT_SYMBOL(kmem_ptr_validate);
2141 * Determine the size of a slab object
2143 unsigned int kmem_cache_size(struct kmem_cache *s)
2145 return s->objsize;
2147 EXPORT_SYMBOL(kmem_cache_size);
2149 const char *kmem_cache_name(struct kmem_cache *s)
2151 return s->name;
2153 EXPORT_SYMBOL(kmem_cache_name);
2156 * Attempt to free all slabs on a node. Return the number of slabs we
2157 * were unable to free.
2159 static int free_list(struct kmem_cache *s, struct kmem_cache_node *n,
2160 struct list_head *list)
2162 int slabs_inuse = 0;
2163 unsigned long flags;
2164 struct page *page, *h;
2166 spin_lock_irqsave(&n->list_lock, flags);
2167 list_for_each_entry_safe(page, h, list, lru)
2168 if (!page->inuse) {
2169 list_del(&page->lru);
2170 discard_slab(s, page);
2171 } else
2172 slabs_inuse++;
2173 spin_unlock_irqrestore(&n->list_lock, flags);
2174 return slabs_inuse;
2178 * Release all resources used by a slab cache.
2180 static inline int kmem_cache_close(struct kmem_cache *s)
2182 int node;
2184 flush_all(s);
2186 /* Attempt to free all objects */
2187 for_each_online_node(node) {
2188 struct kmem_cache_node *n = get_node(s, node);
2190 n->nr_partial -= free_list(s, n, &n->partial);
2191 if (atomic_long_read(&n->nr_slabs))
2192 return 1;
2194 free_kmem_cache_nodes(s);
2195 return 0;
2199 * Close a cache and release the kmem_cache structure
2200 * (must be used for caches created using kmem_cache_create)
2202 void kmem_cache_destroy(struct kmem_cache *s)
2204 down_write(&slub_lock);
2205 s->refcount--;
2206 if (!s->refcount) {
2207 list_del(&s->list);
2208 up_write(&slub_lock);
2209 if (kmem_cache_close(s))
2210 WARN_ON(1);
2211 sysfs_slab_remove(s);
2212 kfree(s);
2213 } else
2214 up_write(&slub_lock);
2216 EXPORT_SYMBOL(kmem_cache_destroy);
2218 /********************************************************************
2219 * Kmalloc subsystem
2220 *******************************************************************/
2222 struct kmem_cache kmalloc_caches[KMALLOC_SHIFT_HIGH + 1] __cacheline_aligned;
2223 EXPORT_SYMBOL(kmalloc_caches);
2225 #ifdef CONFIG_ZONE_DMA
2226 static struct kmem_cache *kmalloc_caches_dma[KMALLOC_SHIFT_HIGH + 1];
2227 #endif
2229 static int __init setup_slub_min_order(char *str)
2231 get_option (&str, &slub_min_order);
2233 return 1;
2236 __setup("slub_min_order=", setup_slub_min_order);
2238 static int __init setup_slub_max_order(char *str)
2240 get_option (&str, &slub_max_order);
2242 return 1;
2245 __setup("slub_max_order=", setup_slub_max_order);
2247 static int __init setup_slub_min_objects(char *str)
2249 get_option (&str, &slub_min_objects);
2251 return 1;
2254 __setup("slub_min_objects=", setup_slub_min_objects);
2256 static int __init setup_slub_nomerge(char *str)
2258 slub_nomerge = 1;
2259 return 1;
2262 __setup("slub_nomerge", setup_slub_nomerge);
2264 static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2265 const char *name, int size, gfp_t gfp_flags)
2267 unsigned int flags = 0;
2269 if (gfp_flags & SLUB_DMA)
2270 flags = SLAB_CACHE_DMA;
2272 down_write(&slub_lock);
2273 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
2274 flags, NULL))
2275 goto panic;
2277 list_add(&s->list, &slab_caches);
2278 up_write(&slub_lock);
2279 if (sysfs_slab_add(s))
2280 goto panic;
2281 return s;
2283 panic:
2284 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2287 #ifdef CONFIG_ZONE_DMA
2289 static void sysfs_add_func(struct work_struct *w)
2291 struct kmem_cache *s;
2293 down_write(&slub_lock);
2294 list_for_each_entry(s, &slab_caches, list) {
2295 if (s->flags & __SYSFS_ADD_DEFERRED) {
2296 s->flags &= ~__SYSFS_ADD_DEFERRED;
2297 sysfs_slab_add(s);
2300 up_write(&slub_lock);
2303 static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2305 static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2307 struct kmem_cache *s;
2308 char *text;
2309 size_t realsize;
2311 s = kmalloc_caches_dma[index];
2312 if (s)
2313 return s;
2315 /* Dynamically create dma cache */
2316 if (flags & __GFP_WAIT)
2317 down_write(&slub_lock);
2318 else {
2319 if (!down_write_trylock(&slub_lock))
2320 goto out;
2323 if (kmalloc_caches_dma[index])
2324 goto unlock_out;
2326 realsize = kmalloc_caches[index].objsize;
2327 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d", (unsigned int)realsize),
2328 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2330 if (!s || !text || !kmem_cache_open(s, flags, text,
2331 realsize, ARCH_KMALLOC_MINALIGN,
2332 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) {
2333 kfree(s);
2334 kfree(text);
2335 goto unlock_out;
2338 list_add(&s->list, &slab_caches);
2339 kmalloc_caches_dma[index] = s;
2341 schedule_work(&sysfs_add_work);
2343 unlock_out:
2344 up_write(&slub_lock);
2345 out:
2346 return kmalloc_caches_dma[index];
2348 #endif
2351 * Conversion table for small slabs sizes / 8 to the index in the
2352 * kmalloc array. This is necessary for slabs < 192 since we have non power
2353 * of two cache sizes there. The size of larger slabs can be determined using
2354 * fls.
2356 static s8 size_index[24] = {
2357 3, /* 8 */
2358 4, /* 16 */
2359 5, /* 24 */
2360 5, /* 32 */
2361 6, /* 40 */
2362 6, /* 48 */
2363 6, /* 56 */
2364 6, /* 64 */
2365 1, /* 72 */
2366 1, /* 80 */
2367 1, /* 88 */
2368 1, /* 96 */
2369 7, /* 104 */
2370 7, /* 112 */
2371 7, /* 120 */
2372 7, /* 128 */
2373 2, /* 136 */
2374 2, /* 144 */
2375 2, /* 152 */
2376 2, /* 160 */
2377 2, /* 168 */
2378 2, /* 176 */
2379 2, /* 184 */
2380 2 /* 192 */
2383 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2385 int index;
2387 if (size <= 192) {
2388 if (!size)
2389 return ZERO_SIZE_PTR;
2391 index = size_index[(size - 1) / 8];
2392 } else {
2393 if (size > KMALLOC_MAX_SIZE)
2394 return NULL;
2396 index = fls(size - 1);
2399 #ifdef CONFIG_ZONE_DMA
2400 if (unlikely((flags & SLUB_DMA)))
2401 return dma_kmalloc_cache(index, flags);
2403 #endif
2404 return &kmalloc_caches[index];
2407 void *__kmalloc(size_t size, gfp_t flags)
2409 struct kmem_cache *s = get_slab(size, flags);
2411 if (ZERO_OR_NULL_PTR(s))
2412 return s;
2414 return slab_alloc(s, flags, -1, __builtin_return_address(0));
2416 EXPORT_SYMBOL(__kmalloc);
2418 #ifdef CONFIG_NUMA
2419 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2421 struct kmem_cache *s = get_slab(size, flags);
2423 if (ZERO_OR_NULL_PTR(s))
2424 return s;
2426 return slab_alloc(s, flags, node, __builtin_return_address(0));
2428 EXPORT_SYMBOL(__kmalloc_node);
2429 #endif
2431 size_t ksize(const void *object)
2433 struct page *page;
2434 struct kmem_cache *s;
2436 if (ZERO_OR_NULL_PTR(object))
2437 return 0;
2439 page = get_object_page(object);
2440 BUG_ON(!page);
2441 s = page->slab;
2442 BUG_ON(!s);
2445 * Debugging requires use of the padding between object
2446 * and whatever may come after it.
2448 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2449 return s->objsize;
2452 * If we have the need to store the freelist pointer
2453 * back there or track user information then we can
2454 * only use the space before that information.
2456 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2457 return s->inuse;
2460 * Else we can use all the padding etc for the allocation
2462 return s->size;
2464 EXPORT_SYMBOL(ksize);
2466 void kfree(const void *x)
2468 struct kmem_cache *s;
2469 struct page *page;
2472 * This has to be an unsigned comparison. According to Linus
2473 * some gcc version treat a pointer as a signed entity. Then
2474 * this comparison would be true for all "negative" pointers
2475 * (which would cover the whole upper half of the address space).
2477 if (ZERO_OR_NULL_PTR(x))
2478 return;
2480 page = virt_to_head_page(x);
2481 s = page->slab;
2483 slab_free(s, page, (void *)x, __builtin_return_address(0));
2485 EXPORT_SYMBOL(kfree);
2488 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2489 * the remaining slabs by the number of items in use. The slabs with the
2490 * most items in use come first. New allocations will then fill those up
2491 * and thus they can be removed from the partial lists.
2493 * The slabs with the least items are placed last. This results in them
2494 * being allocated from last increasing the chance that the last objects
2495 * are freed in them.
2497 int kmem_cache_shrink(struct kmem_cache *s)
2499 int node;
2500 int i;
2501 struct kmem_cache_node *n;
2502 struct page *page;
2503 struct page *t;
2504 struct list_head *slabs_by_inuse =
2505 kmalloc(sizeof(struct list_head) * s->objects, GFP_KERNEL);
2506 unsigned long flags;
2508 if (!slabs_by_inuse)
2509 return -ENOMEM;
2511 flush_all(s);
2512 for_each_online_node(node) {
2513 n = get_node(s, node);
2515 if (!n->nr_partial)
2516 continue;
2518 for (i = 0; i < s->objects; i++)
2519 INIT_LIST_HEAD(slabs_by_inuse + i);
2521 spin_lock_irqsave(&n->list_lock, flags);
2524 * Build lists indexed by the items in use in each slab.
2526 * Note that concurrent frees may occur while we hold the
2527 * list_lock. page->inuse here is the upper limit.
2529 list_for_each_entry_safe(page, t, &n->partial, lru) {
2530 if (!page->inuse && slab_trylock(page)) {
2532 * Must hold slab lock here because slab_free
2533 * may have freed the last object and be
2534 * waiting to release the slab.
2536 list_del(&page->lru);
2537 n->nr_partial--;
2538 slab_unlock(page);
2539 discard_slab(s, page);
2540 } else {
2541 list_move(&page->lru,
2542 slabs_by_inuse + page->inuse);
2547 * Rebuild the partial list with the slabs filled up most
2548 * first and the least used slabs at the end.
2550 for (i = s->objects - 1; i >= 0; i--)
2551 list_splice(slabs_by_inuse + i, n->partial.prev);
2553 spin_unlock_irqrestore(&n->list_lock, flags);
2556 kfree(slabs_by_inuse);
2557 return 0;
2559 EXPORT_SYMBOL(kmem_cache_shrink);
2561 /********************************************************************
2562 * Basic setup of slabs
2563 *******************************************************************/
2565 void __init kmem_cache_init(void)
2567 int i;
2568 int caches = 0;
2570 #ifdef CONFIG_NUMA
2572 * Must first have the slab cache available for the allocations of the
2573 * struct kmem_cache_node's. There is special bootstrap code in
2574 * kmem_cache_open for slab_state == DOWN.
2576 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2577 sizeof(struct kmem_cache_node), GFP_KERNEL);
2578 kmalloc_caches[0].refcount = -1;
2579 caches++;
2580 #endif
2582 /* Able to allocate the per node structures */
2583 slab_state = PARTIAL;
2585 /* Caches that are not of the two-to-the-power-of size */
2586 if (KMALLOC_MIN_SIZE <= 64) {
2587 create_kmalloc_cache(&kmalloc_caches[1],
2588 "kmalloc-96", 96, GFP_KERNEL);
2589 caches++;
2591 if (KMALLOC_MIN_SIZE <= 128) {
2592 create_kmalloc_cache(&kmalloc_caches[2],
2593 "kmalloc-192", 192, GFP_KERNEL);
2594 caches++;
2597 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++) {
2598 create_kmalloc_cache(&kmalloc_caches[i],
2599 "kmalloc", 1 << i, GFP_KERNEL);
2600 caches++;
2605 * Patch up the size_index table if we have strange large alignment
2606 * requirements for the kmalloc array. This is only the case for
2607 * mips it seems. The standard arches will not generate any code here.
2609 * Largest permitted alignment is 256 bytes due to the way we
2610 * handle the index determination for the smaller caches.
2612 * Make sure that nothing crazy happens if someone starts tinkering
2613 * around with ARCH_KMALLOC_MINALIGN
2615 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
2616 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
2618 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
2619 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
2621 slab_state = UP;
2623 /* Provide the correct kmalloc names now that the caches are up */
2624 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++)
2625 kmalloc_caches[i]. name =
2626 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
2628 #ifdef CONFIG_SMP
2629 register_cpu_notifier(&slab_notifier);
2630 #endif
2632 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
2633 nr_cpu_ids * sizeof(struct page *);
2635 printk(KERN_INFO "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
2636 " CPUs=%d, Nodes=%d\n",
2637 caches, cache_line_size(),
2638 slub_min_order, slub_max_order, slub_min_objects,
2639 nr_cpu_ids, nr_node_ids);
2643 * Find a mergeable slab cache
2645 static int slab_unmergeable(struct kmem_cache *s)
2647 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
2648 return 1;
2650 if (s->ctor)
2651 return 1;
2654 * We may have set a slab to be unmergeable during bootstrap.
2656 if (s->refcount < 0)
2657 return 1;
2659 return 0;
2662 static struct kmem_cache *find_mergeable(size_t size,
2663 size_t align, unsigned long flags,
2664 void (*ctor)(void *, struct kmem_cache *, unsigned long))
2666 struct kmem_cache *s;
2668 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
2669 return NULL;
2671 if (ctor)
2672 return NULL;
2674 size = ALIGN(size, sizeof(void *));
2675 align = calculate_alignment(flags, align, size);
2676 size = ALIGN(size, align);
2678 list_for_each_entry(s, &slab_caches, list) {
2679 if (slab_unmergeable(s))
2680 continue;
2682 if (size > s->size)
2683 continue;
2685 if (((flags | slub_debug) & SLUB_MERGE_SAME) !=
2686 (s->flags & SLUB_MERGE_SAME))
2687 continue;
2689 * Check if alignment is compatible.
2690 * Courtesy of Adrian Drzewiecki
2692 if ((s->size & ~(align -1)) != s->size)
2693 continue;
2695 if (s->size - size >= sizeof(void *))
2696 continue;
2698 return s;
2700 return NULL;
2703 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
2704 size_t align, unsigned long flags,
2705 void (*ctor)(void *, struct kmem_cache *, unsigned long))
2707 struct kmem_cache *s;
2709 down_write(&slub_lock);
2710 s = find_mergeable(size, align, flags, ctor);
2711 if (s) {
2712 s->refcount++;
2714 * Adjust the object sizes so that we clear
2715 * the complete object on kzalloc.
2717 s->objsize = max(s->objsize, (int)size);
2718 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
2719 up_write(&slub_lock);
2720 if (sysfs_slab_alias(s, name))
2721 goto err;
2722 return s;
2724 s = kmalloc(kmem_size, GFP_KERNEL);
2725 if (s) {
2726 if (kmem_cache_open(s, GFP_KERNEL, name,
2727 size, align, flags, ctor)) {
2728 list_add(&s->list, &slab_caches);
2729 up_write(&slub_lock);
2730 if (sysfs_slab_add(s))
2731 goto err;
2732 return s;
2734 kfree(s);
2736 up_write(&slub_lock);
2738 err:
2739 if (flags & SLAB_PANIC)
2740 panic("Cannot create slabcache %s\n", name);
2741 else
2742 s = NULL;
2743 return s;
2745 EXPORT_SYMBOL(kmem_cache_create);
2747 #ifdef CONFIG_SMP
2749 * Use the cpu notifier to insure that the cpu slabs are flushed when
2750 * necessary.
2752 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
2753 unsigned long action, void *hcpu)
2755 long cpu = (long)hcpu;
2756 struct kmem_cache *s;
2757 unsigned long flags;
2759 switch (action) {
2760 case CPU_UP_CANCELED:
2761 case CPU_UP_CANCELED_FROZEN:
2762 case CPU_DEAD:
2763 case CPU_DEAD_FROZEN:
2764 down_read(&slub_lock);
2765 list_for_each_entry(s, &slab_caches, list) {
2766 local_irq_save(flags);
2767 __flush_cpu_slab(s, cpu);
2768 local_irq_restore(flags);
2770 up_read(&slub_lock);
2771 break;
2772 default:
2773 break;
2775 return NOTIFY_OK;
2778 static struct notifier_block __cpuinitdata slab_notifier =
2779 { &slab_cpuup_callback, NULL, 0 };
2781 #endif
2783 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller)
2785 struct kmem_cache *s = get_slab(size, gfpflags);
2787 if (ZERO_OR_NULL_PTR(s))
2788 return s;
2790 return slab_alloc(s, gfpflags, -1, caller);
2793 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
2794 int node, void *caller)
2796 struct kmem_cache *s = get_slab(size, gfpflags);
2798 if (ZERO_OR_NULL_PTR(s))
2799 return s;
2801 return slab_alloc(s, gfpflags, node, caller);
2804 #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
2805 static int validate_slab(struct kmem_cache *s, struct page *page,
2806 unsigned long *map)
2808 void *p;
2809 void *addr = page_address(page);
2811 if (!check_slab(s, page) ||
2812 !on_freelist(s, page, NULL))
2813 return 0;
2815 /* Now we know that a valid freelist exists */
2816 bitmap_zero(map, s->objects);
2818 for_each_free_object(p, s, page->freelist) {
2819 set_bit(slab_index(p, s, addr), map);
2820 if (!check_object(s, page, p, 0))
2821 return 0;
2824 for_each_object(p, s, addr)
2825 if (!test_bit(slab_index(p, s, addr), map))
2826 if (!check_object(s, page, p, 1))
2827 return 0;
2828 return 1;
2831 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
2832 unsigned long *map)
2834 if (slab_trylock(page)) {
2835 validate_slab(s, page, map);
2836 slab_unlock(page);
2837 } else
2838 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
2839 s->name, page);
2841 if (s->flags & DEBUG_DEFAULT_FLAGS) {
2842 if (!SlabDebug(page))
2843 printk(KERN_ERR "SLUB %s: SlabDebug not set "
2844 "on slab 0x%p\n", s->name, page);
2845 } else {
2846 if (SlabDebug(page))
2847 printk(KERN_ERR "SLUB %s: SlabDebug set on "
2848 "slab 0x%p\n", s->name, page);
2852 static int validate_slab_node(struct kmem_cache *s,
2853 struct kmem_cache_node *n, unsigned long *map)
2855 unsigned long count = 0;
2856 struct page *page;
2857 unsigned long flags;
2859 spin_lock_irqsave(&n->list_lock, flags);
2861 list_for_each_entry(page, &n->partial, lru) {
2862 validate_slab_slab(s, page, map);
2863 count++;
2865 if (count != n->nr_partial)
2866 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
2867 "counter=%ld\n", s->name, count, n->nr_partial);
2869 if (!(s->flags & SLAB_STORE_USER))
2870 goto out;
2872 list_for_each_entry(page, &n->full, lru) {
2873 validate_slab_slab(s, page, map);
2874 count++;
2876 if (count != atomic_long_read(&n->nr_slabs))
2877 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
2878 "counter=%ld\n", s->name, count,
2879 atomic_long_read(&n->nr_slabs));
2881 out:
2882 spin_unlock_irqrestore(&n->list_lock, flags);
2883 return count;
2886 static long validate_slab_cache(struct kmem_cache *s)
2888 int node;
2889 unsigned long count = 0;
2890 unsigned long *map = kmalloc(BITS_TO_LONGS(s->objects) *
2891 sizeof(unsigned long), GFP_KERNEL);
2893 if (!map)
2894 return -ENOMEM;
2896 flush_all(s);
2897 for_each_online_node(node) {
2898 struct kmem_cache_node *n = get_node(s, node);
2900 count += validate_slab_node(s, n, map);
2902 kfree(map);
2903 return count;
2906 #ifdef SLUB_RESILIENCY_TEST
2907 static void resiliency_test(void)
2909 u8 *p;
2911 printk(KERN_ERR "SLUB resiliency testing\n");
2912 printk(KERN_ERR "-----------------------\n");
2913 printk(KERN_ERR "A. Corruption after allocation\n");
2915 p = kzalloc(16, GFP_KERNEL);
2916 p[16] = 0x12;
2917 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
2918 " 0x12->0x%p\n\n", p + 16);
2920 validate_slab_cache(kmalloc_caches + 4);
2922 /* Hmmm... The next two are dangerous */
2923 p = kzalloc(32, GFP_KERNEL);
2924 p[32 + sizeof(void *)] = 0x34;
2925 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
2926 " 0x34 -> -0x%p\n", p);
2927 printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n");
2929 validate_slab_cache(kmalloc_caches + 5);
2930 p = kzalloc(64, GFP_KERNEL);
2931 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
2932 *p = 0x56;
2933 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
2935 printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n");
2936 validate_slab_cache(kmalloc_caches + 6);
2938 printk(KERN_ERR "\nB. Corruption after free\n");
2939 p = kzalloc(128, GFP_KERNEL);
2940 kfree(p);
2941 *p = 0x78;
2942 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
2943 validate_slab_cache(kmalloc_caches + 7);
2945 p = kzalloc(256, GFP_KERNEL);
2946 kfree(p);
2947 p[50] = 0x9a;
2948 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
2949 validate_slab_cache(kmalloc_caches + 8);
2951 p = kzalloc(512, GFP_KERNEL);
2952 kfree(p);
2953 p[512] = 0xab;
2954 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
2955 validate_slab_cache(kmalloc_caches + 9);
2957 #else
2958 static void resiliency_test(void) {};
2959 #endif
2962 * Generate lists of code addresses where slabcache objects are allocated
2963 * and freed.
2966 struct location {
2967 unsigned long count;
2968 void *addr;
2969 long long sum_time;
2970 long min_time;
2971 long max_time;
2972 long min_pid;
2973 long max_pid;
2974 cpumask_t cpus;
2975 nodemask_t nodes;
2978 struct loc_track {
2979 unsigned long max;
2980 unsigned long count;
2981 struct location *loc;
2984 static void free_loc_track(struct loc_track *t)
2986 if (t->max)
2987 free_pages((unsigned long)t->loc,
2988 get_order(sizeof(struct location) * t->max));
2991 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
2993 struct location *l;
2994 int order;
2996 order = get_order(sizeof(struct location) * max);
2998 l = (void *)__get_free_pages(flags, order);
2999 if (!l)
3000 return 0;
3002 if (t->count) {
3003 memcpy(l, t->loc, sizeof(struct location) * t->count);
3004 free_loc_track(t);
3006 t->max = max;
3007 t->loc = l;
3008 return 1;
3011 static int add_location(struct loc_track *t, struct kmem_cache *s,
3012 const struct track *track)
3014 long start, end, pos;
3015 struct location *l;
3016 void *caddr;
3017 unsigned long age = jiffies - track->when;
3019 start = -1;
3020 end = t->count;
3022 for ( ; ; ) {
3023 pos = start + (end - start + 1) / 2;
3026 * There is nothing at "end". If we end up there
3027 * we need to add something to before end.
3029 if (pos == end)
3030 break;
3032 caddr = t->loc[pos].addr;
3033 if (track->addr == caddr) {
3035 l = &t->loc[pos];
3036 l->count++;
3037 if (track->when) {
3038 l->sum_time += age;
3039 if (age < l->min_time)
3040 l->min_time = age;
3041 if (age > l->max_time)
3042 l->max_time = age;
3044 if (track->pid < l->min_pid)
3045 l->min_pid = track->pid;
3046 if (track->pid > l->max_pid)
3047 l->max_pid = track->pid;
3049 cpu_set(track->cpu, l->cpus);
3051 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3052 return 1;
3055 if (track->addr < caddr)
3056 end = pos;
3057 else
3058 start = pos;
3062 * Not found. Insert new tracking element.
3064 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3065 return 0;
3067 l = t->loc + pos;
3068 if (pos < t->count)
3069 memmove(l + 1, l,
3070 (t->count - pos) * sizeof(struct location));
3071 t->count++;
3072 l->count = 1;
3073 l->addr = track->addr;
3074 l->sum_time = age;
3075 l->min_time = age;
3076 l->max_time = age;
3077 l->min_pid = track->pid;
3078 l->max_pid = track->pid;
3079 cpus_clear(l->cpus);
3080 cpu_set(track->cpu, l->cpus);
3081 nodes_clear(l->nodes);
3082 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3083 return 1;
3086 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3087 struct page *page, enum track_item alloc)
3089 void *addr = page_address(page);
3090 DECLARE_BITMAP(map, s->objects);
3091 void *p;
3093 bitmap_zero(map, s->objects);
3094 for_each_free_object(p, s, page->freelist)
3095 set_bit(slab_index(p, s, addr), map);
3097 for_each_object(p, s, addr)
3098 if (!test_bit(slab_index(p, s, addr), map))
3099 add_location(t, s, get_track(s, p, alloc));
3102 static int list_locations(struct kmem_cache *s, char *buf,
3103 enum track_item alloc)
3105 int n = 0;
3106 unsigned long i;
3107 struct loc_track t = { 0, 0, NULL };
3108 int node;
3110 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3111 GFP_KERNEL))
3112 return sprintf(buf, "Out of memory\n");
3114 /* Push back cpu slabs */
3115 flush_all(s);
3117 for_each_online_node(node) {
3118 struct kmem_cache_node *n = get_node(s, node);
3119 unsigned long flags;
3120 struct page *page;
3122 if (!atomic_long_read(&n->nr_slabs))
3123 continue;
3125 spin_lock_irqsave(&n->list_lock, flags);
3126 list_for_each_entry(page, &n->partial, lru)
3127 process_slab(&t, s, page, alloc);
3128 list_for_each_entry(page, &n->full, lru)
3129 process_slab(&t, s, page, alloc);
3130 spin_unlock_irqrestore(&n->list_lock, flags);
3133 for (i = 0; i < t.count; i++) {
3134 struct location *l = &t.loc[i];
3136 if (n > PAGE_SIZE - 100)
3137 break;
3138 n += sprintf(buf + n, "%7ld ", l->count);
3140 if (l->addr)
3141 n += sprint_symbol(buf + n, (unsigned long)l->addr);
3142 else
3143 n += sprintf(buf + n, "<not-available>");
3145 if (l->sum_time != l->min_time) {
3146 unsigned long remainder;
3148 n += sprintf(buf + n, " age=%ld/%ld/%ld",
3149 l->min_time,
3150 div_long_long_rem(l->sum_time, l->count, &remainder),
3151 l->max_time);
3152 } else
3153 n += sprintf(buf + n, " age=%ld",
3154 l->min_time);
3156 if (l->min_pid != l->max_pid)
3157 n += sprintf(buf + n, " pid=%ld-%ld",
3158 l->min_pid, l->max_pid);
3159 else
3160 n += sprintf(buf + n, " pid=%ld",
3161 l->min_pid);
3163 if (num_online_cpus() > 1 && !cpus_empty(l->cpus) &&
3164 n < PAGE_SIZE - 60) {
3165 n += sprintf(buf + n, " cpus=");
3166 n += cpulist_scnprintf(buf + n, PAGE_SIZE - n - 50,
3167 l->cpus);
3170 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
3171 n < PAGE_SIZE - 60) {
3172 n += sprintf(buf + n, " nodes=");
3173 n += nodelist_scnprintf(buf + n, PAGE_SIZE - n - 50,
3174 l->nodes);
3177 n += sprintf(buf + n, "\n");
3180 free_loc_track(&t);
3181 if (!t.count)
3182 n += sprintf(buf, "No data\n");
3183 return n;
3186 static unsigned long count_partial(struct kmem_cache_node *n)
3188 unsigned long flags;
3189 unsigned long x = 0;
3190 struct page *page;
3192 spin_lock_irqsave(&n->list_lock, flags);
3193 list_for_each_entry(page, &n->partial, lru)
3194 x += page->inuse;
3195 spin_unlock_irqrestore(&n->list_lock, flags);
3196 return x;
3199 enum slab_stat_type {
3200 SL_FULL,
3201 SL_PARTIAL,
3202 SL_CPU,
3203 SL_OBJECTS
3206 #define SO_FULL (1 << SL_FULL)
3207 #define SO_PARTIAL (1 << SL_PARTIAL)
3208 #define SO_CPU (1 << SL_CPU)
3209 #define SO_OBJECTS (1 << SL_OBJECTS)
3211 static unsigned long slab_objects(struct kmem_cache *s,
3212 char *buf, unsigned long flags)
3214 unsigned long total = 0;
3215 int cpu;
3216 int node;
3217 int x;
3218 unsigned long *nodes;
3219 unsigned long *per_cpu;
3221 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3222 per_cpu = nodes + nr_node_ids;
3224 for_each_possible_cpu(cpu) {
3225 struct page *page = s->cpu_slab[cpu];
3226 int node;
3228 if (page) {
3229 node = page_to_nid(page);
3230 if (flags & SO_CPU) {
3231 int x = 0;
3233 if (flags & SO_OBJECTS)
3234 x = page->inuse;
3235 else
3236 x = 1;
3237 total += x;
3238 nodes[node] += x;
3240 per_cpu[node]++;
3244 for_each_online_node(node) {
3245 struct kmem_cache_node *n = get_node(s, node);
3247 if (flags & SO_PARTIAL) {
3248 if (flags & SO_OBJECTS)
3249 x = count_partial(n);
3250 else
3251 x = n->nr_partial;
3252 total += x;
3253 nodes[node] += x;
3256 if (flags & SO_FULL) {
3257 int full_slabs = atomic_long_read(&n->nr_slabs)
3258 - per_cpu[node]
3259 - n->nr_partial;
3261 if (flags & SO_OBJECTS)
3262 x = full_slabs * s->objects;
3263 else
3264 x = full_slabs;
3265 total += x;
3266 nodes[node] += x;
3270 x = sprintf(buf, "%lu", total);
3271 #ifdef CONFIG_NUMA
3272 for_each_online_node(node)
3273 if (nodes[node])
3274 x += sprintf(buf + x, " N%d=%lu",
3275 node, nodes[node]);
3276 #endif
3277 kfree(nodes);
3278 return x + sprintf(buf + x, "\n");
3281 static int any_slab_objects(struct kmem_cache *s)
3283 int node;
3284 int cpu;
3286 for_each_possible_cpu(cpu)
3287 if (s->cpu_slab[cpu])
3288 return 1;
3290 for_each_node(node) {
3291 struct kmem_cache_node *n = get_node(s, node);
3293 if (n->nr_partial || atomic_long_read(&n->nr_slabs))
3294 return 1;
3296 return 0;
3299 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3300 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3302 struct slab_attribute {
3303 struct attribute attr;
3304 ssize_t (*show)(struct kmem_cache *s, char *buf);
3305 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3308 #define SLAB_ATTR_RO(_name) \
3309 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3311 #define SLAB_ATTR(_name) \
3312 static struct slab_attribute _name##_attr = \
3313 __ATTR(_name, 0644, _name##_show, _name##_store)
3315 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3317 return sprintf(buf, "%d\n", s->size);
3319 SLAB_ATTR_RO(slab_size);
3321 static ssize_t align_show(struct kmem_cache *s, char *buf)
3323 return sprintf(buf, "%d\n", s->align);
3325 SLAB_ATTR_RO(align);
3327 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3329 return sprintf(buf, "%d\n", s->objsize);
3331 SLAB_ATTR_RO(object_size);
3333 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3335 return sprintf(buf, "%d\n", s->objects);
3337 SLAB_ATTR_RO(objs_per_slab);
3339 static ssize_t order_show(struct kmem_cache *s, char *buf)
3341 return sprintf(buf, "%d\n", s->order);
3343 SLAB_ATTR_RO(order);
3345 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3347 if (s->ctor) {
3348 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3350 return n + sprintf(buf + n, "\n");
3352 return 0;
3354 SLAB_ATTR_RO(ctor);
3356 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3358 return sprintf(buf, "%d\n", s->refcount - 1);
3360 SLAB_ATTR_RO(aliases);
3362 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3364 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU);
3366 SLAB_ATTR_RO(slabs);
3368 static ssize_t partial_show(struct kmem_cache *s, char *buf)
3370 return slab_objects(s, buf, SO_PARTIAL);
3372 SLAB_ATTR_RO(partial);
3374 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3376 return slab_objects(s, buf, SO_CPU);
3378 SLAB_ATTR_RO(cpu_slabs);
3380 static ssize_t objects_show(struct kmem_cache *s, char *buf)
3382 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU|SO_OBJECTS);
3384 SLAB_ATTR_RO(objects);
3386 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3388 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3391 static ssize_t sanity_checks_store(struct kmem_cache *s,
3392 const char *buf, size_t length)
3394 s->flags &= ~SLAB_DEBUG_FREE;
3395 if (buf[0] == '1')
3396 s->flags |= SLAB_DEBUG_FREE;
3397 return length;
3399 SLAB_ATTR(sanity_checks);
3401 static ssize_t trace_show(struct kmem_cache *s, char *buf)
3403 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3406 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3407 size_t length)
3409 s->flags &= ~SLAB_TRACE;
3410 if (buf[0] == '1')
3411 s->flags |= SLAB_TRACE;
3412 return length;
3414 SLAB_ATTR(trace);
3416 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3418 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3421 static ssize_t reclaim_account_store(struct kmem_cache *s,
3422 const char *buf, size_t length)
3424 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3425 if (buf[0] == '1')
3426 s->flags |= SLAB_RECLAIM_ACCOUNT;
3427 return length;
3429 SLAB_ATTR(reclaim_account);
3431 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3433 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
3435 SLAB_ATTR_RO(hwcache_align);
3437 #ifdef CONFIG_ZONE_DMA
3438 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3440 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3442 SLAB_ATTR_RO(cache_dma);
3443 #endif
3445 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3447 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3449 SLAB_ATTR_RO(destroy_by_rcu);
3451 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3453 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3456 static ssize_t red_zone_store(struct kmem_cache *s,
3457 const char *buf, size_t length)
3459 if (any_slab_objects(s))
3460 return -EBUSY;
3462 s->flags &= ~SLAB_RED_ZONE;
3463 if (buf[0] == '1')
3464 s->flags |= SLAB_RED_ZONE;
3465 calculate_sizes(s);
3466 return length;
3468 SLAB_ATTR(red_zone);
3470 static ssize_t poison_show(struct kmem_cache *s, char *buf)
3472 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
3475 static ssize_t poison_store(struct kmem_cache *s,
3476 const char *buf, size_t length)
3478 if (any_slab_objects(s))
3479 return -EBUSY;
3481 s->flags &= ~SLAB_POISON;
3482 if (buf[0] == '1')
3483 s->flags |= SLAB_POISON;
3484 calculate_sizes(s);
3485 return length;
3487 SLAB_ATTR(poison);
3489 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
3491 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
3494 static ssize_t store_user_store(struct kmem_cache *s,
3495 const char *buf, size_t length)
3497 if (any_slab_objects(s))
3498 return -EBUSY;
3500 s->flags &= ~SLAB_STORE_USER;
3501 if (buf[0] == '1')
3502 s->flags |= SLAB_STORE_USER;
3503 calculate_sizes(s);
3504 return length;
3506 SLAB_ATTR(store_user);
3508 static ssize_t validate_show(struct kmem_cache *s, char *buf)
3510 return 0;
3513 static ssize_t validate_store(struct kmem_cache *s,
3514 const char *buf, size_t length)
3516 int ret = -EINVAL;
3518 if (buf[0] == '1') {
3519 ret = validate_slab_cache(s);
3520 if (ret >= 0)
3521 ret = length;
3523 return ret;
3525 SLAB_ATTR(validate);
3527 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
3529 return 0;
3532 static ssize_t shrink_store(struct kmem_cache *s,
3533 const char *buf, size_t length)
3535 if (buf[0] == '1') {
3536 int rc = kmem_cache_shrink(s);
3538 if (rc)
3539 return rc;
3540 } else
3541 return -EINVAL;
3542 return length;
3544 SLAB_ATTR(shrink);
3546 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
3548 if (!(s->flags & SLAB_STORE_USER))
3549 return -ENOSYS;
3550 return list_locations(s, buf, TRACK_ALLOC);
3552 SLAB_ATTR_RO(alloc_calls);
3554 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
3556 if (!(s->flags & SLAB_STORE_USER))
3557 return -ENOSYS;
3558 return list_locations(s, buf, TRACK_FREE);
3560 SLAB_ATTR_RO(free_calls);
3562 #ifdef CONFIG_NUMA
3563 static ssize_t defrag_ratio_show(struct kmem_cache *s, char *buf)
3565 return sprintf(buf, "%d\n", s->defrag_ratio / 10);
3568 static ssize_t defrag_ratio_store(struct kmem_cache *s,
3569 const char *buf, size_t length)
3571 int n = simple_strtoul(buf, NULL, 10);
3573 if (n < 100)
3574 s->defrag_ratio = n * 10;
3575 return length;
3577 SLAB_ATTR(defrag_ratio);
3578 #endif
3580 static struct attribute * slab_attrs[] = {
3581 &slab_size_attr.attr,
3582 &object_size_attr.attr,
3583 &objs_per_slab_attr.attr,
3584 &order_attr.attr,
3585 &objects_attr.attr,
3586 &slabs_attr.attr,
3587 &partial_attr.attr,
3588 &cpu_slabs_attr.attr,
3589 &ctor_attr.attr,
3590 &aliases_attr.attr,
3591 &align_attr.attr,
3592 &sanity_checks_attr.attr,
3593 &trace_attr.attr,
3594 &hwcache_align_attr.attr,
3595 &reclaim_account_attr.attr,
3596 &destroy_by_rcu_attr.attr,
3597 &red_zone_attr.attr,
3598 &poison_attr.attr,
3599 &store_user_attr.attr,
3600 &validate_attr.attr,
3601 &shrink_attr.attr,
3602 &alloc_calls_attr.attr,
3603 &free_calls_attr.attr,
3604 #ifdef CONFIG_ZONE_DMA
3605 &cache_dma_attr.attr,
3606 #endif
3607 #ifdef CONFIG_NUMA
3608 &defrag_ratio_attr.attr,
3609 #endif
3610 NULL
3613 static struct attribute_group slab_attr_group = {
3614 .attrs = slab_attrs,
3617 static ssize_t slab_attr_show(struct kobject *kobj,
3618 struct attribute *attr,
3619 char *buf)
3621 struct slab_attribute *attribute;
3622 struct kmem_cache *s;
3623 int err;
3625 attribute = to_slab_attr(attr);
3626 s = to_slab(kobj);
3628 if (!attribute->show)
3629 return -EIO;
3631 err = attribute->show(s, buf);
3633 return err;
3636 static ssize_t slab_attr_store(struct kobject *kobj,
3637 struct attribute *attr,
3638 const char *buf, size_t len)
3640 struct slab_attribute *attribute;
3641 struct kmem_cache *s;
3642 int err;
3644 attribute = to_slab_attr(attr);
3645 s = to_slab(kobj);
3647 if (!attribute->store)
3648 return -EIO;
3650 err = attribute->store(s, buf, len);
3652 return err;
3655 static struct sysfs_ops slab_sysfs_ops = {
3656 .show = slab_attr_show,
3657 .store = slab_attr_store,
3660 static struct kobj_type slab_ktype = {
3661 .sysfs_ops = &slab_sysfs_ops,
3664 static int uevent_filter(struct kset *kset, struct kobject *kobj)
3666 struct kobj_type *ktype = get_ktype(kobj);
3668 if (ktype == &slab_ktype)
3669 return 1;
3670 return 0;
3673 static struct kset_uevent_ops slab_uevent_ops = {
3674 .filter = uevent_filter,
3677 static decl_subsys(slab, &slab_ktype, &slab_uevent_ops);
3679 #define ID_STR_LENGTH 64
3681 /* Create a unique string id for a slab cache:
3682 * format
3683 * :[flags-]size:[memory address of kmemcache]
3685 static char *create_unique_id(struct kmem_cache *s)
3687 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
3688 char *p = name;
3690 BUG_ON(!name);
3692 *p++ = ':';
3694 * First flags affecting slabcache operations. We will only
3695 * get here for aliasable slabs so we do not need to support
3696 * too many flags. The flags here must cover all flags that
3697 * are matched during merging to guarantee that the id is
3698 * unique.
3700 if (s->flags & SLAB_CACHE_DMA)
3701 *p++ = 'd';
3702 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3703 *p++ = 'a';
3704 if (s->flags & SLAB_DEBUG_FREE)
3705 *p++ = 'F';
3706 if (p != name + 1)
3707 *p++ = '-';
3708 p += sprintf(p, "%07d", s->size);
3709 BUG_ON(p > name + ID_STR_LENGTH - 1);
3710 return name;
3713 static int sysfs_slab_add(struct kmem_cache *s)
3715 int err;
3716 const char *name;
3717 int unmergeable;
3719 if (slab_state < SYSFS)
3720 /* Defer until later */
3721 return 0;
3723 unmergeable = slab_unmergeable(s);
3724 if (unmergeable) {
3726 * Slabcache can never be merged so we can use the name proper.
3727 * This is typically the case for debug situations. In that
3728 * case we can catch duplicate names easily.
3730 sysfs_remove_link(&slab_subsys.kobj, s->name);
3731 name = s->name;
3732 } else {
3734 * Create a unique name for the slab as a target
3735 * for the symlinks.
3737 name = create_unique_id(s);
3740 kobj_set_kset_s(s, slab_subsys);
3741 kobject_set_name(&s->kobj, name);
3742 kobject_init(&s->kobj);
3743 err = kobject_add(&s->kobj);
3744 if (err)
3745 return err;
3747 err = sysfs_create_group(&s->kobj, &slab_attr_group);
3748 if (err)
3749 return err;
3750 kobject_uevent(&s->kobj, KOBJ_ADD);
3751 if (!unmergeable) {
3752 /* Setup first alias */
3753 sysfs_slab_alias(s, s->name);
3754 kfree(name);
3756 return 0;
3759 static void sysfs_slab_remove(struct kmem_cache *s)
3761 kobject_uevent(&s->kobj, KOBJ_REMOVE);
3762 kobject_del(&s->kobj);
3766 * Need to buffer aliases during bootup until sysfs becomes
3767 * available lest we loose that information.
3769 struct saved_alias {
3770 struct kmem_cache *s;
3771 const char *name;
3772 struct saved_alias *next;
3775 static struct saved_alias *alias_list;
3777 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
3779 struct saved_alias *al;
3781 if (slab_state == SYSFS) {
3783 * If we have a leftover link then remove it.
3785 sysfs_remove_link(&slab_subsys.kobj, name);
3786 return sysfs_create_link(&slab_subsys.kobj,
3787 &s->kobj, name);
3790 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
3791 if (!al)
3792 return -ENOMEM;
3794 al->s = s;
3795 al->name = name;
3796 al->next = alias_list;
3797 alias_list = al;
3798 return 0;
3801 static int __init slab_sysfs_init(void)
3803 struct kmem_cache *s;
3804 int err;
3806 err = subsystem_register(&slab_subsys);
3807 if (err) {
3808 printk(KERN_ERR "Cannot register slab subsystem.\n");
3809 return -ENOSYS;
3812 slab_state = SYSFS;
3814 list_for_each_entry(s, &slab_caches, list) {
3815 err = sysfs_slab_add(s);
3816 BUG_ON(err);
3819 while (alias_list) {
3820 struct saved_alias *al = alias_list;
3822 alias_list = alias_list->next;
3823 err = sysfs_slab_alias(al->s, al->name);
3824 BUG_ON(err);
3825 kfree(al);
3828 resiliency_test();
3829 return 0;
3832 __initcall(slab_sysfs_init);
3833 #endif