Root/mm/slub.c

1/*
2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
4 *
5 * The allocator synchronizes using per slab locks and only
6 * uses a centralized lock to manage a pool of partial slabs.
7 *
8 * (C) 2007 SGI, Christoph Lameter
9 */
10
11#include <linux/mm.h>
12#include <linux/swap.h> /* struct reclaim_state */
13#include <linux/module.h>
14#include <linux/bit_spinlock.h>
15#include <linux/interrupt.h>
16#include <linux/bitops.h>
17#include <linux/slab.h>
18#include <linux/proc_fs.h>
19#include <linux/seq_file.h>
20#include <linux/kmemcheck.h>
21#include <linux/cpu.h>
22#include <linux/cpuset.h>
23#include <linux/mempolicy.h>
24#include <linux/ctype.h>
25#include <linux/debugobjects.h>
26#include <linux/kallsyms.h>
27#include <linux/memory.h>
28#include <linux/math64.h>
29#include <linux/fault-inject.h>
30
31#include <trace/events/kmem.h>
32
33/*
34 * Lock order:
35 * 1. slab_lock(page)
36 * 2. slab->list_lock
37 *
38 * The slab_lock protects operations on the object of a particular
39 * slab and its metadata in the page struct. If the slab lock
40 * has been taken then no allocations nor frees can be performed
41 * on the objects in the slab nor can the slab be added or removed
42 * from the partial or full lists since this would mean modifying
43 * the page_struct of the slab.
44 *
45 * The list_lock protects the partial and full list on each node and
46 * the partial slab counter. If taken then no new slabs may be added or
47 * removed from the lists nor make the number of partial slabs be modified.
48 * (Note that the total number of slabs is an atomic value that may be
49 * modified without taking the list lock).
50 *
51 * The list_lock is a centralized lock and thus we avoid taking it as
52 * much as possible. As long as SLUB does not have to handle partial
53 * slabs, operations can continue without any centralized lock. F.e.
54 * allocating a long series of objects that fill up slabs does not require
55 * the list lock.
56 *
57 * The lock order is sometimes inverted when we are trying to get a slab
58 * off a list. We take the list_lock and then look for a page on the list
59 * to use. While we do that objects in the slabs may be freed. We can
60 * only operate on the slab if we have also taken the slab_lock. So we use
61 * a slab_trylock() on the slab. If trylock was successful then no frees
62 * can occur anymore and we can use the slab for allocations etc. If the
63 * slab_trylock() does not succeed then frees are in progress in the slab and
64 * we must stay away from it for a while since we may cause a bouncing
65 * cacheline if we try to acquire the lock. So go onto the next slab.
66 * If all pages are busy then we may allocate a new slab instead of reusing
67 * a partial slab. A new slab has no one operating on it and thus there is
68 * no danger of cacheline contention.
69 *
70 * Interrupts are disabled during allocation and deallocation in order to
71 * make the slab allocator safe to use in the context of an irq. In addition
72 * interrupts are disabled to ensure that the processor does not change
73 * while handling per_cpu slabs, due to kernel preemption.
74 *
75 * SLUB assigns one slab for allocation to each processor.
76 * Allocations only occur from these slabs called cpu slabs.
77 *
78 * Slabs with free elements are kept on a partial list and during regular
79 * operations no list for full slabs is used. If an object in a full slab is
80 * freed then the slab will show up again on the partial lists.
81 * We track full slabs for debugging purposes though because otherwise we
82 * cannot scan all objects.
83 *
84 * Slabs are freed when they become empty. Teardown and setup is
85 * minimal so we rely on the page allocators per cpu caches for
86 * fast frees and allocs.
87 *
88 * Overloading of page flags that are otherwise used for LRU management.
89 *
90 * PageActive The slab is frozen and exempt from list processing.
91 * This means that the slab is dedicated to a purpose
92 * such as satisfying allocations for a specific
93 * processor. Objects may be freed in the slab while
94 * it is frozen but slab_free will then skip the usual
95 * list operations. It is up to the processor holding
96 * the slab to integrate the slab into the slab lists
97 * when the slab is no longer needed.
98 *
99 * One use of this flag is to mark slabs that are
100 * used for allocations. Then such a slab becomes a cpu
101 * slab. The cpu slab may be equipped with an additional
102 * freelist that allows lockless access to
103 * free objects in addition to the regular freelist
104 * that requires the slab lock.
105 *
106 * PageError Slab requires special handling due to debug
107 * options set. This moves slab handling out of
108 * the fast path and disables lockless freelists.
109 */
110
111#define SLAB_DEBUG_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
112        SLAB_TRACE | SLAB_DEBUG_FREE)
113
114static inline int kmem_cache_debug(struct kmem_cache *s)
115{
116#ifdef CONFIG_SLUB_DEBUG
117    return unlikely(s->flags & SLAB_DEBUG_FLAGS);
118#else
119    return 0;
120#endif
121}
122
123/*
124 * Issues still to be resolved:
125 *
126 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
127 *
128 * - Variable sizing of the per node arrays
129 */
130
131/* Enable to test recovery from slab corruption on boot */
132#undef SLUB_RESILIENCY_TEST
133
134/*
135 * Mininum number of partial slabs. These will be left on the partial
136 * lists even if they are empty. kmem_cache_shrink may reclaim them.
137 */
138#define MIN_PARTIAL 5
139
140/*
141 * Maximum number of desirable partial slabs.
142 * The existence of more partial slabs makes kmem_cache_shrink
143 * sort the partial list by the number of objects in the.
144 */
145#define MAX_PARTIAL 10
146
147#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
148                SLAB_POISON | SLAB_STORE_USER)
149
150/*
151 * Debugging flags that require metadata to be stored in the slab. These get
152 * disabled when slub_debug=O is used and a cache's min order increases with
153 * metadata.
154 */
155#define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
156
157/*
158 * Set of flags that will prevent slab merging
159 */
160#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
161        SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
162        SLAB_FAILSLAB)
163
164#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
165        SLAB_CACHE_DMA | SLAB_NOTRACK)
166
167#define OO_SHIFT 16
168#define OO_MASK ((1 << OO_SHIFT) - 1)
169#define MAX_OBJS_PER_PAGE 65535 /* since page.objects is u16 */
170
171/* Internal SLUB flags */
172#define __OBJECT_POISON 0x80000000UL /* Poison object */
173
174static int kmem_size = sizeof(struct kmem_cache);
175
176#ifdef CONFIG_SMP
177static struct notifier_block slab_notifier;
178#endif
179
180static enum {
181    DOWN, /* No slab functionality available */
182    PARTIAL, /* Kmem_cache_node works */
183    UP, /* Everything works but does not show up in sysfs */
184    SYSFS /* Sysfs up */
185} slab_state = DOWN;
186
187/* A list of all slab caches on the system */
188static DECLARE_RWSEM(slub_lock);
189static LIST_HEAD(slab_caches);
190
191/*
192 * Tracking user of a slab.
193 */
194struct track {
195    unsigned long addr; /* Called from address */
196    int cpu; /* Was running on cpu */
197    int pid; /* Pid context */
198    unsigned long when; /* When did the operation occur */
199};
200
201enum track_item { TRACK_ALLOC, TRACK_FREE };
202
203#ifdef CONFIG_SYSFS
204static int sysfs_slab_add(struct kmem_cache *);
205static int sysfs_slab_alias(struct kmem_cache *, const char *);
206static void sysfs_slab_remove(struct kmem_cache *);
207
208#else
209static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
210static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
211                            { return 0; }
212static inline void sysfs_slab_remove(struct kmem_cache *s)
213{
214    kfree(s->name);
215    kfree(s);
216}
217
218#endif
219
220static inline void stat(const struct kmem_cache *s, enum stat_item si)
221{
222#ifdef CONFIG_SLUB_STATS
223    __this_cpu_inc(s->cpu_slab->stat[si]);
224#endif
225}
226
227/********************************************************************
228 * Core slab cache functions
229 *******************************************************************/
230
231int slab_is_available(void)
232{
233    return slab_state >= UP;
234}
235
236static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
237{
238    return s->node[node];
239}
240
241/* Verify that a pointer has an address that is valid within a slab page */
242static inline int check_valid_pointer(struct kmem_cache *s,
243                struct page *page, const void *object)
244{
245    void *base;
246
247    if (!object)
248        return 1;
249
250    base = page_address(page);
251    if (object < base || object >= base + page->objects * s->size ||
252        (object - base) % s->size) {
253        return 0;
254    }
255
256    return 1;
257}
258
259static inline void *get_freepointer(struct kmem_cache *s, void *object)
260{
261    return *(void **)(object + s->offset);
262}
263
264static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
265{
266    void *p;
267
268#ifdef CONFIG_DEBUG_PAGEALLOC
269    probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
270#else
271    p = get_freepointer(s, object);
272#endif
273    return p;
274}
275
276static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
277{
278    *(void **)(object + s->offset) = fp;
279}
280
281/* Loop over all objects in a slab */
282#define for_each_object(__p, __s, __addr, __objects) \
283    for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
284            __p += (__s)->size)
285
286/* Determine object index from a given position */
287static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
288{
289    return (p - addr) / s->size;
290}
291
292static inline size_t slab_ksize(const struct kmem_cache *s)
293{
294#ifdef CONFIG_SLUB_DEBUG
295    /*
296     * Debugging requires use of the padding between object
297     * and whatever may come after it.
298     */
299    if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
300        return s->objsize;
301
302#endif
303    /*
304     * If we have the need to store the freelist pointer
305     * back there or track user information then we can
306     * only use the space before that information.
307     */
308    if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
309        return s->inuse;
310    /*
311     * Else we can use all the padding etc for the allocation
312     */
313    return s->size;
314}
315
316static inline int order_objects(int order, unsigned long size, int reserved)
317{
318    return ((PAGE_SIZE << order) - reserved) / size;
319}
320
321static inline struct kmem_cache_order_objects oo_make(int order,
322        unsigned long size, int reserved)
323{
324    struct kmem_cache_order_objects x = {
325        (order << OO_SHIFT) + order_objects(order, size, reserved)
326    };
327
328    return x;
329}
330
331static inline int oo_order(struct kmem_cache_order_objects x)
332{
333    return x.x >> OO_SHIFT;
334}
335
336static inline int oo_objects(struct kmem_cache_order_objects x)
337{
338    return x.x & OO_MASK;
339}
340
341#ifdef CONFIG_SLUB_DEBUG
342/*
343 * Determine a map of object in use on a page.
344 *
345 * Slab lock or node listlock must be held to guarantee that the page does
346 * not vanish from under us.
347 */
348static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
349{
350    void *p;
351    void *addr = page_address(page);
352
353    for (p = page->freelist; p; p = get_freepointer(s, p))
354        set_bit(slab_index(p, s, addr), map);
355}
356
357/*
358 * Debug settings:
359 */
360#ifdef CONFIG_SLUB_DEBUG_ON
361static int slub_debug = DEBUG_DEFAULT_FLAGS;
362#else
363static int slub_debug;
364#endif
365
366static char *slub_debug_slabs;
367static int disable_higher_order_debug;
368
369/*
370 * Object debugging
371 */
372static void print_section(char *text, u8 *addr, unsigned int length)
373{
374    int i, offset;
375    int newline = 1;
376    char ascii[17];
377
378    ascii[16] = 0;
379
380    for (i = 0; i < length; i++) {
381        if (newline) {
382            printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
383            newline = 0;
384        }
385        printk(KERN_CONT " %02x", addr[i]);
386        offset = i % 16;
387        ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
388        if (offset == 15) {
389            printk(KERN_CONT " %s\n", ascii);
390            newline = 1;
391        }
392    }
393    if (!newline) {
394        i %= 16;
395        while (i < 16) {
396            printk(KERN_CONT " ");
397            ascii[i] = ' ';
398            i++;
399        }
400        printk(KERN_CONT " %s\n", ascii);
401    }
402}
403
404static struct track *get_track(struct kmem_cache *s, void *object,
405    enum track_item alloc)
406{
407    struct track *p;
408
409    if (s->offset)
410        p = object + s->offset + sizeof(void *);
411    else
412        p = object + s->inuse;
413
414    return p + alloc;
415}
416
417static void set_track(struct kmem_cache *s, void *object,
418            enum track_item alloc, unsigned long addr)
419{
420    struct track *p = get_track(s, object, alloc);
421
422    if (addr) {
423        p->addr = addr;
424        p->cpu = smp_processor_id();
425        p->pid = current->pid;
426        p->when = jiffies;
427    } else
428        memset(p, 0, sizeof(struct track));
429}
430
431static void init_tracking(struct kmem_cache *s, void *object)
432{
433    if (!(s->flags & SLAB_STORE_USER))
434        return;
435
436    set_track(s, object, TRACK_FREE, 0UL);
437    set_track(s, object, TRACK_ALLOC, 0UL);
438}
439
440static void print_track(const char *s, struct track *t)
441{
442    if (!t->addr)
443        return;
444
445    printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
446        s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
447}
448
449static void print_tracking(struct kmem_cache *s, void *object)
450{
451    if (!(s->flags & SLAB_STORE_USER))
452        return;
453
454    print_track("Allocated", get_track(s, object, TRACK_ALLOC));
455    print_track("Freed", get_track(s, object, TRACK_FREE));
456}
457
458static void print_page_info(struct page *page)
459{
460    printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
461        page, page->objects, page->inuse, page->freelist, page->flags);
462
463}
464
465static void slab_bug(struct kmem_cache *s, char *fmt, ...)
466{
467    va_list args;
468    char buf[100];
469
470    va_start(args, fmt);
471    vsnprintf(buf, sizeof(buf), fmt, args);
472    va_end(args);
473    printk(KERN_ERR "========================================"
474            "=====================================\n");
475    printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
476    printk(KERN_ERR "----------------------------------------"
477            "-------------------------------------\n\n");
478}
479
480static void slab_fix(struct kmem_cache *s, char *fmt, ...)
481{
482    va_list args;
483    char buf[100];
484
485    va_start(args, fmt);
486    vsnprintf(buf, sizeof(buf), fmt, args);
487    va_end(args);
488    printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
489}
490
491static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
492{
493    unsigned int off; /* Offset of last byte */
494    u8 *addr = page_address(page);
495
496    print_tracking(s, p);
497
498    print_page_info(page);
499
500    printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
501            p, p - addr, get_freepointer(s, p));
502
503    if (p > addr + 16)
504        print_section("Bytes b4", p - 16, 16);
505
506    print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE));
507
508    if (s->flags & SLAB_RED_ZONE)
509        print_section("Redzone", p + s->objsize,
510            s->inuse - s->objsize);
511
512    if (s->offset)
513        off = s->offset + sizeof(void *);
514    else
515        off = s->inuse;
516
517    if (s->flags & SLAB_STORE_USER)
518        off += 2 * sizeof(struct track);
519
520    if (off != s->size)
521        /* Beginning of the filler is the free pointer */
522        print_section("Padding", p + off, s->size - off);
523
524    dump_stack();
525}
526
527static void object_err(struct kmem_cache *s, struct page *page,
528            u8 *object, char *reason)
529{
530    slab_bug(s, "%s", reason);
531    print_trailer(s, page, object);
532}
533
534static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
535{
536    va_list args;
537    char buf[100];
538
539    va_start(args, fmt);
540    vsnprintf(buf, sizeof(buf), fmt, args);
541    va_end(args);
542    slab_bug(s, "%s", buf);
543    print_page_info(page);
544    dump_stack();
545}
546
547static void init_object(struct kmem_cache *s, void *object, u8 val)
548{
549    u8 *p = object;
550
551    if (s->flags & __OBJECT_POISON) {
552        memset(p, POISON_FREE, s->objsize - 1);
553        p[s->objsize - 1] = POISON_END;
554    }
555
556    if (s->flags & SLAB_RED_ZONE)
557        memset(p + s->objsize, val, s->inuse - s->objsize);
558}
559
560static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
561{
562    while (bytes) {
563        if (*start != (u8)value)
564            return start;
565        start++;
566        bytes--;
567    }
568    return NULL;
569}
570
571static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
572                        void *from, void *to)
573{
574    slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
575    memset(from, data, to - from);
576}
577
578static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
579            u8 *object, char *what,
580            u8 *start, unsigned int value, unsigned int bytes)
581{
582    u8 *fault;
583    u8 *end;
584
585    fault = check_bytes(start, value, bytes);
586    if (!fault)
587        return 1;
588
589    end = start + bytes;
590    while (end > fault && end[-1] == value)
591        end--;
592
593    slab_bug(s, "%s overwritten", what);
594    printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
595                    fault, end - 1, fault[0], value);
596    print_trailer(s, page, object);
597
598    restore_bytes(s, what, value, fault, end);
599    return 0;
600}
601
602/*
603 * Object layout:
604 *
605 * object address
606 * Bytes of the object to be managed.
607 * If the freepointer may overlay the object then the free
608 * pointer is the first word of the object.
609 *
610 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
611 * 0xa5 (POISON_END)
612 *
613 * object + s->objsize
614 * Padding to reach word boundary. This is also used for Redzoning.
615 * Padding is extended by another word if Redzoning is enabled and
616 * objsize == inuse.
617 *
618 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
619 * 0xcc (RED_ACTIVE) for objects in use.
620 *
621 * object + s->inuse
622 * Meta data starts here.
623 *
624 * A. Free pointer (if we cannot overwrite object on free)
625 * B. Tracking data for SLAB_STORE_USER
626 * C. Padding to reach required alignment boundary or at mininum
627 * one word if debugging is on to be able to detect writes
628 * before the word boundary.
629 *
630 * Padding is done using 0x5a (POISON_INUSE)
631 *
632 * object + s->size
633 * Nothing is used beyond s->size.
634 *
635 * If slabcaches are merged then the objsize and inuse boundaries are mostly
636 * ignored. And therefore no slab options that rely on these boundaries
637 * may be used with merged slabcaches.
638 */
639
640static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
641{
642    unsigned long off = s->inuse; /* The end of info */
643
644    if (s->offset)
645        /* Freepointer is placed after the object. */
646        off += sizeof(void *);
647
648    if (s->flags & SLAB_STORE_USER)
649        /* We also have user information there */
650        off += 2 * sizeof(struct track);
651
652    if (s->size == off)
653        return 1;
654
655    return check_bytes_and_report(s, page, p, "Object padding",
656                p + off, POISON_INUSE, s->size - off);
657}
658
659/* Check the pad bytes at the end of a slab page */
660static int slab_pad_check(struct kmem_cache *s, struct page *page)
661{
662    u8 *start;
663    u8 *fault;
664    u8 *end;
665    int length;
666    int remainder;
667
668    if (!(s->flags & SLAB_POISON))
669        return 1;
670
671    start = page_address(page);
672    length = (PAGE_SIZE << compound_order(page)) - s->reserved;
673    end = start + length;
674    remainder = length % s->size;
675    if (!remainder)
676        return 1;
677
678    fault = check_bytes(end - remainder, POISON_INUSE, remainder);
679    if (!fault)
680        return 1;
681    while (end > fault && end[-1] == POISON_INUSE)
682        end--;
683
684    slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
685    print_section("Padding", end - remainder, remainder);
686
687    restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
688    return 0;
689}
690
691static int check_object(struct kmem_cache *s, struct page *page,
692                    void *object, u8 val)
693{
694    u8 *p = object;
695    u8 *endobject = object + s->objsize;
696
697    if (s->flags & SLAB_RED_ZONE) {
698        if (!check_bytes_and_report(s, page, object, "Redzone",
699            endobject, val, s->inuse - s->objsize))
700            return 0;
701    } else {
702        if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
703            check_bytes_and_report(s, page, p, "Alignment padding",
704                endobject, POISON_INUSE, s->inuse - s->objsize);
705        }
706    }
707
708    if (s->flags & SLAB_POISON) {
709        if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
710            (!check_bytes_and_report(s, page, p, "Poison", p,
711                    POISON_FREE, s->objsize - 1) ||
712             !check_bytes_and_report(s, page, p, "Poison",
713                p + s->objsize - 1, POISON_END, 1)))
714            return 0;
715        /*
716         * check_pad_bytes cleans up on its own.
717         */
718        check_pad_bytes(s, page, p);
719    }
720
721    if (!s->offset && val == SLUB_RED_ACTIVE)
722        /*
723         * Object and freepointer overlap. Cannot check
724         * freepointer while object is allocated.
725         */
726        return 1;
727
728    /* Check free pointer validity */
729    if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
730        object_err(s, page, p, "Freepointer corrupt");
731        /*
732         * No choice but to zap it and thus lose the remainder
733         * of the free objects in this slab. May cause
734         * another error because the object count is now wrong.
735         */
736        set_freepointer(s, p, NULL);
737        return 0;
738    }
739    return 1;
740}
741
742static int check_slab(struct kmem_cache *s, struct page *page)
743{
744    int maxobj;
745
746    VM_BUG_ON(!irqs_disabled());
747
748    if (!PageSlab(page)) {
749        slab_err(s, page, "Not a valid slab page");
750        return 0;
751    }
752
753    maxobj = order_objects(compound_order(page), s->size, s->reserved);
754    if (page->objects > maxobj) {
755        slab_err(s, page, "objects %u > max %u",
756            s->name, page->objects, maxobj);
757        return 0;
758    }
759    if (page->inuse > page->objects) {
760        slab_err(s, page, "inuse %u > max %u",
761            s->name, page->inuse, page->objects);
762        return 0;
763    }
764    /* Slab_pad_check fixes things up after itself */
765    slab_pad_check(s, page);
766    return 1;
767}
768
769/*
770 * Determine if a certain object on a page is on the freelist. Must hold the
771 * slab lock to guarantee that the chains are in a consistent state.
772 */
773static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
774{
775    int nr = 0;
776    void *fp = page->freelist;
777    void *object = NULL;
778    unsigned long max_objects;
779
780    while (fp && nr <= page->objects) {
781        if (fp == search)
782            return 1;
783        if (!check_valid_pointer(s, page, fp)) {
784            if (object) {
785                object_err(s, page, object,
786                    "Freechain corrupt");
787                set_freepointer(s, object, NULL);
788                break;
789            } else {
790                slab_err(s, page, "Freepointer corrupt");
791                page->freelist = NULL;
792                page->inuse = page->objects;
793                slab_fix(s, "Freelist cleared");
794                return 0;
795            }
796            break;
797        }
798        object = fp;
799        fp = get_freepointer(s, object);
800        nr++;
801    }
802
803    max_objects = order_objects(compound_order(page), s->size, s->reserved);
804    if (max_objects > MAX_OBJS_PER_PAGE)
805        max_objects = MAX_OBJS_PER_PAGE;
806
807    if (page->objects != max_objects) {
808        slab_err(s, page, "Wrong number of objects. Found %d but "
809            "should be %d", page->objects, max_objects);
810        page->objects = max_objects;
811        slab_fix(s, "Number of objects adjusted.");
812    }
813    if (page->inuse != page->objects - nr) {
814        slab_err(s, page, "Wrong object count. Counter is %d but "
815            "counted were %d", page->inuse, page->objects - nr);
816        page->inuse = page->objects - nr;
817        slab_fix(s, "Object count adjusted.");
818    }
819    return search == NULL;
820}
821
822static void trace(struct kmem_cache *s, struct page *page, void *object,
823                                int alloc)
824{
825    if (s->flags & SLAB_TRACE) {
826        printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
827            s->name,
828            alloc ? "alloc" : "free",
829            object, page->inuse,
830            page->freelist);
831
832        if (!alloc)
833            print_section("Object", (void *)object, s->objsize);
834
835        dump_stack();
836    }
837}
838
839/*
840 * Hooks for other subsystems that check memory allocations. In a typical
841 * production configuration these hooks all should produce no code at all.
842 */
843static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
844{
845    flags &= gfp_allowed_mask;
846    lockdep_trace_alloc(flags);
847    might_sleep_if(flags & __GFP_WAIT);
848
849    return should_failslab(s->objsize, flags, s->flags);
850}
851
852static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
853{
854    flags &= gfp_allowed_mask;
855    kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
856    kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, flags);
857}
858
859static inline void slab_free_hook(struct kmem_cache *s, void *x)
860{
861    kmemleak_free_recursive(x, s->flags);
862
863    /*
864     * Trouble is that we may no longer disable interupts in the fast path
865     * So in order to make the debug calls that expect irqs to be
866     * disabled we need to disable interrupts temporarily.
867     */
868#if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
869    {
870        unsigned long flags;
871
872        local_irq_save(flags);
873        kmemcheck_slab_free(s, x, s->objsize);
874        debug_check_no_locks_freed(x, s->objsize);
875        local_irq_restore(flags);
876    }
877#endif
878    if (!(s->flags & SLAB_DEBUG_OBJECTS))
879        debug_check_no_obj_freed(x, s->objsize);
880}
881
882/*
883 * Tracking of fully allocated slabs for debugging purposes.
884 */
885static void add_full(struct kmem_cache_node *n, struct page *page)
886{
887    spin_lock(&n->list_lock);
888    list_add(&page->lru, &n->full);
889    spin_unlock(&n->list_lock);
890}
891
892static void remove_full(struct kmem_cache *s, struct page *page)
893{
894    struct kmem_cache_node *n;
895
896    if (!(s->flags & SLAB_STORE_USER))
897        return;
898
899    n = get_node(s, page_to_nid(page));
900
901    spin_lock(&n->list_lock);
902    list_del(&page->lru);
903    spin_unlock(&n->list_lock);
904}
905
906/* Tracking of the number of slabs for debugging purposes */
907static inline unsigned long slabs_node(struct kmem_cache *s, int node)
908{
909    struct kmem_cache_node *n = get_node(s, node);
910
911    return atomic_long_read(&n->nr_slabs);
912}
913
914static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
915{
916    return atomic_long_read(&n->nr_slabs);
917}
918
919static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
920{
921    struct kmem_cache_node *n = get_node(s, node);
922
923    /*
924     * May be called early in order to allocate a slab for the
925     * kmem_cache_node structure. Solve the chicken-egg
926     * dilemma by deferring the increment of the count during
927     * bootstrap (see early_kmem_cache_node_alloc).
928     */
929    if (n) {
930        atomic_long_inc(&n->nr_slabs);
931        atomic_long_add(objects, &n->total_objects);
932    }
933}
934static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
935{
936    struct kmem_cache_node *n = get_node(s, node);
937
938    atomic_long_dec(&n->nr_slabs);
939    atomic_long_sub(objects, &n->total_objects);
940}
941
942/* Object debug checks for alloc/free paths */
943static void setup_object_debug(struct kmem_cache *s, struct page *page,
944                                void *object)
945{
946    if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
947        return;
948
949    init_object(s, object, SLUB_RED_INACTIVE);
950    init_tracking(s, object);
951}
952
953static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
954                    void *object, unsigned long addr)
955{
956    if (!check_slab(s, page))
957        goto bad;
958
959    if (!on_freelist(s, page, object)) {
960        object_err(s, page, object, "Object already allocated");
961        goto bad;
962    }
963
964    if (!check_valid_pointer(s, page, object)) {
965        object_err(s, page, object, "Freelist Pointer check fails");
966        goto bad;
967    }
968
969    if (!check_object(s, page, object, SLUB_RED_INACTIVE))
970        goto bad;
971
972    /* Success perform special debug activities for allocs */
973    if (s->flags & SLAB_STORE_USER)
974        set_track(s, object, TRACK_ALLOC, addr);
975    trace(s, page, object, 1);
976    init_object(s, object, SLUB_RED_ACTIVE);
977    return 1;
978
979bad:
980    if (PageSlab(page)) {
981        /*
982         * If this is a slab page then lets do the best we can
983         * to avoid issues in the future. Marking all objects
984         * as used avoids touching the remaining objects.
985         */
986        slab_fix(s, "Marking all objects used");
987        page->inuse = page->objects;
988        page->freelist = NULL;
989    }
990    return 0;
991}
992
993static noinline int free_debug_processing(struct kmem_cache *s,
994         struct page *page, void *object, unsigned long addr)
995{
996    if (!check_slab(s, page))
997        goto fail;
998
999    if (!check_valid_pointer(s, page, object)) {
1000        slab_err(s, page, "Invalid object pointer 0x%p", object);
1001        goto fail;
1002    }
1003
1004    if (on_freelist(s, page, object)) {
1005        object_err(s, page, object, "Object already free");
1006        goto fail;
1007    }
1008
1009    if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1010        return 0;
1011
1012    if (unlikely(s != page->slab)) {
1013        if (!PageSlab(page)) {
1014            slab_err(s, page, "Attempt to free object(0x%p) "
1015                "outside of slab", object);
1016        } else if (!page->slab) {
1017            printk(KERN_ERR
1018                "SLUB <none>: no slab for object 0x%p.\n",
1019                        object);
1020            dump_stack();
1021        } else
1022            object_err(s, page, object,
1023                    "page slab pointer corrupt.");
1024        goto fail;
1025    }
1026
1027    /* Special debug activities for freeing objects */
1028    if (!PageSlubFrozen(page) && !page->freelist)
1029        remove_full(s, page);
1030    if (s->flags & SLAB_STORE_USER)
1031        set_track(s, object, TRACK_FREE, addr);
1032    trace(s, page, object, 0);
1033    init_object(s, object, SLUB_RED_INACTIVE);
1034    return 1;
1035
1036fail:
1037    slab_fix(s, "Object at 0x%p not freed", object);
1038    return 0;
1039}
1040
1041static int __init setup_slub_debug(char *str)
1042{
1043    slub_debug = DEBUG_DEFAULT_FLAGS;
1044    if (*str++ != '=' || !*str)
1045        /*
1046         * No options specified. Switch on full debugging.
1047         */
1048        goto out;
1049
1050    if (*str == ',')
1051        /*
1052         * No options but restriction on slabs. This means full
1053         * debugging for slabs matching a pattern.
1054         */
1055        goto check_slabs;
1056
1057    if (tolower(*str) == 'o') {
1058        /*
1059         * Avoid enabling debugging on caches if its minimum order
1060         * would increase as a result.
1061         */
1062        disable_higher_order_debug = 1;
1063        goto out;
1064    }
1065
1066    slub_debug = 0;
1067    if (*str == '-')
1068        /*
1069         * Switch off all debugging measures.
1070         */
1071        goto out;
1072
1073    /*
1074     * Determine which debug features should be switched on
1075     */
1076    for (; *str && *str != ','; str++) {
1077        switch (tolower(*str)) {
1078        case 'f':
1079            slub_debug |= SLAB_DEBUG_FREE;
1080            break;
1081        case 'z':
1082            slub_debug |= SLAB_RED_ZONE;
1083            break;
1084        case 'p':
1085            slub_debug |= SLAB_POISON;
1086            break;
1087        case 'u':
1088            slub_debug |= SLAB_STORE_USER;
1089            break;
1090        case 't':
1091            slub_debug |= SLAB_TRACE;
1092            break;
1093        case 'a':
1094            slub_debug |= SLAB_FAILSLAB;
1095            break;
1096        default:
1097            printk(KERN_ERR "slub_debug option '%c' "
1098                "unknown. skipped\n", *str);
1099        }
1100    }
1101
1102check_slabs:
1103    if (*str == ',')
1104        slub_debug_slabs = str + 1;
1105out:
1106    return 1;
1107}
1108
1109__setup("slub_debug", setup_slub_debug);
1110
1111static unsigned long kmem_cache_flags(unsigned long objsize,
1112    unsigned long flags, const char *name,
1113    void (*ctor)(void *))
1114{
1115    /*
1116     * Enable debugging if selected on the kernel commandline.
1117     */
1118    if (slub_debug && (!slub_debug_slabs ||
1119        !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1120        flags |= slub_debug;
1121
1122    return flags;
1123}
1124#else
1125static inline void setup_object_debug(struct kmem_cache *s,
1126            struct page *page, void *object) {}
1127
1128static inline int alloc_debug_processing(struct kmem_cache *s,
1129    struct page *page, void *object, unsigned long addr) { return 0; }
1130
1131static inline int free_debug_processing(struct kmem_cache *s,
1132    struct page *page, void *object, unsigned long addr) { return 0; }
1133
1134static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1135            { return 1; }
1136static inline int check_object(struct kmem_cache *s, struct page *page,
1137            void *object, u8 val) { return 1; }
1138static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
1139static inline unsigned long kmem_cache_flags(unsigned long objsize,
1140    unsigned long flags, const char *name,
1141    void (*ctor)(void *))
1142{
1143    return flags;
1144}
1145#define slub_debug 0
1146
1147#define disable_higher_order_debug 0
1148
1149static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1150                            { return 0; }
1151static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1152                            { return 0; }
1153static inline void inc_slabs_node(struct kmem_cache *s, int node,
1154                            int objects) {}
1155static inline void dec_slabs_node(struct kmem_cache *s, int node,
1156                            int objects) {}
1157
1158static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1159                            { return 0; }
1160
1161static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1162        void *object) {}
1163
1164static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1165
1166#endif /* CONFIG_SLUB_DEBUG */
1167
1168/*
1169 * Slab allocation and freeing
1170 */
1171static inline struct page *alloc_slab_page(gfp_t flags, int node,
1172                    struct kmem_cache_order_objects oo)
1173{
1174    int order = oo_order(oo);
1175
1176    flags |= __GFP_NOTRACK;
1177
1178    if (node == NUMA_NO_NODE)
1179        return alloc_pages(flags, order);
1180    else
1181        return alloc_pages_exact_node(node, flags, order);
1182}
1183
1184static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1185{
1186    struct page *page;
1187    struct kmem_cache_order_objects oo = s->oo;
1188    gfp_t alloc_gfp;
1189
1190    flags |= s->allocflags;
1191
1192    /*
1193     * Let the initial higher-order allocation fail under memory pressure
1194     * so we fall-back to the minimum order allocation.
1195     */
1196    alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1197
1198    page = alloc_slab_page(alloc_gfp, node, oo);
1199    if (unlikely(!page)) {
1200        oo = s->min;
1201        /*
1202         * Allocation may have failed due to fragmentation.
1203         * Try a lower order alloc if possible
1204         */
1205        page = alloc_slab_page(flags, node, oo);
1206        if (!page)
1207            return NULL;
1208
1209        stat(s, ORDER_FALLBACK);
1210    }
1211
1212    if (kmemcheck_enabled
1213        && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1214        int pages = 1 << oo_order(oo);
1215
1216        kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1217
1218        /*
1219         * Objects from caches that have a constructor don't get
1220         * cleared when they're allocated, so we need to do it here.
1221         */
1222        if (s->ctor)
1223            kmemcheck_mark_uninitialized_pages(page, pages);
1224        else
1225            kmemcheck_mark_unallocated_pages(page, pages);
1226    }
1227
1228    page->objects = oo_objects(oo);
1229    mod_zone_page_state(page_zone(page),
1230        (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1231        NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1232        1 << oo_order(oo));
1233
1234    return page;
1235}
1236
1237static void setup_object(struct kmem_cache *s, struct page *page,
1238                void *object)
1239{
1240    setup_object_debug(s, page, object);
1241    if (unlikely(s->ctor))
1242        s->ctor(object);
1243}
1244
1245static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1246{
1247    struct page *page;
1248    void *start;
1249    void *last;
1250    void *p;
1251
1252    BUG_ON(flags & GFP_SLAB_BUG_MASK);
1253
1254    page = allocate_slab(s,
1255        flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1256    if (!page)
1257        goto out;
1258
1259    inc_slabs_node(s, page_to_nid(page), page->objects);
1260    page->slab = s;
1261    page->flags |= 1 << PG_slab;
1262
1263    start = page_address(page);
1264
1265    if (unlikely(s->flags & SLAB_POISON))
1266        memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1267
1268    last = start;
1269    for_each_object(p, s, start, page->objects) {
1270        setup_object(s, page, last);
1271        set_freepointer(s, last, p);
1272        last = p;
1273    }
1274    setup_object(s, page, last);
1275    set_freepointer(s, last, NULL);
1276
1277    page->freelist = start;
1278    page->inuse = 0;
1279out:
1280    return page;
1281}
1282
1283static void __free_slab(struct kmem_cache *s, struct page *page)
1284{
1285    int order = compound_order(page);
1286    int pages = 1 << order;
1287
1288    if (kmem_cache_debug(s)) {
1289        void *p;
1290
1291        slab_pad_check(s, page);
1292        for_each_object(p, s, page_address(page),
1293                        page->objects)
1294            check_object(s, page, p, SLUB_RED_INACTIVE);
1295    }
1296
1297    kmemcheck_free_shadow(page, compound_order(page));
1298
1299    mod_zone_page_state(page_zone(page),
1300        (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1301        NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1302        -pages);
1303
1304    __ClearPageSlab(page);
1305    reset_page_mapcount(page);
1306    if (current->reclaim_state)
1307        current->reclaim_state->reclaimed_slab += pages;
1308    __free_pages(page, order);
1309}
1310
1311#define need_reserve_slab_rcu \
1312    (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1313
1314static void rcu_free_slab(struct rcu_head *h)
1315{
1316    struct page *page;
1317
1318    if (need_reserve_slab_rcu)
1319        page = virt_to_head_page(h);
1320    else
1321        page = container_of((struct list_head *)h, struct page, lru);
1322
1323    __free_slab(page->slab, page);
1324}
1325
1326static void free_slab(struct kmem_cache *s, struct page *page)
1327{
1328    if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1329        struct rcu_head *head;
1330
1331        if (need_reserve_slab_rcu) {
1332            int order = compound_order(page);
1333            int offset = (PAGE_SIZE << order) - s->reserved;
1334
1335            VM_BUG_ON(s->reserved != sizeof(*head));
1336            head = page_address(page) + offset;
1337        } else {
1338            /*
1339             * RCU free overloads the RCU head over the LRU
1340             */
1341            head = (void *)&page->lru;
1342        }
1343
1344        call_rcu(head, rcu_free_slab);
1345    } else
1346        __free_slab(s, page);
1347}
1348
1349static void discard_slab(struct kmem_cache *s, struct page *page)
1350{
1351    dec_slabs_node(s, page_to_nid(page), page->objects);
1352    free_slab(s, page);
1353}
1354
1355/*
1356 * Per slab locking using the pagelock
1357 */
1358static __always_inline void slab_lock(struct page *page)
1359{
1360    bit_spin_lock(PG_locked, &page->flags);
1361}
1362
1363static __always_inline void slab_unlock(struct page *page)
1364{
1365    __bit_spin_unlock(PG_locked, &page->flags);
1366}
1367
1368static __always_inline int slab_trylock(struct page *page)
1369{
1370    int rc = 1;
1371
1372    rc = bit_spin_trylock(PG_locked, &page->flags);
1373    return rc;
1374}
1375
1376/*
1377 * Management of partially allocated slabs
1378 */
1379static void add_partial(struct kmem_cache_node *n,
1380                struct page *page, int tail)
1381{
1382    spin_lock(&n->list_lock);
1383    n->nr_partial++;
1384    if (tail)
1385        list_add_tail(&page->lru, &n->partial);
1386    else
1387        list_add(&page->lru, &n->partial);
1388    spin_unlock(&n->list_lock);
1389}
1390
1391static inline void __remove_partial(struct kmem_cache_node *n,
1392                    struct page *page)
1393{
1394    list_del(&page->lru);
1395    n->nr_partial--;
1396}
1397
1398static void remove_partial(struct kmem_cache *s, struct page *page)
1399{
1400    struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1401
1402    spin_lock(&n->list_lock);
1403    __remove_partial(n, page);
1404    spin_unlock(&n->list_lock);
1405}
1406
1407/*
1408 * Lock slab and remove from the partial list.
1409 *
1410 * Must hold list_lock.
1411 */
1412static inline int lock_and_freeze_slab(struct kmem_cache_node *n,
1413                            struct page *page)
1414{
1415    if (slab_trylock(page)) {
1416        __remove_partial(n, page);
1417        __SetPageSlubFrozen(page);
1418        return 1;
1419    }
1420    return 0;
1421}
1422
1423/*
1424 * Try to allocate a partial slab from a specific node.
1425 */
1426static struct page *get_partial_node(struct kmem_cache_node *n)
1427{
1428    struct page *page;
1429
1430    /*
1431     * Racy check. If we mistakenly see no partial slabs then we
1432     * just allocate an empty slab. If we mistakenly try to get a
1433     * partial slab and there is none available then get_partials()
1434     * will return NULL.
1435     */
1436    if (!n || !n->nr_partial)
1437        return NULL;
1438
1439    spin_lock(&n->list_lock);
1440    list_for_each_entry(page, &n->partial, lru)
1441        if (lock_and_freeze_slab(n, page))
1442            goto out;
1443    page = NULL;
1444out:
1445    spin_unlock(&n->list_lock);
1446    return page;
1447}
1448
1449/*
1450 * Get a page from somewhere. Search in increasing NUMA distances.
1451 */
1452static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1453{
1454#ifdef CONFIG_NUMA
1455    struct zonelist *zonelist;
1456    struct zoneref *z;
1457    struct zone *zone;
1458    enum zone_type high_zoneidx = gfp_zone(flags);
1459    struct page *page;
1460
1461    /*
1462     * The defrag ratio allows a configuration of the tradeoffs between
1463     * inter node defragmentation and node local allocations. A lower
1464     * defrag_ratio increases the tendency to do local allocations
1465     * instead of attempting to obtain partial slabs from other nodes.
1466     *
1467     * If the defrag_ratio is set to 0 then kmalloc() always
1468     * returns node local objects. If the ratio is higher then kmalloc()
1469     * may return off node objects because partial slabs are obtained
1470     * from other nodes and filled up.
1471     *
1472     * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1473     * defrag_ratio = 1000) then every (well almost) allocation will
1474     * first attempt to defrag slab caches on other nodes. This means
1475     * scanning over all nodes to look for partial slabs which may be
1476     * expensive if we do it every time we are trying to find a slab
1477     * with available objects.
1478     */
1479    if (!s->remote_node_defrag_ratio ||
1480            get_cycles() % 1024 > s->remote_node_defrag_ratio)
1481        return NULL;
1482
1483    get_mems_allowed();
1484    zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1485    for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1486        struct kmem_cache_node *n;
1487
1488        n = get_node(s, zone_to_nid(zone));
1489
1490        if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1491                n->nr_partial > s->min_partial) {
1492            page = get_partial_node(n);
1493            if (page) {
1494                put_mems_allowed();
1495                return page;
1496            }
1497        }
1498    }
1499    put_mems_allowed();
1500#endif
1501    return NULL;
1502}
1503
1504/*
1505 * Get a partial page, lock it and return it.
1506 */
1507static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1508{
1509    struct page *page;
1510    int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1511
1512    page = get_partial_node(get_node(s, searchnode));
1513    if (page || node != NUMA_NO_NODE)
1514        return page;
1515
1516    return get_any_partial(s, flags);
1517}
1518
1519/*
1520 * Move a page back to the lists.
1521 *
1522 * Must be called with the slab lock held.
1523 *
1524 * On exit the slab lock will have been dropped.
1525 */
1526static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
1527    __releases(bitlock)
1528{
1529    struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1530
1531    __ClearPageSlubFrozen(page);
1532    if (page->inuse) {
1533
1534        if (page->freelist) {
1535            add_partial(n, page, tail);
1536            stat(s, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1537        } else {
1538            stat(s, DEACTIVATE_FULL);
1539            if (kmem_cache_debug(s) && (s->flags & SLAB_STORE_USER))
1540                add_full(n, page);
1541        }
1542        slab_unlock(page);
1543    } else {
1544        stat(s, DEACTIVATE_EMPTY);
1545        if (n->nr_partial < s->min_partial) {
1546            /*
1547             * Adding an empty slab to the partial slabs in order
1548             * to avoid page allocator overhead. This slab needs
1549             * to come after the other slabs with objects in
1550             * so that the others get filled first. That way the
1551             * size of the partial list stays small.
1552             *
1553             * kmem_cache_shrink can reclaim any empty slabs from
1554             * the partial list.
1555             */
1556            add_partial(n, page, 1);
1557            slab_unlock(page);
1558        } else {
1559            slab_unlock(page);
1560            stat(s, FREE_SLAB);
1561            discard_slab(s, page);
1562        }
1563    }
1564}
1565
1566#ifdef CONFIG_PREEMPT
1567/*
1568 * Calculate the next globally unique transaction for disambiguiation
1569 * during cmpxchg. The transactions start with the cpu number and are then
1570 * incremented by CONFIG_NR_CPUS.
1571 */
1572#define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1573#else
1574/*
1575 * No preemption supported therefore also no need to check for
1576 * different cpus.
1577 */
1578#define TID_STEP 1
1579#endif
1580
1581static inline unsigned long next_tid(unsigned long tid)
1582{
1583    return tid + TID_STEP;
1584}
1585
1586static inline unsigned int tid_to_cpu(unsigned long tid)
1587{
1588    return tid % TID_STEP;
1589}
1590
1591static inline unsigned long tid_to_event(unsigned long tid)
1592{
1593    return tid / TID_STEP;
1594}
1595
1596static inline unsigned int init_tid(int cpu)
1597{
1598    return cpu;
1599}
1600
1601static inline void note_cmpxchg_failure(const char *n,
1602        const struct kmem_cache *s, unsigned long tid)
1603{
1604#ifdef SLUB_DEBUG_CMPXCHG
1605    unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1606
1607    printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1608
1609#ifdef CONFIG_PREEMPT
1610    if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1611        printk("due to cpu change %d -> %d\n",
1612            tid_to_cpu(tid), tid_to_cpu(actual_tid));
1613    else
1614#endif
1615    if (tid_to_event(tid) != tid_to_event(actual_tid))
1616        printk("due to cpu running other code. Event %ld->%ld\n",
1617            tid_to_event(tid), tid_to_event(actual_tid));
1618    else
1619        printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1620            actual_tid, tid, next_tid(tid));
1621#endif
1622    stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1623}
1624
1625void init_kmem_cache_cpus(struct kmem_cache *s)
1626{
1627    int cpu;
1628
1629    for_each_possible_cpu(cpu)
1630        per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1631}
1632/*
1633 * Remove the cpu slab
1634 */
1635static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1636    __releases(bitlock)
1637{
1638    struct page *page = c->page;
1639    int tail = 1;
1640
1641    if (page->freelist)
1642        stat(s, DEACTIVATE_REMOTE_FREES);
1643    /*
1644     * Merge cpu freelist into slab freelist. Typically we get here
1645     * because both freelists are empty. So this is unlikely
1646     * to occur.
1647     */
1648    while (unlikely(c->freelist)) {
1649        void **object;
1650
1651        tail = 0; /* Hot objects. Put the slab first */
1652
1653        /* Retrieve object from cpu_freelist */
1654        object = c->freelist;
1655        c->freelist = get_freepointer(s, c->freelist);
1656
1657        /* And put onto the regular freelist */
1658        set_freepointer(s, object, page->freelist);
1659        page->freelist = object;
1660        page->inuse--;
1661    }
1662    c->page = NULL;
1663    c->tid = next_tid(c->tid);
1664    unfreeze_slab(s, page, tail);
1665}
1666
1667static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1668{
1669    stat(s, CPUSLAB_FLUSH);
1670    slab_lock(c->page);
1671    deactivate_slab(s, c);
1672}
1673
1674/*
1675 * Flush cpu slab.
1676 *
1677 * Called from IPI handler with interrupts disabled.
1678 */
1679static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1680{
1681    struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
1682
1683    if (likely(c && c->page))
1684        flush_slab(s, c);
1685}
1686
1687static void flush_cpu_slab(void *d)
1688{
1689    struct kmem_cache *s = d;
1690
1691    __flush_cpu_slab(s, smp_processor_id());
1692}
1693
1694static void flush_all(struct kmem_cache *s)
1695{
1696    on_each_cpu(flush_cpu_slab, s, 1);
1697}
1698
1699/*
1700 * Check if the objects in a per cpu structure fit numa
1701 * locality expectations.
1702 */
1703static inline int node_match(struct kmem_cache_cpu *c, int node)
1704{
1705#ifdef CONFIG_NUMA
1706    if (node != NUMA_NO_NODE && c->node != node)
1707        return 0;
1708#endif
1709    return 1;
1710}
1711
1712static int count_free(struct page *page)
1713{
1714    return page->objects - page->inuse;
1715}
1716
1717static unsigned long count_partial(struct kmem_cache_node *n,
1718                    int (*get_count)(struct page *))
1719{
1720    unsigned long flags;
1721    unsigned long x = 0;
1722    struct page *page;
1723
1724    spin_lock_irqsave(&n->list_lock, flags);
1725    list_for_each_entry(page, &n->partial, lru)
1726        x += get_count(page);
1727    spin_unlock_irqrestore(&n->list_lock, flags);
1728    return x;
1729}
1730
1731static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
1732{
1733#ifdef CONFIG_SLUB_DEBUG
1734    return atomic_long_read(&n->total_objects);
1735#else
1736    return 0;
1737#endif
1738}
1739
1740static noinline void
1741slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
1742{
1743    int node;
1744
1745    printk(KERN_WARNING
1746        "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1747        nid, gfpflags);
1748    printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
1749        "default order: %d, min order: %d\n", s->name, s->objsize,
1750        s->size, oo_order(s->oo), oo_order(s->min));
1751
1752    if (oo_order(s->min) > get_order(s->objsize))
1753        printk(KERN_WARNING " %s debugging increased min order, use "
1754               "slub_debug=O to disable.\n", s->name);
1755
1756    for_each_online_node(node) {
1757        struct kmem_cache_node *n = get_node(s, node);
1758        unsigned long nr_slabs;
1759        unsigned long nr_objs;
1760        unsigned long nr_free;
1761
1762        if (!n)
1763            continue;
1764
1765        nr_free = count_partial(n, count_free);
1766        nr_slabs = node_nr_slabs(n);
1767        nr_objs = node_nr_objs(n);
1768
1769        printk(KERN_WARNING
1770            " node %d: slabs: %ld, objs: %ld, free: %ld\n",
1771            node, nr_slabs, nr_objs, nr_free);
1772    }
1773}
1774
1775/*
1776 * Slow path. The lockless freelist is empty or we need to perform
1777 * debugging duties.
1778 *
1779 * Interrupts are disabled.
1780 *
1781 * Processing is still very fast if new objects have been freed to the
1782 * regular freelist. In that case we simply take over the regular freelist
1783 * as the lockless freelist and zap the regular freelist.
1784 *
1785 * If that is not working then we fall back to the partial lists. We take the
1786 * first element of the freelist as the object to allocate now and move the
1787 * rest of the freelist to the lockless freelist.
1788 *
1789 * And if we were unable to get a new slab from the partial slab lists then
1790 * we need to allocate a new slab. This is the slowest path since it involves
1791 * a call to the page allocator and the setup of a new slab.
1792 */
1793static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
1794              unsigned long addr, struct kmem_cache_cpu *c)
1795{
1796    void **object;
1797    struct page *page;
1798    unsigned long flags;
1799
1800    local_irq_save(flags);
1801#ifdef CONFIG_PREEMPT
1802    /*
1803     * We may have been preempted and rescheduled on a different
1804     * cpu before disabling interrupts. Need to reload cpu area
1805     * pointer.
1806     */
1807    c = this_cpu_ptr(s->cpu_slab);
1808#endif
1809
1810    /* We handle __GFP_ZERO in the caller */
1811    gfpflags &= ~__GFP_ZERO;
1812
1813    page = c->page;
1814    if (!page)
1815        goto new_slab;
1816
1817    slab_lock(page);
1818    if (unlikely(!node_match(c, node)))
1819        goto another_slab;
1820
1821    stat(s, ALLOC_REFILL);
1822
1823load_freelist:
1824    object = page->freelist;
1825    if (unlikely(!object))
1826        goto another_slab;
1827    if (kmem_cache_debug(s))
1828        goto debug;
1829
1830    c->freelist = get_freepointer(s, object);
1831    page->inuse = page->objects;
1832    page->freelist = NULL;
1833
1834    slab_unlock(page);
1835    c->tid = next_tid(c->tid);
1836    local_irq_restore(flags);
1837    stat(s, ALLOC_SLOWPATH);
1838    return object;
1839
1840another_slab:
1841    deactivate_slab(s, c);
1842
1843new_slab:
1844    page = get_partial(s, gfpflags, node);
1845    if (page) {
1846        stat(s, ALLOC_FROM_PARTIAL);
1847        c->node = page_to_nid(page);
1848        c->page = page;
1849        goto load_freelist;
1850    }
1851
1852    gfpflags &= gfp_allowed_mask;
1853    if (gfpflags & __GFP_WAIT)
1854        local_irq_enable();
1855
1856    page = new_slab(s, gfpflags, node);
1857
1858    if (gfpflags & __GFP_WAIT)
1859        local_irq_disable();
1860
1861    if (page) {
1862        c = __this_cpu_ptr(s->cpu_slab);
1863        stat(s, ALLOC_SLAB);
1864        if (c->page)
1865            flush_slab(s, c);
1866
1867        slab_lock(page);
1868        __SetPageSlubFrozen(page);
1869        c->node = page_to_nid(page);
1870        c->page = page;
1871        goto load_freelist;
1872    }
1873    if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
1874        slab_out_of_memory(s, gfpflags, node);
1875    local_irq_restore(flags);
1876    return NULL;
1877debug:
1878    if (!alloc_debug_processing(s, page, object, addr))
1879        goto another_slab;
1880
1881    page->inuse++;
1882    page->freelist = get_freepointer(s, object);
1883    deactivate_slab(s, c);
1884    c->page = NULL;
1885    c->node = NUMA_NO_NODE;
1886    local_irq_restore(flags);
1887    return object;
1888}
1889
1890/*
1891 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1892 * have the fastpath folded into their functions. So no function call
1893 * overhead for requests that can be satisfied on the fastpath.
1894 *
1895 * The fastpath works by first checking if the lockless freelist can be used.
1896 * If not then __slab_alloc is called for slow processing.
1897 *
1898 * Otherwise we can simply pick the next object from the lockless free list.
1899 */
1900static __always_inline void *slab_alloc(struct kmem_cache *s,
1901        gfp_t gfpflags, int node, unsigned long addr)
1902{
1903    void **object;
1904    struct kmem_cache_cpu *c;
1905    unsigned long tid;
1906
1907    if (slab_pre_alloc_hook(s, gfpflags))
1908        return NULL;
1909
1910redo:
1911
1912    /*
1913     * Must read kmem_cache cpu data via this cpu ptr. Preemption is
1914     * enabled. We may switch back and forth between cpus while
1915     * reading from one cpu area. That does not matter as long
1916     * as we end up on the original cpu again when doing the cmpxchg.
1917     */
1918    c = __this_cpu_ptr(s->cpu_slab);
1919
1920    /*
1921     * The transaction ids are globally unique per cpu and per operation on
1922     * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
1923     * occurs on the right processor and that there was no operation on the
1924     * linked list in between.
1925     */
1926    tid = c->tid;
1927    barrier();
1928
1929    object = c->freelist;
1930    if (unlikely(!object || !node_match(c, node)))
1931
1932        object = __slab_alloc(s, gfpflags, node, addr, c);
1933
1934    else {
1935        /*
1936         * The cmpxchg will only match if there was no additional
1937         * operation and if we are on the right processor.
1938         *
1939         * The cmpxchg does the following atomically (without lock semantics!)
1940         * 1. Relocate first pointer to the current per cpu area.
1941         * 2. Verify that tid and freelist have not been changed
1942         * 3. If they were not changed replace tid and freelist
1943         *
1944         * Since this is without lock semantics the protection is only against
1945         * code executing on this cpu *not* from access by other cpus.
1946         */
1947        if (unlikely(!irqsafe_cpu_cmpxchg_double(
1948                s->cpu_slab->freelist, s->cpu_slab->tid,
1949                object, tid,
1950                get_freepointer_safe(s, object), next_tid(tid)))) {
1951
1952            note_cmpxchg_failure("slab_alloc", s, tid);
1953            goto redo;
1954        }
1955        stat(s, ALLOC_FASTPATH);
1956    }
1957
1958    if (unlikely(gfpflags & __GFP_ZERO) && object)
1959        memset(object, 0, s->objsize);
1960
1961    slab_post_alloc_hook(s, gfpflags, object);
1962
1963    return object;
1964}
1965
1966void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1967{
1968    void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
1969
1970    trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
1971
1972    return ret;
1973}
1974EXPORT_SYMBOL(kmem_cache_alloc);
1975
1976#ifdef CONFIG_TRACING
1977void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
1978{
1979    void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
1980    trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
1981    return ret;
1982}
1983EXPORT_SYMBOL(kmem_cache_alloc_trace);
1984
1985void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
1986{
1987    void *ret = kmalloc_order(size, flags, order);
1988    trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
1989    return ret;
1990}
1991EXPORT_SYMBOL(kmalloc_order_trace);
1992#endif
1993
1994#ifdef CONFIG_NUMA
1995void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1996{
1997    void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
1998
1999    trace_kmem_cache_alloc_node(_RET_IP_, ret,
2000                    s->objsize, s->size, gfpflags, node);
2001
2002    return ret;
2003}
2004EXPORT_SYMBOL(kmem_cache_alloc_node);
2005
2006#ifdef CONFIG_TRACING
2007void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2008                    gfp_t gfpflags,
2009                    int node, size_t size)
2010{
2011    void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2012
2013    trace_kmalloc_node(_RET_IP_, ret,
2014               size, s->size, gfpflags, node);
2015    return ret;
2016}
2017EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2018#endif
2019#endif
2020
2021/*
2022 * Slow patch handling. This may still be called frequently since objects
2023 * have a longer lifetime than the cpu slabs in most processing loads.
2024 *
2025 * So we still attempt to reduce cache line usage. Just take the slab
2026 * lock and free the item. If there is no additional partial page
2027 * handling required then we can return immediately.
2028 */
2029static void __slab_free(struct kmem_cache *s, struct page *page,
2030            void *x, unsigned long addr)
2031{
2032    void *prior;
2033    void **object = (void *)x;
2034    unsigned long flags;
2035
2036    local_irq_save(flags);
2037    slab_lock(page);
2038    stat(s, FREE_SLOWPATH);
2039
2040    if (kmem_cache_debug(s) && !free_debug_processing(s, page, x, addr))
2041        goto out_unlock;
2042
2043    prior = page->freelist;
2044    set_freepointer(s, object, prior);
2045    page->freelist = object;
2046    page->inuse--;
2047
2048    if (unlikely(PageSlubFrozen(page))) {
2049        stat(s, FREE_FROZEN);
2050        goto out_unlock;
2051    }
2052
2053    if (unlikely(!page->inuse))
2054        goto slab_empty;
2055
2056    /*
2057     * Objects left in the slab. If it was not on the partial list before
2058     * then add it.
2059     */
2060    if (unlikely(!prior)) {
2061        add_partial(get_node(s, page_to_nid(page)), page, 1);
2062        stat(s, FREE_ADD_PARTIAL);
2063    }
2064
2065out_unlock:
2066    slab_unlock(page);
2067    local_irq_restore(flags);
2068    return;
2069
2070slab_empty:
2071    if (prior) {
2072        /*
2073         * Slab still on the partial list.
2074         */
2075        remove_partial(s, page);
2076        stat(s, FREE_REMOVE_PARTIAL);
2077    }
2078    slab_unlock(page);
2079    local_irq_restore(flags);
2080    stat(s, FREE_SLAB);
2081    discard_slab(s, page);
2082}
2083
2084/*
2085 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2086 * can perform fastpath freeing without additional function calls.
2087 *
2088 * The fastpath is only possible if we are freeing to the current cpu slab
2089 * of this processor. This typically the case if we have just allocated
2090 * the item before.
2091 *
2092 * If fastpath is not possible then fall back to __slab_free where we deal
2093 * with all sorts of special processing.
2094 */
2095static __always_inline void slab_free(struct kmem_cache *s,
2096            struct page *page, void *x, unsigned long addr)
2097{
2098    void **object = (void *)x;
2099    struct kmem_cache_cpu *c;
2100    unsigned long tid;
2101
2102    slab_free_hook(s, x);
2103
2104redo:
2105
2106    /*
2107     * Determine the currently cpus per cpu slab.
2108     * The cpu may change afterward. However that does not matter since
2109     * data is retrieved via this pointer. If we are on the same cpu
2110     * during the cmpxchg then the free will succedd.
2111     */
2112    c = __this_cpu_ptr(s->cpu_slab);
2113
2114    tid = c->tid;
2115    barrier();
2116
2117    if (likely(page == c->page)) {
2118        set_freepointer(s, object, c->freelist);
2119
2120        if (unlikely(!irqsafe_cpu_cmpxchg_double(
2121                s->cpu_slab->freelist, s->cpu_slab->tid,
2122                c->freelist, tid,
2123                object, next_tid(tid)))) {
2124
2125            note_cmpxchg_failure("slab_free", s, tid);
2126            goto redo;
2127        }
2128        stat(s, FREE_FASTPATH);
2129    } else
2130        __slab_free(s, page, x, addr);
2131
2132}
2133
2134void kmem_cache_free(struct kmem_cache *s, void *x)
2135{
2136    struct page *page;
2137
2138    page = virt_to_head_page(x);
2139
2140    slab_free(s, page, x, _RET_IP_);
2141
2142    trace_kmem_cache_free(_RET_IP_, x);
2143}
2144EXPORT_SYMBOL(kmem_cache_free);
2145
2146/*
2147 * Object placement in a slab is made very easy because we always start at
2148 * offset 0. If we tune the size of the object to the alignment then we can
2149 * get the required alignment by putting one properly sized object after
2150 * another.
2151 *
2152 * Notice that the allocation order determines the sizes of the per cpu
2153 * caches. Each processor has always one slab available for allocations.
2154 * Increasing the allocation order reduces the number of times that slabs
2155 * must be moved on and off the partial lists and is therefore a factor in
2156 * locking overhead.
2157 */
2158
2159/*
2160 * Mininum / Maximum order of slab pages. This influences locking overhead
2161 * and slab fragmentation. A higher order reduces the number of partial slabs
2162 * and increases the number of allocations possible without having to
2163 * take the list_lock.
2164 */
2165static int slub_min_order;
2166static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2167static int slub_min_objects;
2168
2169/*
2170 * Merge control. If this is set then no merging of slab caches will occur.
2171 * (Could be removed. This was introduced to pacify the merge skeptics.)
2172 */
2173static int slub_nomerge;
2174
2175/*
2176 * Calculate the order of allocation given an slab object size.
2177 *
2178 * The order of allocation has significant impact on performance and other
2179 * system components. Generally order 0 allocations should be preferred since
2180 * order 0 does not cause fragmentation in the page allocator. Larger objects
2181 * be problematic to put into order 0 slabs because there may be too much
2182 * unused space left. We go to a higher order if more than 1/16th of the slab
2183 * would be wasted.
2184 *
2185 * In order to reach satisfactory performance we must ensure that a minimum
2186 * number of objects is in one slab. Otherwise we may generate too much
2187 * activity on the partial lists which requires taking the list_lock. This is
2188 * less a concern for large slabs though which are rarely used.
2189 *
2190 * slub_max_order specifies the order where we begin to stop considering the
2191 * number of objects in a slab as critical. If we reach slub_max_order then
2192 * we try to keep the page order as low as possible. So we accept more waste
2193 * of space in favor of a small page order.
2194 *
2195 * Higher order allocations also allow the placement of more objects in a
2196 * slab and thereby reduce object handling overhead. If the user has
2197 * requested a higher mininum order then we start with that one instead of
2198 * the smallest order which will fit the object.
2199 */
2200static inline int slab_order(int size, int min_objects,
2201                int max_order, int fract_leftover, int reserved)
2202{
2203    int order;
2204    int rem;
2205    int min_order = slub_min_order;
2206
2207    if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2208        return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2209
2210    for (order = max(min_order,
2211                fls(min_objects * size - 1) - PAGE_SHIFT);
2212            order <= max_order; order++) {
2213
2214        unsigned long slab_size = PAGE_SIZE << order;
2215
2216        if (slab_size < min_objects * size + reserved)
2217            continue;
2218
2219        rem = (slab_size - reserved) % size;
2220
2221        if (rem <= slab_size / fract_leftover)
2222            break;
2223
2224    }
2225
2226    return order;
2227}
2228
2229static inline int calculate_order(int size, int reserved)
2230{
2231    int order;
2232    int min_objects;
2233    int fraction;
2234    int max_objects;
2235
2236    /*
2237     * Attempt to find best configuration for a slab. This
2238     * works by first attempting to generate a layout with
2239     * the best configuration and backing off gradually.
2240     *
2241     * First we reduce the acceptable waste in a slab. Then
2242     * we reduce the minimum objects required in a slab.
2243     */
2244    min_objects = slub_min_objects;
2245    if (!min_objects)
2246        min_objects = 4 * (fls(nr_cpu_ids) + 1);
2247    max_objects = order_objects(slub_max_order, size, reserved);
2248    min_objects = min(min_objects, max_objects);
2249
2250    while (min_objects > 1) {
2251        fraction = 16;
2252        while (fraction >= 4) {
2253            order = slab_order(size, min_objects,
2254                    slub_max_order, fraction, reserved);
2255            if (order <= slub_max_order)
2256                return order;
2257            fraction /= 2;
2258        }
2259        min_objects--;
2260    }
2261
2262    /*
2263     * We were unable to place multiple objects in a slab. Now
2264     * lets see if we can place a single object there.
2265     */
2266    order = slab_order(size, 1, slub_max_order, 1, reserved);
2267    if (order <= slub_max_order)
2268        return order;
2269
2270    /*
2271     * Doh this slab cannot be placed using slub_max_order.
2272     */
2273    order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2274    if (order < MAX_ORDER)
2275        return order;
2276    return -ENOSYS;
2277}
2278
2279/*
2280 * Figure out what the alignment of the objects will be.
2281 */
2282static unsigned long calculate_alignment(unsigned long flags,
2283        unsigned long align, unsigned long size)
2284{
2285    /*
2286     * If the user wants hardware cache aligned objects then follow that
2287     * suggestion if the object is sufficiently large.
2288     *
2289     * The hardware cache alignment cannot override the specified
2290     * alignment though. If that is greater then use it.
2291     */
2292    if (flags & SLAB_HWCACHE_ALIGN) {
2293        unsigned long ralign = cache_line_size();
2294        while (size <= ralign / 2)
2295            ralign /= 2;
2296        align = max(align, ralign);
2297    }
2298
2299    if (align < ARCH_SLAB_MINALIGN)
2300        align = ARCH_SLAB_MINALIGN;
2301
2302    return ALIGN(align, sizeof(void *));
2303}
2304
2305static void
2306init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
2307{
2308    n->nr_partial = 0;
2309    spin_lock_init(&n->list_lock);
2310    INIT_LIST_HEAD(&n->partial);
2311#ifdef CONFIG_SLUB_DEBUG
2312    atomic_long_set(&n->nr_slabs, 0);
2313    atomic_long_set(&n->total_objects, 0);
2314    INIT_LIST_HEAD(&n->full);
2315#endif
2316}
2317
2318static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2319{
2320    BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2321            SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2322
2323    /*
2324     * Must align to double word boundary for the double cmpxchg
2325     * instructions to work; see __pcpu_double_call_return_bool().
2326     */
2327    s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2328                     2 * sizeof(void *));
2329
2330    if (!s->cpu_slab)
2331        return 0;
2332
2333    init_kmem_cache_cpus(s);
2334
2335    return 1;
2336}
2337
2338static struct kmem_cache *kmem_cache_node;
2339
2340/*
2341 * No kmalloc_node yet so do it by hand. We know that this is the first
2342 * slab on the node for this slabcache. There are no concurrent accesses
2343 * possible.
2344 *
2345 * Note that this function only works on the kmalloc_node_cache
2346 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2347 * memory on a fresh node that has no slab structures yet.
2348 */
2349static void early_kmem_cache_node_alloc(int node)
2350{
2351    struct page *page;
2352    struct kmem_cache_node *n;
2353    unsigned long flags;
2354
2355    BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2356
2357    page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2358
2359    BUG_ON(!page);
2360    if (page_to_nid(page) != node) {
2361        printk(KERN_ERR "SLUB: Unable to allocate memory from "
2362                "node %d\n", node);
2363        printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2364                "in order to be able to continue\n");
2365    }
2366
2367    n = page->freelist;
2368    BUG_ON(!n);
2369    page->freelist = get_freepointer(kmem_cache_node, n);
2370    page->inuse++;
2371    kmem_cache_node->node[node] = n;
2372#ifdef CONFIG_SLUB_DEBUG
2373    init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2374    init_tracking(kmem_cache_node, n);
2375#endif
2376    init_kmem_cache_node(n, kmem_cache_node);
2377    inc_slabs_node(kmem_cache_node, node, page->objects);
2378
2379    /*
2380     * lockdep requires consistent irq usage for each lock
2381     * so even though there cannot be a race this early in
2382     * the boot sequence, we still disable irqs.
2383     */
2384    local_irq_save(flags);
2385    add_partial(n, page, 0);
2386    local_irq_restore(flags);
2387}
2388
2389static void free_kmem_cache_nodes(struct kmem_cache *s)
2390{
2391    int node;
2392
2393    for_each_node_state(node, N_NORMAL_MEMORY) {
2394        struct kmem_cache_node *n = s->node[node];
2395
2396        if (n)
2397            kmem_cache_free(kmem_cache_node, n);
2398
2399        s->node[node] = NULL;
2400    }
2401}
2402
2403static int init_kmem_cache_nodes(struct kmem_cache *s)
2404{
2405    int node;
2406
2407    for_each_node_state(node, N_NORMAL_MEMORY) {
2408        struct kmem_cache_node *n;
2409
2410        if (slab_state == DOWN) {
2411            early_kmem_cache_node_alloc(node);
2412            continue;
2413        }
2414        n = kmem_cache_alloc_node(kmem_cache_node,
2415                        GFP_KERNEL, node);
2416
2417        if (!n) {
2418            free_kmem_cache_nodes(s);
2419            return 0;
2420        }
2421
2422        s->node[node] = n;
2423        init_kmem_cache_node(n, s);
2424    }
2425    return 1;
2426}
2427
2428static void set_min_partial(struct kmem_cache *s, unsigned long min)
2429{
2430    if (min < MIN_PARTIAL)
2431        min = MIN_PARTIAL;
2432    else if (min > MAX_PARTIAL)
2433        min = MAX_PARTIAL;
2434    s->min_partial = min;
2435}
2436
2437/*
2438 * calculate_sizes() determines the order and the distribution of data within
2439 * a slab object.
2440 */
2441static int calculate_sizes(struct kmem_cache *s, int forced_order)
2442{
2443    unsigned long flags = s->flags;
2444    unsigned long size = s->objsize;
2445    unsigned long align = s->align;
2446    int order;
2447
2448    /*
2449     * Round up object size to the next word boundary. We can only
2450     * place the free pointer at word boundaries and this determines
2451     * the possible location of the free pointer.
2452     */
2453    size = ALIGN(size, sizeof(void *));
2454
2455#ifdef CONFIG_SLUB_DEBUG
2456    /*
2457     * Determine if we can poison the object itself. If the user of
2458     * the slab may touch the object after free or before allocation
2459     * then we should never poison the object itself.
2460     */
2461    if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2462            !s->ctor)
2463        s->flags |= __OBJECT_POISON;
2464    else
2465        s->flags &= ~__OBJECT_POISON;
2466
2467
2468    /*
2469     * If we are Redzoning then check if there is some space between the
2470     * end of the object and the free pointer. If not then add an
2471     * additional word to have some bytes to store Redzone information.
2472     */
2473    if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2474        size += sizeof(void *);
2475#endif
2476
2477    /*
2478     * With that we have determined the number of bytes in actual use
2479     * by the object. This is the potential offset to the free pointer.
2480     */
2481    s->inuse = size;
2482
2483    if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2484        s->ctor)) {
2485        /*
2486         * Relocate free pointer after the object if it is not
2487         * permitted to overwrite the first word of the object on
2488         * kmem_cache_free.
2489         *
2490         * This is the case if we do RCU, have a constructor or
2491         * destructor or are poisoning the objects.
2492         */
2493        s->offset = size;
2494        size += sizeof(void *);
2495    }
2496
2497#ifdef CONFIG_SLUB_DEBUG
2498    if (flags & SLAB_STORE_USER)
2499        /*
2500         * Need to store information about allocs and frees after
2501         * the object.
2502         */
2503        size += 2 * sizeof(struct track);
2504
2505    if (flags & SLAB_RED_ZONE)
2506        /*
2507         * Add some empty padding so that we can catch
2508         * overwrites from earlier objects rather than let
2509         * tracking information or the free pointer be
2510         * corrupted if a user writes before the start
2511         * of the object.
2512         */
2513        size += sizeof(void *);
2514#endif
2515
2516    /*
2517     * Determine the alignment based on various parameters that the
2518     * user specified and the dynamic determination of cache line size
2519     * on bootup.
2520     */
2521    align = calculate_alignment(flags, align, s->objsize);
2522    s->align = align;
2523
2524    /*
2525     * SLUB stores one object immediately after another beginning from
2526     * offset 0. In order to align the objects we have to simply size
2527     * each object to conform to the alignment.
2528     */
2529    size = ALIGN(size, align);
2530    s->size = size;
2531    if (forced_order >= 0)
2532        order = forced_order;
2533    else
2534        order = calculate_order(size, s->reserved);
2535
2536    if (order < 0)
2537        return 0;
2538
2539    s->allocflags = 0;
2540    if (order)
2541        s->allocflags |= __GFP_COMP;
2542
2543    if (s->flags & SLAB_CACHE_DMA)
2544        s->allocflags |= SLUB_DMA;
2545
2546    if (s->flags & SLAB_RECLAIM_ACCOUNT)
2547        s->allocflags |= __GFP_RECLAIMABLE;
2548
2549    /*
2550     * Determine the number of objects per slab
2551     */
2552    s->oo = oo_make(order, size, s->reserved);
2553    s->min = oo_make(get_order(size), size, s->reserved);
2554    if (oo_objects(s->oo) > oo_objects(s->max))
2555        s->max = s->oo;
2556
2557    return !!oo_objects(s->oo);
2558
2559}
2560
2561static int kmem_cache_open(struct kmem_cache *s,
2562        const char *name, size_t size,
2563        size_t align, unsigned long flags,
2564        void (*ctor)(void *))
2565{
2566    memset(s, 0, kmem_size);
2567    s->name = name;
2568    s->ctor = ctor;
2569    s->objsize = size;
2570    s->align = align;
2571    s->flags = kmem_cache_flags(size, flags, name, ctor);
2572    s->reserved = 0;
2573
2574    if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
2575        s->reserved = sizeof(struct rcu_head);
2576
2577    if (!calculate_sizes(s, -1))
2578        goto error;
2579    if (disable_higher_order_debug) {
2580        /*
2581         * Disable debugging flags that store metadata if the min slab
2582         * order increased.
2583         */
2584        if (get_order(s->size) > get_order(s->objsize)) {
2585            s->flags &= ~DEBUG_METADATA_FLAGS;
2586            s->offset = 0;
2587            if (!calculate_sizes(s, -1))
2588                goto error;
2589        }
2590    }
2591
2592    /*
2593     * The larger the object size is, the more pages we want on the partial
2594     * list to avoid pounding the page allocator excessively.
2595     */
2596    set_min_partial(s, ilog2(s->size));
2597    s->refcount = 1;
2598#ifdef CONFIG_NUMA
2599    s->remote_node_defrag_ratio = 1000;
2600#endif
2601    if (!init_kmem_cache_nodes(s))
2602        goto error;
2603
2604    if (alloc_kmem_cache_cpus(s))
2605        return 1;
2606
2607    free_kmem_cache_nodes(s);
2608error:
2609    if (flags & SLAB_PANIC)
2610        panic("Cannot create slab %s size=%lu realsize=%u "
2611            "order=%u offset=%u flags=%lx\n",
2612            s->name, (unsigned long)size, s->size, oo_order(s->oo),
2613            s->offset, flags);
2614    return 0;
2615}
2616
2617/*
2618 * Determine the size of a slab object
2619 */
2620unsigned int kmem_cache_size(struct kmem_cache *s)
2621{
2622    return s->objsize;
2623}
2624EXPORT_SYMBOL(kmem_cache_size);
2625
2626static void list_slab_objects(struct kmem_cache *s, struct page *page,
2627                            const char *text)
2628{
2629#ifdef CONFIG_SLUB_DEBUG
2630    void *addr = page_address(page);
2631    void *p;
2632    unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
2633                     sizeof(long), GFP_ATOMIC);
2634    if (!map)
2635        return;
2636    slab_err(s, page, "%s", text);
2637    slab_lock(page);
2638
2639    get_map(s, page, map);
2640    for_each_object(p, s, addr, page->objects) {
2641
2642        if (!test_bit(slab_index(p, s, addr), map)) {
2643            printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2644                            p, p - addr);
2645            print_tracking(s, p);
2646        }
2647    }
2648    slab_unlock(page);
2649    kfree(map);
2650#endif
2651}
2652
2653/*
2654 * Attempt to free all partial slabs on a node.
2655 */
2656static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2657{
2658    unsigned long flags;
2659    struct page *page, *h;
2660
2661    spin_lock_irqsave(&n->list_lock, flags);
2662    list_for_each_entry_safe(page, h, &n->partial, lru) {
2663        if (!page->inuse) {
2664            __remove_partial(n, page);
2665            discard_slab(s, page);
2666        } else {
2667            list_slab_objects(s, page,
2668                "Objects remaining on kmem_cache_close()");
2669        }
2670    }
2671    spin_unlock_irqrestore(&n->list_lock, flags);
2672}
2673
2674/*
2675 * Release all resources used by a slab cache.
2676 */
2677static inline int kmem_cache_close(struct kmem_cache *s)
2678{
2679    int node;
2680
2681    flush_all(s);
2682    free_percpu(s->cpu_slab);
2683    /* Attempt to free all objects */
2684    for_each_node_state(node, N_NORMAL_MEMORY) {
2685        struct kmem_cache_node *n = get_node(s, node);
2686
2687        free_partial(s, n);
2688        if (n->nr_partial || slabs_node(s, node))
2689            return 1;
2690    }
2691    free_kmem_cache_nodes(s);
2692    return 0;
2693}
2694
2695/*
2696 * Close a cache and release the kmem_cache structure
2697 * (must be used for caches created using kmem_cache_create)
2698 */
2699void kmem_cache_destroy(struct kmem_cache *s)
2700{
2701    down_write(&slub_lock);
2702    s->refcount--;
2703    if (!s->refcount) {
2704        list_del(&s->list);
2705        if (kmem_cache_close(s)) {
2706            printk(KERN_ERR "SLUB %s: %s called for cache that "
2707                "still has objects.\n", s->name, __func__);
2708            dump_stack();
2709        }
2710        if (s->flags & SLAB_DESTROY_BY_RCU)
2711            rcu_barrier();
2712        sysfs_slab_remove(s);
2713    }
2714    up_write(&slub_lock);
2715}
2716EXPORT_SYMBOL(kmem_cache_destroy);
2717
2718/********************************************************************
2719 * Kmalloc subsystem
2720 *******************************************************************/
2721
2722struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
2723EXPORT_SYMBOL(kmalloc_caches);
2724
2725static struct kmem_cache *kmem_cache;
2726
2727#ifdef CONFIG_ZONE_DMA
2728static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
2729#endif
2730
2731static int __init setup_slub_min_order(char *str)
2732{
2733    get_option(&str, &slub_min_order);
2734
2735    return 1;
2736}
2737
2738__setup("slub_min_order=", setup_slub_min_order);
2739
2740static int __init setup_slub_max_order(char *str)
2741{
2742    get_option(&str, &slub_max_order);
2743    slub_max_order = min(slub_max_order, MAX_ORDER - 1);
2744
2745    return 1;
2746}
2747
2748__setup("slub_max_order=", setup_slub_max_order);
2749
2750static int __init setup_slub_min_objects(char *str)
2751{
2752    get_option(&str, &slub_min_objects);
2753
2754    return 1;
2755}
2756
2757__setup("slub_min_objects=", setup_slub_min_objects);
2758
2759static int __init setup_slub_nomerge(char *str)
2760{
2761    slub_nomerge = 1;
2762    return 1;
2763}
2764
2765__setup("slub_nomerge", setup_slub_nomerge);
2766
2767static struct kmem_cache *__init create_kmalloc_cache(const char *name,
2768                        int size, unsigned int flags)
2769{
2770    struct kmem_cache *s;
2771
2772    s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
2773
2774    /*
2775     * This function is called with IRQs disabled during early-boot on
2776     * single CPU so there's no need to take slub_lock here.
2777     */
2778    if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN,
2779                                flags, NULL))
2780        goto panic;
2781
2782    list_add(&s->list, &slab_caches);
2783    return s;
2784
2785panic:
2786    panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2787    return NULL;
2788}
2789
2790/*
2791 * Conversion table for small slabs sizes / 8 to the index in the
2792 * kmalloc array. This is necessary for slabs < 192 since we have non power
2793 * of two cache sizes there. The size of larger slabs can be determined using
2794 * fls.
2795 */
2796static s8 size_index[24] = {
2797    3, /* 8 */
2798    4, /* 16 */
2799    5, /* 24 */
2800    5, /* 32 */
2801    6, /* 40 */
2802    6, /* 48 */
2803    6, /* 56 */
2804    6, /* 64 */
2805    1, /* 72 */
2806    1, /* 80 */
2807    1, /* 88 */
2808    1, /* 96 */
2809    7, /* 104 */
2810    7, /* 112 */
2811    7, /* 120 */
2812    7, /* 128 */
2813    2, /* 136 */
2814    2, /* 144 */
2815    2, /* 152 */
2816    2, /* 160 */
2817    2, /* 168 */
2818    2, /* 176 */
2819    2, /* 184 */
2820    2 /* 192 */
2821};
2822
2823static inline int size_index_elem(size_t bytes)
2824{
2825    return (bytes - 1) / 8;
2826}
2827
2828static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2829{
2830    int index;
2831
2832    if (size <= 192) {
2833        if (!size)
2834            return ZERO_SIZE_PTR;
2835
2836        index = size_index[size_index_elem(size)];
2837    } else
2838        index = fls(size - 1);
2839
2840#ifdef CONFIG_ZONE_DMA
2841    if (unlikely((flags & SLUB_DMA)))
2842        return kmalloc_dma_caches[index];
2843
2844#endif
2845    return kmalloc_caches[index];
2846}
2847
2848void *__kmalloc(size_t size, gfp_t flags)
2849{
2850    struct kmem_cache *s;
2851    void *ret;
2852
2853    if (unlikely(size > SLUB_MAX_SIZE))
2854        return kmalloc_large(size, flags);
2855
2856    s = get_slab(size, flags);
2857
2858    if (unlikely(ZERO_OR_NULL_PTR(s)))
2859        return s;
2860
2861    ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_);
2862
2863    trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
2864
2865    return ret;
2866}
2867EXPORT_SYMBOL(__kmalloc);
2868
2869#ifdef CONFIG_NUMA
2870static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2871{
2872    struct page *page;
2873    void *ptr = NULL;
2874
2875    flags |= __GFP_COMP | __GFP_NOTRACK;
2876    page = alloc_pages_node(node, flags, get_order(size));
2877    if (page)
2878        ptr = page_address(page);
2879
2880    kmemleak_alloc(ptr, size, 1, flags);
2881    return ptr;
2882}
2883
2884void *__kmalloc_node(size_t size, gfp_t flags, int node)
2885{
2886    struct kmem_cache *s;
2887    void *ret;
2888
2889    if (unlikely(size > SLUB_MAX_SIZE)) {
2890        ret = kmalloc_large_node(size, flags, node);
2891
2892        trace_kmalloc_node(_RET_IP_, ret,
2893                   size, PAGE_SIZE << get_order(size),
2894                   flags, node);
2895
2896        return ret;
2897    }
2898
2899    s = get_slab(size, flags);
2900
2901    if (unlikely(ZERO_OR_NULL_PTR(s)))
2902        return s;
2903
2904    ret = slab_alloc(s, flags, node, _RET_IP_);
2905
2906    trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
2907
2908    return ret;
2909}
2910EXPORT_SYMBOL(__kmalloc_node);
2911#endif
2912
2913size_t ksize(const void *object)
2914{
2915    struct page *page;
2916
2917    if (unlikely(object == ZERO_SIZE_PTR))
2918        return 0;
2919
2920    page = virt_to_head_page(object);
2921
2922    if (unlikely(!PageSlab(page))) {
2923        WARN_ON(!PageCompound(page));
2924        return PAGE_SIZE << compound_order(page);
2925    }
2926
2927    return slab_ksize(page->slab);
2928}
2929EXPORT_SYMBOL(ksize);
2930
2931void kfree(const void *x)
2932{
2933    struct page *page;
2934    void *object = (void *)x;
2935
2936    trace_kfree(_RET_IP_, x);
2937
2938    if (unlikely(ZERO_OR_NULL_PTR(x)))
2939        return;
2940
2941    page = virt_to_head_page(x);
2942    if (unlikely(!PageSlab(page))) {
2943        BUG_ON(!PageCompound(page));
2944        kmemleak_free(x);
2945        put_page(page);
2946        return;
2947    }
2948    slab_free(page->slab, page, object, _RET_IP_);
2949}
2950EXPORT_SYMBOL(kfree);
2951
2952/*
2953 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2954 * the remaining slabs by the number of items in use. The slabs with the
2955 * most items in use come first. New allocations will then fill those up
2956 * and thus they can be removed from the partial lists.
2957 *
2958 * The slabs with the least items are placed last. This results in them
2959 * being allocated from last increasing the chance that the last objects
2960 * are freed in them.
2961 */
2962int kmem_cache_shrink(struct kmem_cache *s)
2963{
2964    int node;
2965    int i;
2966    struct kmem_cache_node *n;
2967    struct page *page;
2968    struct page *t;
2969    int objects = oo_objects(s->max);
2970    struct list_head *slabs_by_inuse =
2971        kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
2972    unsigned long flags;
2973
2974    if (!slabs_by_inuse)
2975        return -ENOMEM;
2976
2977    flush_all(s);
2978    for_each_node_state(node, N_NORMAL_MEMORY) {
2979        n = get_node(s, node);
2980
2981        if (!n->nr_partial)
2982            continue;
2983
2984        for (i = 0; i < objects; i++)
2985            INIT_LIST_HEAD(slabs_by_inuse + i);
2986
2987        spin_lock_irqsave(&n->list_lock, flags);
2988
2989        /*
2990         * Build lists indexed by the items in use in each slab.
2991         *
2992         * Note that concurrent frees may occur while we hold the
2993         * list_lock. page->inuse here is the upper limit.
2994         */
2995        list_for_each_entry_safe(page, t, &n->partial, lru) {
2996            if (!page->inuse && slab_trylock(page)) {
2997                /*
2998                 * Must hold slab lock here because slab_free
2999                 * may have freed the last object and be
3000                 * waiting to release the slab.
3001                 */
3002                __remove_partial(n, page);
3003                slab_unlock(page);
3004                discard_slab(s, page);
3005            } else {
3006                list_move(&page->lru,
3007                slabs_by_inuse + page->inuse);
3008            }
3009        }
3010
3011        /*
3012         * Rebuild the partial list with the slabs filled up most
3013         * first and the least used slabs at the end.
3014         */
3015        for (i = objects - 1; i >= 0; i--)
3016            list_splice(slabs_by_inuse + i, n->partial.prev);
3017
3018        spin_unlock_irqrestore(&n->list_lock, flags);
3019    }
3020
3021    kfree(slabs_by_inuse);
3022    return 0;
3023}
3024EXPORT_SYMBOL(kmem_cache_shrink);
3025
3026#if defined(CONFIG_MEMORY_HOTPLUG)
3027static int slab_mem_going_offline_callback(void *arg)
3028{
3029    struct kmem_cache *s;
3030
3031    down_read(&slub_lock);
3032    list_for_each_entry(s, &slab_caches, list)
3033        kmem_cache_shrink(s);
3034    up_read(&slub_lock);
3035
3036    return 0;
3037}
3038
3039static void slab_mem_offline_callback(void *arg)
3040{
3041    struct kmem_cache_node *n;
3042    struct kmem_cache *s;
3043    struct memory_notify *marg = arg;
3044    int offline_node;
3045
3046    offline_node = marg->status_change_nid;
3047
3048    /*
3049     * If the node still has available memory. we need kmem_cache_node
3050     * for it yet.
3051     */
3052    if (offline_node < 0)
3053        return;
3054
3055    down_read(&slub_lock);
3056    list_for_each_entry(s, &slab_caches, list) {
3057        n = get_node(s, offline_node);
3058        if (n) {
3059            /*
3060             * if n->nr_slabs > 0, slabs still exist on the node
3061             * that is going down. We were unable to free them,
3062             * and offline_pages() function shouldn't call this
3063             * callback. So, we must fail.
3064             */
3065            BUG_ON(slabs_node(s, offline_node));
3066
3067            s->node[offline_node] = NULL;
3068            kmem_cache_free(kmem_cache_node, n);
3069        }
3070    }
3071    up_read(&slub_lock);
3072}
3073
3074static int slab_mem_going_online_callback(void *arg)
3075{
3076    struct kmem_cache_node *n;
3077    struct kmem_cache *s;
3078    struct memory_notify *marg = arg;
3079    int nid = marg->status_change_nid;
3080    int ret = 0;
3081
3082    /*
3083     * If the node's memory is already available, then kmem_cache_node is
3084     * already created. Nothing to do.
3085     */
3086    if (nid < 0)
3087        return 0;
3088
3089    /*
3090     * We are bringing a node online. No memory is available yet. We must
3091     * allocate a kmem_cache_node structure in order to bring the node
3092     * online.
3093     */
3094    down_read(&slub_lock);
3095    list_for_each_entry(s, &slab_caches, list) {
3096        /*
3097         * XXX: kmem_cache_alloc_node will fallback to other nodes
3098         * since memory is not yet available from the node that
3099         * is brought up.
3100         */
3101        n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3102        if (!n) {
3103            ret = -ENOMEM;
3104            goto out;
3105        }
3106        init_kmem_cache_node(n, s);
3107        s->node[nid] = n;
3108    }
3109out:
3110    up_read(&slub_lock);
3111    return ret;
3112}
3113
3114static int slab_memory_callback(struct notifier_block *self,
3115                unsigned long action, void *arg)
3116{
3117    int ret = 0;
3118
3119    switch (action) {
3120    case MEM_GOING_ONLINE:
3121        ret = slab_mem_going_online_callback(arg);
3122        break;
3123    case MEM_GOING_OFFLINE:
3124        ret = slab_mem_going_offline_callback(arg);
3125        break;
3126    case MEM_OFFLINE:
3127    case MEM_CANCEL_ONLINE:
3128        slab_mem_offline_callback(arg);
3129        break;
3130    case MEM_ONLINE:
3131    case MEM_CANCEL_OFFLINE:
3132        break;
3133    }
3134    if (ret)
3135        ret = notifier_from_errno(ret);
3136    else
3137        ret = NOTIFY_OK;
3138    return ret;
3139}
3140
3141#endif /* CONFIG_MEMORY_HOTPLUG */
3142
3143/********************************************************************
3144 * Basic setup of slabs
3145 *******************************************************************/
3146
3147/*
3148 * Used for early kmem_cache structures that were allocated using
3149 * the page allocator
3150 */
3151
3152static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s)
3153{
3154    int node;
3155
3156    list_add(&s->list, &slab_caches);
3157    s->refcount = -1;
3158
3159    for_each_node_state(node, N_NORMAL_MEMORY) {
3160        struct kmem_cache_node *n = get_node(s, node);
3161        struct page *p;
3162
3163        if (n) {
3164            list_for_each_entry(p, &n->partial, lru)
3165                p->slab = s;
3166
3167#ifdef CONFIG_SLUB_DEBUG
3168            list_for_each_entry(p, &n->full, lru)
3169                p->slab = s;
3170#endif
3171        }
3172    }
3173}
3174
3175void __init kmem_cache_init(void)
3176{
3177    int i;
3178    int caches = 0;
3179    struct kmem_cache *temp_kmem_cache;
3180    int order;
3181    struct kmem_cache *temp_kmem_cache_node;
3182    unsigned long kmalloc_size;
3183
3184    kmem_size = offsetof(struct kmem_cache, node) +
3185                nr_node_ids * sizeof(struct kmem_cache_node *);
3186
3187    /* Allocate two kmem_caches from the page allocator */
3188    kmalloc_size = ALIGN(kmem_size, cache_line_size());
3189    order = get_order(2 * kmalloc_size);
3190    kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order);
3191
3192    /*
3193     * Must first have the slab cache available for the allocations of the
3194     * struct kmem_cache_node's. There is special bootstrap code in
3195     * kmem_cache_open for slab_state == DOWN.
3196     */
3197    kmem_cache_node = (void *)kmem_cache + kmalloc_size;
3198
3199    kmem_cache_open(kmem_cache_node, "kmem_cache_node",
3200        sizeof(struct kmem_cache_node),
3201        0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3202
3203    hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3204
3205    /* Able to allocate the per node structures */
3206    slab_state = PARTIAL;
3207
3208    temp_kmem_cache = kmem_cache;
3209    kmem_cache_open(kmem_cache, "kmem_cache", kmem_size,
3210        0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3211    kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3212    memcpy(kmem_cache, temp_kmem_cache, kmem_size);
3213
3214    /*
3215     * Allocate kmem_cache_node properly from the kmem_cache slab.
3216     * kmem_cache_node is separately allocated so no need to
3217     * update any list pointers.
3218     */
3219    temp_kmem_cache_node = kmem_cache_node;
3220
3221    kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3222    memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size);
3223
3224    kmem_cache_bootstrap_fixup(kmem_cache_node);
3225
3226    caches++;
3227    kmem_cache_bootstrap_fixup(kmem_cache);
3228    caches++;
3229    /* Free temporary boot structure */
3230    free_pages((unsigned long)temp_kmem_cache, order);
3231
3232    /* Now we can use the kmem_cache to allocate kmalloc slabs */
3233
3234    /*
3235     * Patch up the size_index table if we have strange large alignment
3236     * requirements for the kmalloc array. This is only the case for
3237     * MIPS it seems. The standard arches will not generate any code here.
3238     *
3239     * Largest permitted alignment is 256 bytes due to the way we
3240     * handle the index determination for the smaller caches.
3241     *
3242     * Make sure that nothing crazy happens if someone starts tinkering
3243     * around with ARCH_KMALLOC_MINALIGN
3244     */
3245    BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3246        (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3247
3248    for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3249        int elem = size_index_elem(i);
3250        if (elem >= ARRAY_SIZE(size_index))
3251            break;
3252        size_index[elem] = KMALLOC_SHIFT_LOW;
3253    }
3254
3255    if (KMALLOC_MIN_SIZE == 64) {
3256        /*
3257         * The 96 byte size cache is not used if the alignment
3258         * is 64 byte.
3259         */
3260        for (i = 64 + 8; i <= 96; i += 8)
3261            size_index[size_index_elem(i)] = 7;
3262    } else if (KMALLOC_MIN_SIZE == 128) {
3263        /*
3264         * The 192 byte sized cache is not used if the alignment
3265         * is 128 byte. Redirect kmalloc to use the 256 byte cache
3266         * instead.
3267         */
3268        for (i = 128 + 8; i <= 192; i += 8)
3269            size_index[size_index_elem(i)] = 8;
3270    }
3271
3272    /* Caches that are not of the two-to-the-power-of size */
3273    if (KMALLOC_MIN_SIZE <= 32) {
3274        kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3275        caches++;
3276    }
3277
3278    if (KMALLOC_MIN_SIZE <= 64) {
3279        kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3280        caches++;
3281    }
3282
3283    for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3284        kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3285        caches++;
3286    }
3287
3288    slab_state = UP;
3289
3290    /* Provide the correct kmalloc names now that the caches are up */
3291    if (KMALLOC_MIN_SIZE <= 32) {
3292        kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3293        BUG_ON(!kmalloc_caches[1]->name);
3294    }
3295
3296    if (KMALLOC_MIN_SIZE <= 64) {
3297        kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3298        BUG_ON(!kmalloc_caches[2]->name);
3299    }
3300
3301    for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3302        char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3303
3304        BUG_ON(!s);
3305        kmalloc_caches[i]->name = s;
3306    }
3307
3308#ifdef CONFIG_SMP
3309    register_cpu_notifier(&slab_notifier);
3310#endif
3311
3312#ifdef CONFIG_ZONE_DMA
3313    for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3314        struct kmem_cache *s = kmalloc_caches[i];
3315
3316        if (s && s->size) {
3317            char *name = kasprintf(GFP_NOWAIT,
3318                 "dma-kmalloc-%d", s->objsize);
3319
3320            BUG_ON(!name);
3321            kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3322                s->objsize, SLAB_CACHE_DMA);
3323        }
3324    }
3325#endif
3326    printk(KERN_INFO
3327        "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3328        " CPUs=%d, Nodes=%d\n",
3329        caches, cache_line_size(),
3330        slub_min_order, slub_max_order, slub_min_objects,
3331        nr_cpu_ids, nr_node_ids);
3332}
3333
3334void __init kmem_cache_init_late(void)
3335{
3336}
3337
3338/*
3339 * Find a mergeable slab cache
3340 */
3341static int slab_unmergeable(struct kmem_cache *s)
3342{
3343    if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3344        return 1;
3345
3346    if (s->ctor)
3347        return 1;
3348
3349    /*
3350     * We may have set a slab to be unmergeable during bootstrap.
3351     */
3352    if (s->refcount < 0)
3353        return 1;
3354
3355    return 0;
3356}
3357
3358static struct kmem_cache *find_mergeable(size_t size,
3359        size_t align, unsigned long flags, const char *name,
3360        void (*ctor)(void *))
3361{
3362    struct kmem_cache *s;
3363
3364    if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3365        return NULL;
3366
3367    if (ctor)
3368        return NULL;
3369
3370    size = ALIGN(size, sizeof(void *));
3371    align = calculate_alignment(flags, align, size);
3372    size = ALIGN(size, align);
3373    flags = kmem_cache_flags(size, flags, name, NULL);
3374
3375    list_for_each_entry(s, &slab_caches, list) {
3376        if (slab_unmergeable(s))
3377            continue;
3378
3379        if (size > s->size)
3380            continue;
3381
3382        if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3383                continue;
3384        /*
3385         * Check if alignment is compatible.
3386         * Courtesy of Adrian Drzewiecki
3387         */
3388        if ((s->size & ~(align - 1)) != s->size)
3389            continue;
3390
3391        if (s->size - size >= sizeof(void *))
3392            continue;
3393
3394        return s;
3395    }
3396    return NULL;
3397}
3398
3399struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3400        size_t align, unsigned long flags, void (*ctor)(void *))
3401{
3402    struct kmem_cache *s;
3403    char *n;
3404
3405    if (WARN_ON(!name))
3406        return NULL;
3407
3408    down_write(&slub_lock);
3409    s = find_mergeable(size, align, flags, name, ctor);
3410    if (s) {
3411        s->refcount++;
3412        /*
3413         * Adjust the object sizes so that we clear
3414         * the complete object on kzalloc.
3415         */
3416        s->objsize = max(s->objsize, (int)size);
3417        s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3418
3419        if (sysfs_slab_alias(s, name)) {
3420            s->refcount--;
3421            goto err;
3422        }
3423        up_write(&slub_lock);
3424        return s;
3425    }
3426
3427    n = kstrdup(name, GFP_KERNEL);
3428    if (!n)
3429        goto err;
3430
3431    s = kmalloc(kmem_size, GFP_KERNEL);
3432    if (s) {
3433        if (kmem_cache_open(s, n,
3434                size, align, flags, ctor)) {
3435            list_add(&s->list, &slab_caches);
3436            if (sysfs_slab_add(s)) {
3437                list_del(&s->list);
3438                kfree(n);
3439                kfree(s);
3440                goto err;
3441            }
3442            up_write(&slub_lock);
3443            return s;
3444        }
3445        kfree(n);
3446        kfree(s);
3447    }
3448err:
3449    up_write(&slub_lock);
3450
3451    if (flags & SLAB_PANIC)
3452        panic("Cannot create slabcache %s\n", name);
3453    else
3454        s = NULL;
3455    return s;
3456}
3457EXPORT_SYMBOL(kmem_cache_create);
3458
3459#ifdef CONFIG_SMP
3460/*
3461 * Use the cpu notifier to insure that the cpu slabs are flushed when
3462 * necessary.
3463 */
3464static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3465        unsigned long action, void *hcpu)
3466{
3467    long cpu = (long)hcpu;
3468    struct kmem_cache *s;
3469    unsigned long flags;
3470
3471    switch (action) {
3472    case CPU_UP_CANCELED:
3473    case CPU_UP_CANCELED_FROZEN:
3474    case CPU_DEAD:
3475    case CPU_DEAD_FROZEN:
3476        down_read(&slub_lock);
3477        list_for_each_entry(s, &slab_caches, list) {
3478            local_irq_save(flags);
3479            __flush_cpu_slab(s, cpu);
3480            local_irq_restore(flags);
3481        }
3482        up_read(&slub_lock);
3483        break;
3484    default:
3485        break;
3486    }
3487    return NOTIFY_OK;
3488}
3489
3490static struct notifier_block __cpuinitdata slab_notifier = {
3491    .notifier_call = slab_cpuup_callback
3492};
3493
3494#endif
3495
3496void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3497{
3498    struct kmem_cache *s;
3499    void *ret;
3500
3501    if (unlikely(size > SLUB_MAX_SIZE))
3502        return kmalloc_large(size, gfpflags);
3503
3504    s = get_slab(size, gfpflags);
3505
3506    if (unlikely(ZERO_OR_NULL_PTR(s)))
3507        return s;
3508
3509    ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller);
3510
3511    /* Honor the call site pointer we received. */
3512    trace_kmalloc(caller, ret, size, s->size, gfpflags);
3513
3514    return ret;
3515}
3516
3517#ifdef CONFIG_NUMA
3518void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3519                    int node, unsigned long caller)
3520{
3521    struct kmem_cache *s;
3522    void *ret;
3523
3524    if (unlikely(size > SLUB_MAX_SIZE)) {
3525        ret = kmalloc_large_node(size, gfpflags, node);
3526
3527        trace_kmalloc_node(caller, ret,
3528                   size, PAGE_SIZE << get_order(size),
3529                   gfpflags, node);
3530
3531        return ret;
3532    }
3533
3534    s = get_slab(size, gfpflags);
3535
3536    if (unlikely(ZERO_OR_NULL_PTR(s)))
3537        return s;
3538
3539    ret = slab_alloc(s, gfpflags, node, caller);
3540
3541    /* Honor the call site pointer we received. */
3542    trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3543
3544    return ret;
3545}
3546#endif
3547
3548#ifdef CONFIG_SYSFS
3549static int count_inuse(struct page *page)
3550{
3551    return page->inuse;
3552}
3553
3554static int count_total(struct page *page)
3555{
3556    return page->objects;
3557}
3558#endif
3559
3560#ifdef CONFIG_SLUB_DEBUG
3561static int validate_slab(struct kmem_cache *s, struct page *page,
3562                        unsigned long *map)
3563{
3564    void *p;
3565    void *addr = page_address(page);
3566
3567    if (!check_slab(s, page) ||
3568            !on_freelist(s, page, NULL))
3569        return 0;
3570
3571    /* Now we know that a valid freelist exists */
3572    bitmap_zero(map, page->objects);
3573
3574    get_map(s, page, map);
3575    for_each_object(p, s, addr, page->objects) {
3576        if (test_bit(slab_index(p, s, addr), map))
3577            if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3578                return 0;
3579    }
3580
3581    for_each_object(p, s, addr, page->objects)
3582        if (!test_bit(slab_index(p, s, addr), map))
3583            if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3584                return 0;
3585    return 1;
3586}
3587
3588static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3589                        unsigned long *map)
3590{
3591    if (slab_trylock(page)) {
3592        validate_slab(s, page, map);
3593        slab_unlock(page);
3594    } else
3595        printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3596            s->name, page);
3597}
3598
3599static int validate_slab_node(struct kmem_cache *s,
3600        struct kmem_cache_node *n, unsigned long *map)
3601{
3602    unsigned long count = 0;
3603    struct page *page;
3604    unsigned long flags;
3605
3606    spin_lock_irqsave(&n->list_lock, flags);
3607
3608    list_for_each_entry(page, &n->partial, lru) {
3609        validate_slab_slab(s, page, map);
3610        count++;
3611    }
3612    if (count != n->nr_partial)
3613        printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3614            "counter=%ld\n", s->name, count, n->nr_partial);
3615
3616    if (!(s->flags & SLAB_STORE_USER))
3617        goto out;
3618
3619    list_for_each_entry(page, &n->full, lru) {
3620        validate_slab_slab(s, page, map);
3621        count++;
3622    }
3623    if (count != atomic_long_read(&n->nr_slabs))
3624        printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3625            "counter=%ld\n", s->name, count,
3626            atomic_long_read(&n->nr_slabs));
3627
3628out:
3629    spin_unlock_irqrestore(&n->list_lock, flags);
3630    return count;
3631}
3632
3633static long validate_slab_cache(struct kmem_cache *s)
3634{
3635    int node;
3636    unsigned long count = 0;
3637    unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3638                sizeof(unsigned long), GFP_KERNEL);
3639
3640    if (!map)
3641        return -ENOMEM;
3642
3643    flush_all(s);
3644    for_each_node_state(node, N_NORMAL_MEMORY) {
3645        struct kmem_cache_node *n = get_node(s, node);
3646
3647        count += validate_slab_node(s, n, map);
3648    }
3649    kfree(map);
3650    return count;
3651}
3652/*
3653 * Generate lists of code addresses where slabcache objects are allocated
3654 * and freed.
3655 */
3656
3657struct location {
3658    unsigned long count;
3659    unsigned long addr;
3660    long long sum_time;
3661    long min_time;
3662    long max_time;
3663    long min_pid;
3664    long max_pid;
3665    DECLARE_BITMAP(cpus, NR_CPUS);
3666    nodemask_t nodes;
3667};
3668
3669struct loc_track {
3670    unsigned long max;
3671    unsigned long count;
3672    struct location *loc;
3673};
3674
3675static void free_loc_track(struct loc_track *t)
3676{
3677    if (t->max)
3678        free_pages((unsigned long)t->loc,
3679            get_order(sizeof(struct location) * t->max));
3680}
3681
3682static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3683{
3684    struct location *l;
3685    int order;
3686
3687    order = get_order(sizeof(struct location) * max);
3688
3689    l = (void *)__get_free_pages(flags, order);
3690    if (!l)
3691        return 0;
3692
3693    if (t->count) {
3694        memcpy(l, t->loc, sizeof(struct location) * t->count);
3695        free_loc_track(t);
3696    }
3697    t->max = max;
3698    t->loc = l;
3699    return 1;
3700}
3701
3702static int add_location(struct loc_track *t, struct kmem_cache *s,
3703                const struct track *track)
3704{
3705    long start, end, pos;
3706    struct location *l;
3707    unsigned long caddr;
3708    unsigned long age = jiffies - track->when;
3709
3710    start = -1;
3711    end = t->count;
3712
3713    for ( ; ; ) {
3714        pos = start + (end - start + 1) / 2;
3715
3716        /*
3717         * There is nothing at "end". If we end up there
3718         * we need to add something to before end.
3719         */
3720        if (pos == end)
3721            break;
3722
3723        caddr = t->loc[pos].addr;
3724        if (track->addr == caddr) {
3725
3726            l = &t->loc[pos];
3727            l->count++;
3728            if (track->when) {
3729                l->sum_time += age;
3730                if (age < l->min_time)
3731                    l->min_time = age;
3732                if (age > l->max_time)
3733                    l->max_time = age;
3734
3735                if (track->pid < l->min_pid)
3736                    l->min_pid = track->pid;
3737                if (track->pid > l->max_pid)
3738                    l->max_pid = track->pid;
3739
3740                cpumask_set_cpu(track->cpu,
3741                        to_cpumask(l->cpus));
3742            }
3743            node_set(page_to_nid(virt_to_page(track)), l->nodes);
3744            return 1;
3745        }
3746
3747        if (track->addr < caddr)
3748            end = pos;
3749        else
3750            start = pos;
3751    }
3752
3753    /*
3754     * Not found. Insert new tracking element.
3755     */
3756    if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3757        return 0;
3758
3759    l = t->loc + pos;
3760    if (pos < t->count)
3761        memmove(l + 1, l,
3762            (t->count - pos) * sizeof(struct location));
3763    t->count++;
3764    l->count = 1;
3765    l->addr = track->addr;
3766    l->sum_time = age;
3767    l->min_time = age;
3768    l->max_time = age;
3769    l->min_pid = track->pid;
3770    l->max_pid = track->pid;
3771    cpumask_clear(to_cpumask(l->cpus));
3772    cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
3773    nodes_clear(l->nodes);
3774    node_set(page_to_nid(virt_to_page(track)), l->nodes);
3775    return 1;
3776}
3777
3778static void process_slab(struct loc_track *t, struct kmem_cache *s,
3779        struct page *page, enum track_item alloc,
3780        unsigned long *map)
3781{
3782    void *addr = page_address(page);
3783    void *p;
3784
3785    bitmap_zero(map, page->objects);
3786    get_map(s, page, map);
3787
3788    for_each_object(p, s, addr, page->objects)
3789        if (!test_bit(slab_index(p, s, addr), map))
3790            add_location(t, s, get_track(s, p, alloc));
3791}
3792
3793static int list_locations(struct kmem_cache *s, char *buf,
3794                    enum track_item alloc)
3795{
3796    int len = 0;
3797    unsigned long i;
3798    struct loc_track t = { 0, 0, NULL };
3799    int node;
3800    unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3801                     sizeof(unsigned long), GFP_KERNEL);
3802
3803    if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3804                     GFP_TEMPORARY)) {
3805        kfree(map);
3806        return sprintf(buf, "Out of memory\n");
3807    }
3808    /* Push back cpu slabs */
3809    flush_all(s);
3810
3811    for_each_node_state(node, N_NORMAL_MEMORY) {
3812        struct kmem_cache_node *n = get_node(s, node);
3813        unsigned long flags;
3814        struct page *page;
3815
3816        if (!atomic_long_read(&n->nr_slabs))
3817            continue;
3818
3819        spin_lock_irqsave(&n->list_lock, flags);
3820        list_for_each_entry(page, &n->partial, lru)
3821            process_slab(&t, s, page, alloc, map);
3822        list_for_each_entry(page, &n->full, lru)
3823            process_slab(&t, s, page, alloc, map);
3824        spin_unlock_irqrestore(&n->list_lock, flags);
3825    }
3826
3827    for (i = 0; i < t.count; i++) {
3828        struct location *l = &t.loc[i];
3829
3830        if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
3831            break;
3832        len += sprintf(buf + len, "%7ld ", l->count);
3833
3834        if (l->addr)
3835            len += sprintf(buf + len, "%pS", (void *)l->addr);
3836        else
3837            len += sprintf(buf + len, "<not-available>");
3838
3839        if (l->sum_time != l->min_time) {
3840            len += sprintf(buf + len, " age=%ld/%ld/%ld",
3841                l->min_time,
3842                (long)div_u64(l->sum_time, l->count),
3843                l->max_time);
3844        } else
3845            len += sprintf(buf + len, " age=%ld",
3846                l->min_time);
3847
3848        if (l->min_pid != l->max_pid)
3849            len += sprintf(buf + len, " pid=%ld-%ld",
3850                l->min_pid, l->max_pid);
3851        else
3852            len += sprintf(buf + len, " pid=%ld",
3853                l->min_pid);
3854
3855        if (num_online_cpus() > 1 &&
3856                !cpumask_empty(to_cpumask(l->cpus)) &&
3857                len < PAGE_SIZE - 60) {
3858            len += sprintf(buf + len, " cpus=");
3859            len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3860                         to_cpumask(l->cpus));
3861        }
3862
3863        if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
3864                len < PAGE_SIZE - 60) {
3865            len += sprintf(buf + len, " nodes=");
3866            len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3867                    l->nodes);
3868        }
3869
3870        len += sprintf(buf + len, "\n");
3871    }
3872
3873    free_loc_track(&t);
3874    kfree(map);
3875    if (!t.count)
3876        len += sprintf(buf, "No data\n");
3877    return len;
3878}
3879#endif
3880
3881#ifdef SLUB_RESILIENCY_TEST
3882static void resiliency_test(void)
3883{
3884    u8 *p;
3885
3886    BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
3887
3888    printk(KERN_ERR "SLUB resiliency testing\n");
3889    printk(KERN_ERR "-----------------------\n");
3890    printk(KERN_ERR "A. Corruption after allocation\n");
3891
3892    p = kzalloc(16, GFP_KERNEL);
3893    p[16] = 0x12;
3894    printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3895            " 0x12->0x%p\n\n", p + 16);
3896
3897    validate_slab_cache(kmalloc_caches[4]);
3898
3899    /* Hmmm... The next two are dangerous */
3900    p = kzalloc(32, GFP_KERNEL);
3901    p[32 + sizeof(void *)] = 0x34;
3902    printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3903            " 0x34 -> -0x%p\n", p);
3904    printk(KERN_ERR
3905        "If allocated object is overwritten then not detectable\n\n");
3906
3907    validate_slab_cache(kmalloc_caches[5]);
3908    p = kzalloc(64, GFP_KERNEL);
3909    p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3910    *p = 0x56;
3911    printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3912                                    p);
3913    printk(KERN_ERR
3914        "If allocated object is overwritten then not detectable\n\n");
3915    validate_slab_cache(kmalloc_caches[6]);
3916
3917    printk(KERN_ERR "\nB. Corruption after free\n");
3918    p = kzalloc(128, GFP_KERNEL);
3919    kfree(p);
3920    *p = 0x78;
3921    printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3922    validate_slab_cache(kmalloc_caches[7]);
3923
3924    p = kzalloc(256, GFP_KERNEL);
3925    kfree(p);
3926    p[50] = 0x9a;
3927    printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3928            p);
3929    validate_slab_cache(kmalloc_caches[8]);
3930
3931    p = kzalloc(512, GFP_KERNEL);
3932    kfree(p);
3933    p[512] = 0xab;
3934    printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3935    validate_slab_cache(kmalloc_caches[9]);
3936}
3937#else
3938#ifdef CONFIG_SYSFS
3939static void resiliency_test(void) {};
3940#endif
3941#endif
3942
3943#ifdef CONFIG_SYSFS
3944enum slab_stat_type {
3945    SL_ALL, /* All slabs */
3946    SL_PARTIAL, /* Only partially allocated slabs */
3947    SL_CPU, /* Only slabs used for cpu caches */
3948    SL_OBJECTS, /* Determine allocated objects not slabs */
3949    SL_TOTAL /* Determine object capacity not slabs */
3950};
3951
3952#define SO_ALL (1 << SL_ALL)
3953#define SO_PARTIAL (1 << SL_PARTIAL)
3954#define SO_CPU (1 << SL_CPU)
3955#define SO_OBJECTS (1 << SL_OBJECTS)
3956#define SO_TOTAL (1 << SL_TOTAL)
3957
3958static ssize_t show_slab_objects(struct kmem_cache *s,
3959                char *buf, unsigned long flags)
3960{
3961    unsigned long total = 0;
3962    int node;
3963    int x;
3964    unsigned long *nodes;
3965    unsigned long *per_cpu;
3966
3967    nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3968    if (!nodes)
3969        return -ENOMEM;
3970    per_cpu = nodes + nr_node_ids;
3971
3972    if (flags & SO_CPU) {
3973        int cpu;
3974
3975        for_each_possible_cpu(cpu) {
3976            struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
3977
3978            if (!c || c->node < 0)
3979                continue;
3980
3981            if (c->page) {
3982                    if (flags & SO_TOTAL)
3983                        x = c->page->objects;
3984                else if (flags & SO_OBJECTS)
3985                    x = c->page->inuse;
3986                else
3987                    x = 1;
3988
3989                total += x;
3990                nodes[c->node] += x;
3991            }
3992            per_cpu[c->node]++;
3993        }
3994    }
3995
3996    lock_memory_hotplug();
3997#ifdef CONFIG_SLUB_DEBUG
3998    if (flags & SO_ALL) {
3999        for_each_node_state(node, N_NORMAL_MEMORY) {
4000            struct kmem_cache_node *n = get_node(s, node);
4001
4002        if (flags & SO_TOTAL)
4003            x = atomic_long_read(&n->total_objects);
4004        else if (flags & SO_OBJECTS)
4005            x = atomic_long_read(&n->total_objects) -
4006                count_partial(n, count_free);
4007
4008            else
4009                x = atomic_long_read(&n->nr_slabs);
4010            total += x;
4011            nodes[node] += x;
4012        }
4013
4014    } else
4015#endif
4016    if (flags & SO_PARTIAL) {
4017        for_each_node_state(node, N_NORMAL_MEMORY) {
4018            struct kmem_cache_node *n = get_node(s, node);
4019
4020            if (flags & SO_TOTAL)
4021                x = count_partial(n, count_total);
4022            else if (flags & SO_OBJECTS)
4023                x = count_partial(n, count_inuse);
4024            else
4025                x = n->nr_partial;
4026            total += x;
4027            nodes[node] += x;
4028        }
4029    }
4030    x = sprintf(buf, "%lu", total);
4031#ifdef CONFIG_NUMA
4032    for_each_node_state(node, N_NORMAL_MEMORY)
4033        if (nodes[node])
4034            x += sprintf(buf + x, " N%d=%lu",
4035                    node, nodes[node]);
4036#endif
4037    unlock_memory_hotplug();
4038    kfree(nodes);
4039    return x + sprintf(buf + x, "\n");
4040}
4041
4042#ifdef CONFIG_SLUB_DEBUG
4043static int any_slab_objects(struct kmem_cache *s)
4044{
4045    int node;
4046
4047    for_each_online_node(node) {
4048        struct kmem_cache_node *n = get_node(s, node);
4049
4050        if (!n)
4051            continue;
4052
4053        if (atomic_long_read(&n->total_objects))
4054            return 1;
4055    }
4056    return 0;
4057}
4058#endif
4059
4060#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4061#define to_slab(n) container_of(n, struct kmem_cache, kobj);
4062
4063struct slab_attribute {
4064    struct attribute attr;
4065    ssize_t (*show)(struct kmem_cache *s, char *buf);
4066    ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4067};
4068
4069#define SLAB_ATTR_RO(_name) \
4070    static struct slab_attribute _name##_attr = __ATTR_RO(_name)
4071
4072#define SLAB_ATTR(_name) \
4073    static struct slab_attribute _name##_attr = \
4074    __ATTR(_name, 0644, _name##_show, _name##_store)
4075
4076static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4077{
4078    return sprintf(buf, "%d\n", s->size);
4079}
4080SLAB_ATTR_RO(slab_size);
4081
4082static ssize_t align_show(struct kmem_cache *s, char *buf)
4083{
4084    return sprintf(buf, "%d\n", s->align);
4085}
4086SLAB_ATTR_RO(align);
4087
4088static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4089{
4090    return sprintf(buf, "%d\n", s->objsize);
4091}
4092SLAB_ATTR_RO(object_size);
4093
4094static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4095{
4096    return sprintf(buf, "%d\n", oo_objects(s->oo));
4097}
4098SLAB_ATTR_RO(objs_per_slab);
4099
4100static ssize_t order_store(struct kmem_cache *s,
4101                const char *buf, size_t length)
4102{
4103    unsigned long order;
4104    int err;
4105
4106    err = strict_strtoul(buf, 10, &order);
4107    if (err)
4108        return err;
4109
4110    if (order > slub_max_order || order < slub_min_order)
4111        return -EINVAL;
4112
4113    calculate_sizes(s, order);
4114    return length;
4115}
4116
4117static ssize_t order_show(struct kmem_cache *s, char *buf)
4118{
4119    return sprintf(buf, "%d\n", oo_order(s->oo));
4120}
4121SLAB_ATTR(order);
4122
4123static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4124{
4125    return sprintf(buf, "%lu\n", s->min_partial);
4126}
4127
4128static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4129                 size_t length)
4130{
4131    unsigned long min;
4132    int err;
4133
4134    err = strict_strtoul(buf, 10, &min);
4135    if (err)
4136        return err;
4137
4138    set_min_partial(s, min);
4139    return length;
4140}
4141SLAB_ATTR(min_partial);
4142
4143static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4144{
4145    if (!s->ctor)
4146        return 0;
4147    return sprintf(buf, "%pS\n", s->ctor);
4148}
4149SLAB_ATTR_RO(ctor);
4150
4151static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4152{
4153    return sprintf(buf, "%d\n", s->refcount - 1);
4154}
4155SLAB_ATTR_RO(aliases);
4156
4157static ssize_t partial_show(struct kmem_cache *s, char *buf)
4158{
4159    return show_slab_objects(s, buf, SO_PARTIAL);
4160}
4161SLAB_ATTR_RO(partial);
4162
4163static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4164{
4165    return show_slab_objects(s, buf, SO_CPU);
4166}
4167SLAB_ATTR_RO(cpu_slabs);
4168
4169static ssize_t objects_show(struct kmem_cache *s, char *buf)
4170{
4171    return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4172}
4173SLAB_ATTR_RO(objects);
4174
4175static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4176{
4177    return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4178}
4179SLAB_ATTR_RO(objects_partial);
4180
4181static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4182{
4183    return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4184}
4185
4186static ssize_t reclaim_account_store(struct kmem_cache *s,
4187                const char *buf, size_t length)
4188{
4189    s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4190    if (buf[0] == '1')
4191        s->flags |= SLAB_RECLAIM_ACCOUNT;
4192    return length;
4193}
4194SLAB_ATTR(reclaim_account);
4195
4196static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4197{
4198    return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4199}
4200SLAB_ATTR_RO(hwcache_align);
4201
4202#ifdef CONFIG_ZONE_DMA
4203static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4204{
4205    return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4206}
4207SLAB_ATTR_RO(cache_dma);
4208#endif
4209
4210static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4211{
4212    return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4213}
4214SLAB_ATTR_RO(destroy_by_rcu);
4215
4216static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4217{
4218    return sprintf(buf, "%d\n", s->reserved);
4219}
4220SLAB_ATTR_RO(reserved);
4221
4222#ifdef CONFIG_SLUB_DEBUG
4223static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4224{
4225    return show_slab_objects(s, buf, SO_ALL);
4226}
4227SLAB_ATTR_RO(slabs);
4228
4229static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4230{
4231    return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4232}
4233SLAB_ATTR_RO(total_objects);
4234
4235static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4236{
4237    return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4238}
4239
4240static ssize_t sanity_checks_store(struct kmem_cache *s,
4241                const char *buf, size_t length)
4242{
4243    s->flags &= ~SLAB_DEBUG_FREE;
4244    if (buf[0] == '1')
4245        s->flags |= SLAB_DEBUG_FREE;
4246    return length;
4247}
4248SLAB_ATTR(sanity_checks);
4249
4250static ssize_t trace_show(struct kmem_cache *s, char *buf)
4251{
4252    return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4253}
4254
4255static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4256                            size_t length)
4257{
4258    s->flags &= ~SLAB_TRACE;
4259    if (buf[0] == '1')
4260        s->flags |= SLAB_TRACE;
4261    return length;
4262}
4263SLAB_ATTR(trace);
4264
4265static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4266{
4267    return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4268}
4269
4270static ssize_t red_zone_store(struct kmem_cache *s,
4271                const char *buf, size_t length)
4272{
4273    if (any_slab_objects(s))
4274        return -EBUSY;
4275
4276    s->flags &= ~SLAB_RED_ZONE;
4277    if (buf[0] == '1')
4278        s->flags |= SLAB_RED_ZONE;
4279    calculate_sizes(s, -1);
4280    return length;
4281}
4282SLAB_ATTR(red_zone);
4283
4284static ssize_t poison_show(struct kmem_cache *s, char *buf)
4285{
4286    return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4287}
4288
4289static ssize_t poison_store(struct kmem_cache *s,
4290                const char *buf, size_t length)
4291{
4292    if (any_slab_objects(s))
4293        return -EBUSY;
4294
4295    s->flags &= ~SLAB_POISON;
4296    if (buf[0] == '1')
4297        s->flags |= SLAB_POISON;
4298    calculate_sizes(s, -1);
4299    return length;
4300}
4301SLAB_ATTR(poison);
4302
4303static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4304{
4305    return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4306}
4307
4308static ssize_t store_user_store(struct kmem_cache *s,
4309                const char *buf, size_t length)
4310{
4311    if (any_slab_objects(s))
4312        return -EBUSY;
4313
4314    s->flags &= ~SLAB_STORE_USER;
4315    if (buf[0] == '1')
4316        s->flags |= SLAB_STORE_USER;
4317    calculate_sizes(s, -1);
4318    return length;
4319}
4320SLAB_ATTR(store_user);
4321
4322static ssize_t validate_show(struct kmem_cache *s, char *buf)
4323{
4324    return 0;
4325}
4326
4327static ssize_t validate_store(struct kmem_cache *s,
4328            const char *buf, size_t length)
4329{
4330    int ret = -EINVAL;
4331
4332    if (buf[0] == '1') {
4333        ret = validate_slab_cache(s);
4334        if (ret >= 0)
4335            ret = length;
4336    }
4337    return ret;
4338}
4339SLAB_ATTR(validate);
4340
4341static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4342{
4343    if (!(s->flags & SLAB_STORE_USER))
4344        return -ENOSYS;
4345    return list_locations(s, buf, TRACK_ALLOC);
4346}
4347SLAB_ATTR_RO(alloc_calls);
4348
4349static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4350{
4351    if (!(s->flags & SLAB_STORE_USER))
4352        return -ENOSYS;
4353    return list_locations(s, buf, TRACK_FREE);
4354}
4355SLAB_ATTR_RO(free_calls);
4356#endif /* CONFIG_SLUB_DEBUG */
4357
4358#ifdef CONFIG_FAILSLAB
4359static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4360{
4361    return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4362}
4363
4364static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4365                            size_t length)
4366{
4367    s->flags &= ~SLAB_FAILSLAB;
4368    if (buf[0] == '1')
4369        s->flags |= SLAB_FAILSLAB;
4370    return length;
4371}
4372SLAB_ATTR(failslab);
4373#endif
4374
4375static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4376{
4377    return 0;
4378}
4379
4380static ssize_t shrink_store(struct kmem_cache *s,
4381            const char *buf, size_t length)
4382{
4383    if (buf[0] == '1') {
4384        int rc = kmem_cache_shrink(s);
4385
4386        if (rc)
4387            return rc;
4388    } else
4389        return -EINVAL;
4390    return length;
4391}
4392SLAB_ATTR(shrink);
4393
4394#ifdef CONFIG_NUMA
4395static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4396{
4397    return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4398}
4399
4400static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4401                const char *buf, size_t length)
4402{
4403    unsigned long ratio;
4404    int err;
4405
4406    err = strict_strtoul(buf, 10, &ratio);
4407    if (err)
4408        return err;
4409
4410    if (ratio <= 100)
4411        s->remote_node_defrag_ratio = ratio * 10;
4412
4413    return length;
4414}
4415SLAB_ATTR(remote_node_defrag_ratio);
4416#endif
4417
4418#ifdef CONFIG_SLUB_STATS
4419static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4420{
4421    unsigned long sum = 0;
4422    int cpu;
4423    int len;
4424    int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4425
4426    if (!data)
4427        return -ENOMEM;
4428
4429    for_each_online_cpu(cpu) {
4430        unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4431
4432        data[cpu] = x;
4433        sum += x;
4434    }
4435
4436    len = sprintf(buf, "%lu", sum);
4437
4438#ifdef CONFIG_SMP
4439    for_each_online_cpu(cpu) {
4440        if (data[cpu] && len < PAGE_SIZE - 20)
4441            len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4442    }
4443#endif
4444    kfree(data);
4445    return len + sprintf(buf + len, "\n");
4446}
4447
4448static void clear_stat(struct kmem_cache *s, enum stat_item si)
4449{
4450    int cpu;
4451
4452    for_each_online_cpu(cpu)
4453        per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4454}
4455
4456#define STAT_ATTR(si, text) \
4457static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4458{ \
4459    return show_stat(s, buf, si); \
4460} \
4461static ssize_t text##_store(struct kmem_cache *s, \
4462                const char *buf, size_t length) \
4463{ \
4464    if (buf[0] != '0') \
4465        return -EINVAL; \
4466    clear_stat(s, si); \
4467    return length; \
4468} \
4469SLAB_ATTR(text); \
4470
4471STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4472STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4473STAT_ATTR(FREE_FASTPATH, free_fastpath);
4474STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4475STAT_ATTR(FREE_FROZEN, free_frozen);
4476STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4477STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4478STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4479STAT_ATTR(ALLOC_SLAB, alloc_slab);
4480STAT_ATTR(ALLOC_REFILL, alloc_refill);
4481STAT_ATTR(FREE_SLAB, free_slab);
4482STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4483STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4484STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4485STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4486STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4487STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4488STAT_ATTR(ORDER_FALLBACK, order_fallback);
4489#endif
4490
4491static struct attribute *slab_attrs[] = {
4492    &slab_size_attr.attr,
4493    &object_size_attr.attr,
4494    &objs_per_slab_attr.attr,
4495    &order_attr.attr,
4496    &min_partial_attr.attr,
4497    &objects_attr.attr,
4498    &objects_partial_attr.attr,
4499    &partial_attr.attr,
4500    &cpu_slabs_attr.attr,
4501    &ctor_attr.attr,
4502    &aliases_attr.attr,
4503    &align_attr.attr,
4504    &hwcache_align_attr.attr,
4505    &reclaim_account_attr.attr,
4506    &destroy_by_rcu_attr.attr,
4507    &shrink_attr.attr,
4508    &reserved_attr.attr,
4509#ifdef CONFIG_SLUB_DEBUG
4510    &total_objects_attr.attr,
4511    &slabs_attr.attr,
4512    &sanity_checks_attr.attr,
4513    &trace_attr.attr,
4514    &red_zone_attr.attr,
4515    &poison_attr.attr,
4516    &store_user_attr.attr,
4517    &validate_attr.attr,
4518    &alloc_calls_attr.attr,
4519    &free_calls_attr.attr,
4520#endif
4521#ifdef CONFIG_ZONE_DMA
4522    &cache_dma_attr.attr,
4523#endif
4524#ifdef CONFIG_NUMA
4525    &remote_node_defrag_ratio_attr.attr,
4526#endif
4527#ifdef CONFIG_SLUB_STATS
4528    &alloc_fastpath_attr.attr,
4529    &alloc_slowpath_attr.attr,
4530    &free_fastpath_attr.attr,
4531    &free_slowpath_attr.attr,
4532    &free_frozen_attr.attr,
4533    &free_add_partial_attr.attr,
4534    &free_remove_partial_attr.attr,
4535    &alloc_from_partial_attr.attr,
4536    &alloc_slab_attr.attr,
4537    &alloc_refill_attr.attr,
4538    &free_slab_attr.attr,
4539    &cpuslab_flush_attr.attr,
4540    &deactivate_full_attr.attr,
4541    &deactivate_empty_attr.attr,
4542    &deactivate_to_head_attr.attr,
4543    &deactivate_to_tail_attr.attr,
4544    &deactivate_remote_frees_attr.attr,
4545    &order_fallback_attr.attr,
4546#endif
4547#ifdef CONFIG_FAILSLAB
4548    &failslab_attr.attr,
4549#endif
4550
4551    NULL
4552};
4553
4554static struct attribute_group slab_attr_group = {
4555    .attrs = slab_attrs,
4556};
4557
4558static ssize_t slab_attr_show(struct kobject *kobj,
4559                struct attribute *attr,
4560                char *buf)
4561{
4562    struct slab_attribute *attribute;
4563    struct kmem_cache *s;
4564    int err;
4565
4566    attribute = to_slab_attr(attr);
4567    s = to_slab(kobj);
4568
4569    if (!attribute->show)
4570        return -EIO;
4571
4572    err = attribute->show(s, buf);
4573
4574    return err;
4575}
4576
4577static ssize_t slab_attr_store(struct kobject *kobj,
4578                struct attribute *attr,
4579                const char *buf, size_t len)
4580{
4581    struct slab_attribute *attribute;
4582    struct kmem_cache *s;
4583    int err;
4584
4585    attribute = to_slab_attr(attr);
4586    s = to_slab(kobj);
4587
4588    if (!attribute->store)
4589        return -EIO;
4590
4591    err = attribute->store(s, buf, len);
4592
4593    return err;
4594}
4595
4596static void kmem_cache_release(struct kobject *kobj)
4597{
4598    struct kmem_cache *s = to_slab(kobj);
4599
4600    kfree(s->name);
4601    kfree(s);
4602}
4603
4604static const struct sysfs_ops slab_sysfs_ops = {
4605    .show = slab_attr_show,
4606    .store = slab_attr_store,
4607};
4608
4609static struct kobj_type slab_ktype = {
4610    .sysfs_ops = &slab_sysfs_ops,
4611    .release = kmem_cache_release
4612};
4613
4614static int uevent_filter(struct kset *kset, struct kobject *kobj)
4615{
4616    struct kobj_type *ktype = get_ktype(kobj);
4617
4618    if (ktype == &slab_ktype)
4619        return 1;
4620    return 0;
4621}
4622
4623static const struct kset_uevent_ops slab_uevent_ops = {
4624    .filter = uevent_filter,
4625};
4626
4627static struct kset *slab_kset;
4628
4629#define ID_STR_LENGTH 64
4630
4631/* Create a unique string id for a slab cache:
4632 *
4633 * Format :[flags-]size
4634 */
4635static char *create_unique_id(struct kmem_cache *s)
4636{
4637    char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4638    char *p = name;
4639
4640    BUG_ON(!name);
4641
4642    *p++ = ':';
4643    /*
4644     * First flags affecting slabcache operations. We will only
4645     * get here for aliasable slabs so we do not need to support
4646     * too many flags. The flags here must cover all flags that
4647     * are matched during merging to guarantee that the id is
4648     * unique.
4649     */
4650    if (s->flags & SLAB_CACHE_DMA)
4651        *p++ = 'd';
4652    if (s->flags & SLAB_RECLAIM_ACCOUNT)
4653        *p++ = 'a';
4654    if (s->flags & SLAB_DEBUG_FREE)
4655        *p++ = 'F';
4656    if (!(s->flags & SLAB_NOTRACK))
4657        *p++ = 't';
4658    if (p != name + 1)
4659        *p++ = '-';
4660    p += sprintf(p, "%07d", s->size);
4661    BUG_ON(p > name + ID_STR_LENGTH - 1);
4662    return name;
4663}
4664
4665static int sysfs_slab_add(struct kmem_cache *s)
4666{
4667    int err;
4668    const char *name;
4669    int unmergeable;
4670
4671    if (slab_state < SYSFS)
4672        /* Defer until later */
4673        return 0;
4674
4675    unmergeable = slab_unmergeable(s);
4676    if (unmergeable) {
4677        /*
4678         * Slabcache can never be merged so we can use the name proper.
4679         * This is typically the case for debug situations. In that
4680         * case we can catch duplicate names easily.
4681         */
4682        sysfs_remove_link(&slab_kset->kobj, s->name);
4683        name = s->name;
4684    } else {
4685        /*
4686         * Create a unique name for the slab as a target
4687         * for the symlinks.
4688         */
4689        name = create_unique_id(s);
4690    }
4691
4692    s->kobj.kset = slab_kset;
4693    err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4694    if (err) {
4695        kobject_put(&s->kobj);
4696        return err;
4697    }
4698
4699    err = sysfs_create_group(&s->kobj, &slab_attr_group);
4700    if (err) {
4701        kobject_del(&s->kobj);
4702        kobject_put(&s->kobj);
4703        return err;
4704    }
4705    kobject_uevent(&s->kobj, KOBJ_ADD);
4706    if (!unmergeable) {
4707        /* Setup first alias */
4708        sysfs_slab_alias(s, s->name);
4709        kfree(name);
4710    }
4711    return 0;
4712}
4713
4714static void sysfs_slab_remove(struct kmem_cache *s)
4715{
4716    if (slab_state < SYSFS)
4717        /*
4718         * Sysfs has not been setup yet so no need to remove the
4719         * cache from sysfs.
4720         */
4721        return;
4722
4723    kobject_uevent(&s->kobj, KOBJ_REMOVE);
4724    kobject_del(&s->kobj);
4725    kobject_put(&s->kobj);
4726}
4727
4728/*
4729 * Need to buffer aliases during bootup until sysfs becomes
4730 * available lest we lose that information.
4731 */
4732struct saved_alias {
4733    struct kmem_cache *s;
4734    const char *name;
4735    struct saved_alias *next;
4736};
4737
4738static struct saved_alias *alias_list;
4739
4740static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4741{
4742    struct saved_alias *al;
4743
4744    if (slab_state == SYSFS) {
4745        /*
4746         * If we have a leftover link then remove it.
4747         */
4748        sysfs_remove_link(&slab_kset->kobj, name);
4749        return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
4750    }
4751
4752    al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4753    if (!al)
4754        return -ENOMEM;
4755
4756    al->s = s;
4757    al->name = name;
4758    al->next = alias_list;
4759    alias_list = al;
4760    return 0;
4761}
4762
4763static int __init slab_sysfs_init(void)
4764{
4765    struct kmem_cache *s;
4766    int err;
4767
4768    down_write(&slub_lock);
4769
4770    slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
4771    if (!slab_kset) {
4772        up_write(&slub_lock);
4773        printk(KERN_ERR "Cannot register slab subsystem.\n");
4774        return -ENOSYS;
4775    }
4776
4777    slab_state = SYSFS;
4778
4779    list_for_each_entry(s, &slab_caches, list) {
4780        err = sysfs_slab_add(s);
4781        if (err)
4782            printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4783                        " to sysfs\n", s->name);
4784    }
4785
4786    while (alias_list) {
4787        struct saved_alias *al = alias_list;
4788
4789        alias_list = alias_list->next;
4790        err = sysfs_slab_alias(al->s, al->name);
4791        if (err)
4792            printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4793                    " %s to sysfs\n", s->name);
4794        kfree(al);
4795    }
4796
4797    up_write(&slub_lock);
4798    resiliency_test();
4799    return 0;
4800}
4801
4802__initcall(slab_sysfs_init);
4803#endif /* CONFIG_SYSFS */
4804
4805/*
4806 * The /proc/slabinfo ABI
4807 */
4808#ifdef CONFIG_SLABINFO
4809static void print_slabinfo_header(struct seq_file *m)
4810{
4811    seq_puts(m, "slabinfo - version: 2.1\n");
4812    seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4813         "<objperslab> <pagesperslab>");
4814    seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4815    seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4816    seq_putc(m, '\n');
4817}
4818
4819static void *s_start(struct seq_file *m, loff_t *pos)
4820{
4821    loff_t n = *pos;
4822
4823    down_read(&slub_lock);
4824    if (!n)
4825        print_slabinfo_header(m);
4826
4827    return seq_list_start(&slab_caches, *pos);
4828}
4829
4830static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4831{
4832    return seq_list_next(p, &slab_caches, pos);
4833}
4834
4835static void s_stop(struct seq_file *m, void *p)
4836{
4837    up_read(&slub_lock);
4838}
4839
4840static int s_show(struct seq_file *m, void *p)
4841{
4842    unsigned long nr_partials = 0;
4843    unsigned long nr_slabs = 0;
4844    unsigned long nr_inuse = 0;
4845    unsigned long nr_objs = 0;
4846    unsigned long nr_free = 0;
4847    struct kmem_cache *s;
4848    int node;
4849
4850    s = list_entry(p, struct kmem_cache, list);
4851
4852    for_each_online_node(node) {
4853        struct kmem_cache_node *n = get_node(s, node);
4854
4855        if (!n)
4856            continue;
4857
4858        nr_partials += n->nr_partial;
4859        nr_slabs += atomic_long_read(&n->nr_slabs);
4860        nr_objs += atomic_long_read(&n->total_objects);
4861        nr_free += count_partial(n, count_free);
4862    }
4863
4864    nr_inuse = nr_objs - nr_free;
4865
4866    seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4867           nr_objs, s->size, oo_objects(s->oo),
4868           (1 << oo_order(s->oo)));
4869    seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4870    seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4871           0UL);
4872    seq_putc(m, '\n');
4873    return 0;
4874}
4875
4876static const struct seq_operations slabinfo_op = {
4877    .start = s_start,
4878    .next = s_next,
4879    .stop = s_stop,
4880    .show = s_show,
4881};
4882
4883static int slabinfo_open(struct inode *inode, struct file *file)
4884{
4885    return seq_open(file, &slabinfo_op);
4886}
4887
4888static const struct file_operations proc_slabinfo_operations = {
4889    .open = slabinfo_open,
4890    .read = seq_read,
4891    .llseek = seq_lseek,
4892    .release = seq_release,
4893};
4894
4895static int __init slab_proc_init(void)
4896{
4897    proc_create("slabinfo", S_IRUGO, NULL, &proc_slabinfo_operations);
4898    return 0;
4899}
4900module_init(slab_proc_init);
4901#endif /* CONFIG_SLABINFO */
4902

Archive Download this file



interactive