Loading...
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 or atomic operatios
6 * and only uses a centralized lock to manage a pool of partial slabs.
7 *
8 * (C) 2007 SGI, Christoph Lameter
9 * (C) 2011 Linux Foundation, Christoph Lameter
10 */
11
12#include <linux/mm.h>
13#include <linux/swap.h> /* struct reclaim_state */
14#include <linux/module.h>
15#include <linux/bit_spinlock.h>
16#include <linux/interrupt.h>
17#include <linux/bitops.h>
18#include <linux/slab.h>
19#include <linux/proc_fs.h>
20#include <linux/seq_file.h>
21#include <linux/kmemcheck.h>
22#include <linux/cpu.h>
23#include <linux/cpuset.h>
24#include <linux/mempolicy.h>
25#include <linux/ctype.h>
26#include <linux/debugobjects.h>
27#include <linux/kallsyms.h>
28#include <linux/memory.h>
29#include <linux/math64.h>
30#include <linux/fault-inject.h>
31#include <linux/stacktrace.h>
32
33#include <trace/events/kmem.h>
34
35/*
36 * Lock order:
37 * 1. slub_lock (Global Semaphore)
38 * 2. node->list_lock
39 * 3. slab_lock(page) (Only on some arches and for debugging)
40 *
41 * slub_lock
42 *
43 * The role of the slub_lock is to protect the list of all the slabs
44 * and to synchronize major metadata changes to slab cache structures.
45 *
46 * The slab_lock is only used for debugging and on arches that do not
47 * have the ability to do a cmpxchg_double. It only protects the second
48 * double word in the page struct. Meaning
49 * A. page->freelist -> List of object free in a page
50 * B. page->counters -> Counters of objects
51 * C. page->frozen -> frozen state
52 *
53 * If a slab is frozen then it is exempt from list management. It is not
54 * on any list. The processor that froze the slab is the one who can
55 * perform list operations on the page. Other processors may put objects
56 * onto the freelist but the processor that froze the slab is the only
57 * one that can retrieve the objects from the page's freelist.
58 *
59 * The list_lock protects the partial and full list on each node and
60 * the partial slab counter. If taken then no new slabs may be added or
61 * removed from the lists nor make the number of partial slabs be modified.
62 * (Note that the total number of slabs is an atomic value that may be
63 * modified without taking the list lock).
64 *
65 * The list_lock is a centralized lock and thus we avoid taking it as
66 * much as possible. As long as SLUB does not have to handle partial
67 * slabs, operations can continue without any centralized lock. F.e.
68 * allocating a long series of objects that fill up slabs does not require
69 * the list lock.
70 * Interrupts are disabled during allocation and deallocation in order to
71 * make the slab allocator safe to use in the context of an irq. In addition
72 * interrupts are disabled to ensure that the processor does not change
73 * while handling per_cpu slabs, due to kernel preemption.
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/* Enable to log cmpxchg failures */
135#undef SLUB_DEBUG_CMPXCHG
136
137/*
138 * Mininum number of partial slabs. These will be left on the partial
139 * lists even if they are empty. kmem_cache_shrink may reclaim them.
140 */
141#define MIN_PARTIAL 5
142
143/*
144 * Maximum number of desirable partial slabs.
145 * The existence of more partial slabs makes kmem_cache_shrink
146 * sort the partial list by the number of objects in the.
147 */
148#define MAX_PARTIAL 10
149
150#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
151 SLAB_POISON | SLAB_STORE_USER)
152
153/*
154 * Debugging flags that require metadata to be stored in the slab. These get
155 * disabled when slub_debug=O is used and a cache's min order increases with
156 * metadata.
157 */
158#define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
159
160/*
161 * Set of flags that will prevent slab merging
162 */
163#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
164 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
165 SLAB_FAILSLAB)
166
167#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
168 SLAB_CACHE_DMA | SLAB_NOTRACK)
169
170#define OO_SHIFT 16
171#define OO_MASK ((1 << OO_SHIFT) - 1)
172#define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
173
174/* Internal SLUB flags */
175#define __OBJECT_POISON 0x80000000UL /* Poison object */
176#define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
177
178static int kmem_size = sizeof(struct kmem_cache);
179
180#ifdef CONFIG_SMP
181static struct notifier_block slab_notifier;
182#endif
183
184static enum {
185 DOWN, /* No slab functionality available */
186 PARTIAL, /* Kmem_cache_node works */
187 UP, /* Everything works but does not show up in sysfs */
188 SYSFS /* Sysfs up */
189} slab_state = DOWN;
190
191/* A list of all slab caches on the system */
192static DECLARE_RWSEM(slub_lock);
193static LIST_HEAD(slab_caches);
194
195/*
196 * Tracking user of a slab.
197 */
198#define TRACK_ADDRS_COUNT 16
199struct track {
200 unsigned long addr; /* Called from address */
201#ifdef CONFIG_STACKTRACE
202 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
203#endif
204 int cpu; /* Was running on cpu */
205 int pid; /* Pid context */
206 unsigned long when; /* When did the operation occur */
207};
208
209enum track_item { TRACK_ALLOC, TRACK_FREE };
210
211#ifdef CONFIG_SYSFS
212static int sysfs_slab_add(struct kmem_cache *);
213static int sysfs_slab_alias(struct kmem_cache *, const char *);
214static void sysfs_slab_remove(struct kmem_cache *);
215
216#else
217static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
218static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
219 { return 0; }
220static inline void sysfs_slab_remove(struct kmem_cache *s)
221{
222 kfree(s->name);
223 kfree(s);
224}
225
226#endif
227
228static inline void stat(const struct kmem_cache *s, enum stat_item si)
229{
230#ifdef CONFIG_SLUB_STATS
231 __this_cpu_inc(s->cpu_slab->stat[si]);
232#endif
233}
234
235/********************************************************************
236 * Core slab cache functions
237 *******************************************************************/
238
239int slab_is_available(void)
240{
241 return slab_state >= UP;
242}
243
244static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
245{
246 return s->node[node];
247}
248
249/* Verify that a pointer has an address that is valid within a slab page */
250static inline int check_valid_pointer(struct kmem_cache *s,
251 struct page *page, const void *object)
252{
253 void *base;
254
255 if (!object)
256 return 1;
257
258 base = page_address(page);
259 if (object < base || object >= base + page->objects * s->size ||
260 (object - base) % s->size) {
261 return 0;
262 }
263
264 return 1;
265}
266
267static inline void *get_freepointer(struct kmem_cache *s, void *object)
268{
269 return *(void **)(object + s->offset);
270}
271
272static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
273{
274 void *p;
275
276#ifdef CONFIG_DEBUG_PAGEALLOC
277 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
278#else
279 p = get_freepointer(s, object);
280#endif
281 return p;
282}
283
284static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
285{
286 *(void **)(object + s->offset) = fp;
287}
288
289/* Loop over all objects in a slab */
290#define for_each_object(__p, __s, __addr, __objects) \
291 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
292 __p += (__s)->size)
293
294/* Determine object index from a given position */
295static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
296{
297 return (p - addr) / s->size;
298}
299
300static inline size_t slab_ksize(const struct kmem_cache *s)
301{
302#ifdef CONFIG_SLUB_DEBUG
303 /*
304 * Debugging requires use of the padding between object
305 * and whatever may come after it.
306 */
307 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
308 return s->objsize;
309
310#endif
311 /*
312 * If we have the need to store the freelist pointer
313 * back there or track user information then we can
314 * only use the space before that information.
315 */
316 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
317 return s->inuse;
318 /*
319 * Else we can use all the padding etc for the allocation
320 */
321 return s->size;
322}
323
324static inline int order_objects(int order, unsigned long size, int reserved)
325{
326 return ((PAGE_SIZE << order) - reserved) / size;
327}
328
329static inline struct kmem_cache_order_objects oo_make(int order,
330 unsigned long size, int reserved)
331{
332 struct kmem_cache_order_objects x = {
333 (order << OO_SHIFT) + order_objects(order, size, reserved)
334 };
335
336 return x;
337}
338
339static inline int oo_order(struct kmem_cache_order_objects x)
340{
341 return x.x >> OO_SHIFT;
342}
343
344static inline int oo_objects(struct kmem_cache_order_objects x)
345{
346 return x.x & OO_MASK;
347}
348
349/*
350 * Per slab locking using the pagelock
351 */
352static __always_inline void slab_lock(struct page *page)
353{
354 bit_spin_lock(PG_locked, &page->flags);
355}
356
357static __always_inline void slab_unlock(struct page *page)
358{
359 __bit_spin_unlock(PG_locked, &page->flags);
360}
361
362/* Interrupts must be disabled (for the fallback code to work right) */
363static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
364 void *freelist_old, unsigned long counters_old,
365 void *freelist_new, unsigned long counters_new,
366 const char *n)
367{
368 VM_BUG_ON(!irqs_disabled());
369#ifdef CONFIG_CMPXCHG_DOUBLE
370 if (s->flags & __CMPXCHG_DOUBLE) {
371 if (cmpxchg_double(&page->freelist,
372 freelist_old, counters_old,
373 freelist_new, counters_new))
374 return 1;
375 } else
376#endif
377 {
378 slab_lock(page);
379 if (page->freelist == freelist_old && page->counters == counters_old) {
380 page->freelist = freelist_new;
381 page->counters = counters_new;
382 slab_unlock(page);
383 return 1;
384 }
385 slab_unlock(page);
386 }
387
388 cpu_relax();
389 stat(s, CMPXCHG_DOUBLE_FAIL);
390
391#ifdef SLUB_DEBUG_CMPXCHG
392 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
393#endif
394
395 return 0;
396}
397
398static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
399 void *freelist_old, unsigned long counters_old,
400 void *freelist_new, unsigned long counters_new,
401 const char *n)
402{
403#ifdef CONFIG_CMPXCHG_DOUBLE
404 if (s->flags & __CMPXCHG_DOUBLE) {
405 if (cmpxchg_double(&page->freelist,
406 freelist_old, counters_old,
407 freelist_new, counters_new))
408 return 1;
409 } else
410#endif
411 {
412 unsigned long flags;
413
414 local_irq_save(flags);
415 slab_lock(page);
416 if (page->freelist == freelist_old && page->counters == counters_old) {
417 page->freelist = freelist_new;
418 page->counters = counters_new;
419 slab_unlock(page);
420 local_irq_restore(flags);
421 return 1;
422 }
423 slab_unlock(page);
424 local_irq_restore(flags);
425 }
426
427 cpu_relax();
428 stat(s, CMPXCHG_DOUBLE_FAIL);
429
430#ifdef SLUB_DEBUG_CMPXCHG
431 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
432#endif
433
434 return 0;
435}
436
437#ifdef CONFIG_SLUB_DEBUG
438/*
439 * Determine a map of object in use on a page.
440 *
441 * Node listlock must be held to guarantee that the page does
442 * not vanish from under us.
443 */
444static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
445{
446 void *p;
447 void *addr = page_address(page);
448
449 for (p = page->freelist; p; p = get_freepointer(s, p))
450 set_bit(slab_index(p, s, addr), map);
451}
452
453/*
454 * Debug settings:
455 */
456#ifdef CONFIG_SLUB_DEBUG_ON
457static int slub_debug = DEBUG_DEFAULT_FLAGS;
458#else
459static int slub_debug;
460#endif
461
462static char *slub_debug_slabs;
463static int disable_higher_order_debug;
464
465/*
466 * Object debugging
467 */
468static void print_section(char *text, u8 *addr, unsigned int length)
469{
470 int i, offset;
471 int newline = 1;
472 char ascii[17];
473
474 ascii[16] = 0;
475
476 for (i = 0; i < length; i++) {
477 if (newline) {
478 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
479 newline = 0;
480 }
481 printk(KERN_CONT " %02x", addr[i]);
482 offset = i % 16;
483 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
484 if (offset == 15) {
485 printk(KERN_CONT " %s\n", ascii);
486 newline = 1;
487 }
488 }
489 if (!newline) {
490 i %= 16;
491 while (i < 16) {
492 printk(KERN_CONT " ");
493 ascii[i] = ' ';
494 i++;
495 }
496 printk(KERN_CONT " %s\n", ascii);
497 }
498}
499
500static struct track *get_track(struct kmem_cache *s, void *object,
501 enum track_item alloc)
502{
503 struct track *p;
504
505 if (s->offset)
506 p = object + s->offset + sizeof(void *);
507 else
508 p = object + s->inuse;
509
510 return p + alloc;
511}
512
513static void set_track(struct kmem_cache *s, void *object,
514 enum track_item alloc, unsigned long addr)
515{
516 struct track *p = get_track(s, object, alloc);
517
518 if (addr) {
519#ifdef CONFIG_STACKTRACE
520 struct stack_trace trace;
521 int i;
522
523 trace.nr_entries = 0;
524 trace.max_entries = TRACK_ADDRS_COUNT;
525 trace.entries = p->addrs;
526 trace.skip = 3;
527 save_stack_trace(&trace);
528
529 /* See rant in lockdep.c */
530 if (trace.nr_entries != 0 &&
531 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
532 trace.nr_entries--;
533
534 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
535 p->addrs[i] = 0;
536#endif
537 p->addr = addr;
538 p->cpu = smp_processor_id();
539 p->pid = current->pid;
540 p->when = jiffies;
541 } else
542 memset(p, 0, sizeof(struct track));
543}
544
545static void init_tracking(struct kmem_cache *s, void *object)
546{
547 if (!(s->flags & SLAB_STORE_USER))
548 return;
549
550 set_track(s, object, TRACK_FREE, 0UL);
551 set_track(s, object, TRACK_ALLOC, 0UL);
552}
553
554static void print_track(const char *s, struct track *t)
555{
556 if (!t->addr)
557 return;
558
559 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
560 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
561#ifdef CONFIG_STACKTRACE
562 {
563 int i;
564 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
565 if (t->addrs[i])
566 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
567 else
568 break;
569 }
570#endif
571}
572
573static void print_tracking(struct kmem_cache *s, void *object)
574{
575 if (!(s->flags & SLAB_STORE_USER))
576 return;
577
578 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
579 print_track("Freed", get_track(s, object, TRACK_FREE));
580}
581
582static void print_page_info(struct page *page)
583{
584 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
585 page, page->objects, page->inuse, page->freelist, page->flags);
586
587}
588
589static void slab_bug(struct kmem_cache *s, char *fmt, ...)
590{
591 va_list args;
592 char buf[100];
593
594 va_start(args, fmt);
595 vsnprintf(buf, sizeof(buf), fmt, args);
596 va_end(args);
597 printk(KERN_ERR "========================================"
598 "=====================================\n");
599 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
600 printk(KERN_ERR "----------------------------------------"
601 "-------------------------------------\n\n");
602}
603
604static void slab_fix(struct kmem_cache *s, char *fmt, ...)
605{
606 va_list args;
607 char buf[100];
608
609 va_start(args, fmt);
610 vsnprintf(buf, sizeof(buf), fmt, args);
611 va_end(args);
612 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
613}
614
615static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
616{
617 unsigned int off; /* Offset of last byte */
618 u8 *addr = page_address(page);
619
620 print_tracking(s, p);
621
622 print_page_info(page);
623
624 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
625 p, p - addr, get_freepointer(s, p));
626
627 if (p > addr + 16)
628 print_section("Bytes b4", p - 16, 16);
629
630 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE));
631
632 if (s->flags & SLAB_RED_ZONE)
633 print_section("Redzone", p + s->objsize,
634 s->inuse - s->objsize);
635
636 if (s->offset)
637 off = s->offset + sizeof(void *);
638 else
639 off = s->inuse;
640
641 if (s->flags & SLAB_STORE_USER)
642 off += 2 * sizeof(struct track);
643
644 if (off != s->size)
645 /* Beginning of the filler is the free pointer */
646 print_section("Padding", p + off, s->size - off);
647
648 dump_stack();
649}
650
651static void object_err(struct kmem_cache *s, struct page *page,
652 u8 *object, char *reason)
653{
654 slab_bug(s, "%s", reason);
655 print_trailer(s, page, object);
656}
657
658static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
659{
660 va_list args;
661 char buf[100];
662
663 va_start(args, fmt);
664 vsnprintf(buf, sizeof(buf), fmt, args);
665 va_end(args);
666 slab_bug(s, "%s", buf);
667 print_page_info(page);
668 dump_stack();
669}
670
671static void init_object(struct kmem_cache *s, void *object, u8 val)
672{
673 u8 *p = object;
674
675 if (s->flags & __OBJECT_POISON) {
676 memset(p, POISON_FREE, s->objsize - 1);
677 p[s->objsize - 1] = POISON_END;
678 }
679
680 if (s->flags & SLAB_RED_ZONE)
681 memset(p + s->objsize, val, s->inuse - s->objsize);
682}
683
684static u8 *check_bytes8(u8 *start, u8 value, unsigned int bytes)
685{
686 while (bytes) {
687 if (*start != value)
688 return start;
689 start++;
690 bytes--;
691 }
692 return NULL;
693}
694
695static u8 *check_bytes(u8 *start, u8 value, unsigned int bytes)
696{
697 u64 value64;
698 unsigned int words, prefix;
699
700 if (bytes <= 16)
701 return check_bytes8(start, value, bytes);
702
703 value64 = value | value << 8 | value << 16 | value << 24;
704 value64 = (value64 & 0xffffffff) | value64 << 32;
705 prefix = 8 - ((unsigned long)start) % 8;
706
707 if (prefix) {
708 u8 *r = check_bytes8(start, value, prefix);
709 if (r)
710 return r;
711 start += prefix;
712 bytes -= prefix;
713 }
714
715 words = bytes / 8;
716
717 while (words) {
718 if (*(u64 *)start != value64)
719 return check_bytes8(start, value, 8);
720 start += 8;
721 words--;
722 }
723
724 return check_bytes8(start, value, bytes % 8);
725}
726
727static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
728 void *from, void *to)
729{
730 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
731 memset(from, data, to - from);
732}
733
734static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
735 u8 *object, char *what,
736 u8 *start, unsigned int value, unsigned int bytes)
737{
738 u8 *fault;
739 u8 *end;
740
741 fault = check_bytes(start, value, bytes);
742 if (!fault)
743 return 1;
744
745 end = start + bytes;
746 while (end > fault && end[-1] == value)
747 end--;
748
749 slab_bug(s, "%s overwritten", what);
750 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
751 fault, end - 1, fault[0], value);
752 print_trailer(s, page, object);
753
754 restore_bytes(s, what, value, fault, end);
755 return 0;
756}
757
758/*
759 * Object layout:
760 *
761 * object address
762 * Bytes of the object to be managed.
763 * If the freepointer may overlay the object then the free
764 * pointer is the first word of the object.
765 *
766 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
767 * 0xa5 (POISON_END)
768 *
769 * object + s->objsize
770 * Padding to reach word boundary. This is also used for Redzoning.
771 * Padding is extended by another word if Redzoning is enabled and
772 * objsize == inuse.
773 *
774 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
775 * 0xcc (RED_ACTIVE) for objects in use.
776 *
777 * object + s->inuse
778 * Meta data starts here.
779 *
780 * A. Free pointer (if we cannot overwrite object on free)
781 * B. Tracking data for SLAB_STORE_USER
782 * C. Padding to reach required alignment boundary or at mininum
783 * one word if debugging is on to be able to detect writes
784 * before the word boundary.
785 *
786 * Padding is done using 0x5a (POISON_INUSE)
787 *
788 * object + s->size
789 * Nothing is used beyond s->size.
790 *
791 * If slabcaches are merged then the objsize and inuse boundaries are mostly
792 * ignored. And therefore no slab options that rely on these boundaries
793 * may be used with merged slabcaches.
794 */
795
796static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
797{
798 unsigned long off = s->inuse; /* The end of info */
799
800 if (s->offset)
801 /* Freepointer is placed after the object. */
802 off += sizeof(void *);
803
804 if (s->flags & SLAB_STORE_USER)
805 /* We also have user information there */
806 off += 2 * sizeof(struct track);
807
808 if (s->size == off)
809 return 1;
810
811 return check_bytes_and_report(s, page, p, "Object padding",
812 p + off, POISON_INUSE, s->size - off);
813}
814
815/* Check the pad bytes at the end of a slab page */
816static int slab_pad_check(struct kmem_cache *s, struct page *page)
817{
818 u8 *start;
819 u8 *fault;
820 u8 *end;
821 int length;
822 int remainder;
823
824 if (!(s->flags & SLAB_POISON))
825 return 1;
826
827 start = page_address(page);
828 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
829 end = start + length;
830 remainder = length % s->size;
831 if (!remainder)
832 return 1;
833
834 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
835 if (!fault)
836 return 1;
837 while (end > fault && end[-1] == POISON_INUSE)
838 end--;
839
840 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
841 print_section("Padding", end - remainder, remainder);
842
843 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
844 return 0;
845}
846
847static int check_object(struct kmem_cache *s, struct page *page,
848 void *object, u8 val)
849{
850 u8 *p = object;
851 u8 *endobject = object + s->objsize;
852
853 if (s->flags & SLAB_RED_ZONE) {
854 if (!check_bytes_and_report(s, page, object, "Redzone",
855 endobject, val, s->inuse - s->objsize))
856 return 0;
857 } else {
858 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
859 check_bytes_and_report(s, page, p, "Alignment padding",
860 endobject, POISON_INUSE, s->inuse - s->objsize);
861 }
862 }
863
864 if (s->flags & SLAB_POISON) {
865 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
866 (!check_bytes_and_report(s, page, p, "Poison", p,
867 POISON_FREE, s->objsize - 1) ||
868 !check_bytes_and_report(s, page, p, "Poison",
869 p + s->objsize - 1, POISON_END, 1)))
870 return 0;
871 /*
872 * check_pad_bytes cleans up on its own.
873 */
874 check_pad_bytes(s, page, p);
875 }
876
877 if (!s->offset && val == SLUB_RED_ACTIVE)
878 /*
879 * Object and freepointer overlap. Cannot check
880 * freepointer while object is allocated.
881 */
882 return 1;
883
884 /* Check free pointer validity */
885 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
886 object_err(s, page, p, "Freepointer corrupt");
887 /*
888 * No choice but to zap it and thus lose the remainder
889 * of the free objects in this slab. May cause
890 * another error because the object count is now wrong.
891 */
892 set_freepointer(s, p, NULL);
893 return 0;
894 }
895 return 1;
896}
897
898static int check_slab(struct kmem_cache *s, struct page *page)
899{
900 int maxobj;
901
902 VM_BUG_ON(!irqs_disabled());
903
904 if (!PageSlab(page)) {
905 slab_err(s, page, "Not a valid slab page");
906 return 0;
907 }
908
909 maxobj = order_objects(compound_order(page), s->size, s->reserved);
910 if (page->objects > maxobj) {
911 slab_err(s, page, "objects %u > max %u",
912 s->name, page->objects, maxobj);
913 return 0;
914 }
915 if (page->inuse > page->objects) {
916 slab_err(s, page, "inuse %u > max %u",
917 s->name, page->inuse, page->objects);
918 return 0;
919 }
920 /* Slab_pad_check fixes things up after itself */
921 slab_pad_check(s, page);
922 return 1;
923}
924
925/*
926 * Determine if a certain object on a page is on the freelist. Must hold the
927 * slab lock to guarantee that the chains are in a consistent state.
928 */
929static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
930{
931 int nr = 0;
932 void *fp;
933 void *object = NULL;
934 unsigned long max_objects;
935
936 fp = page->freelist;
937 while (fp && nr <= page->objects) {
938 if (fp == search)
939 return 1;
940 if (!check_valid_pointer(s, page, fp)) {
941 if (object) {
942 object_err(s, page, object,
943 "Freechain corrupt");
944 set_freepointer(s, object, NULL);
945 break;
946 } else {
947 slab_err(s, page, "Freepointer corrupt");
948 page->freelist = NULL;
949 page->inuse = page->objects;
950 slab_fix(s, "Freelist cleared");
951 return 0;
952 }
953 break;
954 }
955 object = fp;
956 fp = get_freepointer(s, object);
957 nr++;
958 }
959
960 max_objects = order_objects(compound_order(page), s->size, s->reserved);
961 if (max_objects > MAX_OBJS_PER_PAGE)
962 max_objects = MAX_OBJS_PER_PAGE;
963
964 if (page->objects != max_objects) {
965 slab_err(s, page, "Wrong number of objects. Found %d but "
966 "should be %d", page->objects, max_objects);
967 page->objects = max_objects;
968 slab_fix(s, "Number of objects adjusted.");
969 }
970 if (page->inuse != page->objects - nr) {
971 slab_err(s, page, "Wrong object count. Counter is %d but "
972 "counted were %d", page->inuse, page->objects - nr);
973 page->inuse = page->objects - nr;
974 slab_fix(s, "Object count adjusted.");
975 }
976 return search == NULL;
977}
978
979static void trace(struct kmem_cache *s, struct page *page, void *object,
980 int alloc)
981{
982 if (s->flags & SLAB_TRACE) {
983 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
984 s->name,
985 alloc ? "alloc" : "free",
986 object, page->inuse,
987 page->freelist);
988
989 if (!alloc)
990 print_section("Object", (void *)object, s->objsize);
991
992 dump_stack();
993 }
994}
995
996/*
997 * Hooks for other subsystems that check memory allocations. In a typical
998 * production configuration these hooks all should produce no code at all.
999 */
1000static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1001{
1002 flags &= gfp_allowed_mask;
1003 lockdep_trace_alloc(flags);
1004 might_sleep_if(flags & __GFP_WAIT);
1005
1006 return should_failslab(s->objsize, flags, s->flags);
1007}
1008
1009static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
1010{
1011 flags &= gfp_allowed_mask;
1012 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
1013 kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, flags);
1014}
1015
1016static inline void slab_free_hook(struct kmem_cache *s, void *x)
1017{
1018 kmemleak_free_recursive(x, s->flags);
1019
1020 /*
1021 * Trouble is that we may no longer disable interupts in the fast path
1022 * So in order to make the debug calls that expect irqs to be
1023 * disabled we need to disable interrupts temporarily.
1024 */
1025#if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1026 {
1027 unsigned long flags;
1028
1029 local_irq_save(flags);
1030 kmemcheck_slab_free(s, x, s->objsize);
1031 debug_check_no_locks_freed(x, s->objsize);
1032 local_irq_restore(flags);
1033 }
1034#endif
1035 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1036 debug_check_no_obj_freed(x, s->objsize);
1037}
1038
1039/*
1040 * Tracking of fully allocated slabs for debugging purposes.
1041 *
1042 * list_lock must be held.
1043 */
1044static void add_full(struct kmem_cache *s,
1045 struct kmem_cache_node *n, struct page *page)
1046{
1047 if (!(s->flags & SLAB_STORE_USER))
1048 return;
1049
1050 list_add(&page->lru, &n->full);
1051}
1052
1053/*
1054 * list_lock must be held.
1055 */
1056static void remove_full(struct kmem_cache *s, struct page *page)
1057{
1058 if (!(s->flags & SLAB_STORE_USER))
1059 return;
1060
1061 list_del(&page->lru);
1062}
1063
1064/* Tracking of the number of slabs for debugging purposes */
1065static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1066{
1067 struct kmem_cache_node *n = get_node(s, node);
1068
1069 return atomic_long_read(&n->nr_slabs);
1070}
1071
1072static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1073{
1074 return atomic_long_read(&n->nr_slabs);
1075}
1076
1077static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1078{
1079 struct kmem_cache_node *n = get_node(s, node);
1080
1081 /*
1082 * May be called early in order to allocate a slab for the
1083 * kmem_cache_node structure. Solve the chicken-egg
1084 * dilemma by deferring the increment of the count during
1085 * bootstrap (see early_kmem_cache_node_alloc).
1086 */
1087 if (n) {
1088 atomic_long_inc(&n->nr_slabs);
1089 atomic_long_add(objects, &n->total_objects);
1090 }
1091}
1092static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1093{
1094 struct kmem_cache_node *n = get_node(s, node);
1095
1096 atomic_long_dec(&n->nr_slabs);
1097 atomic_long_sub(objects, &n->total_objects);
1098}
1099
1100/* Object debug checks for alloc/free paths */
1101static void setup_object_debug(struct kmem_cache *s, struct page *page,
1102 void *object)
1103{
1104 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1105 return;
1106
1107 init_object(s, object, SLUB_RED_INACTIVE);
1108 init_tracking(s, object);
1109}
1110
1111static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
1112 void *object, unsigned long addr)
1113{
1114 if (!check_slab(s, page))
1115 goto bad;
1116
1117 if (!check_valid_pointer(s, page, object)) {
1118 object_err(s, page, object, "Freelist Pointer check fails");
1119 goto bad;
1120 }
1121
1122 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1123 goto bad;
1124
1125 /* Success perform special debug activities for allocs */
1126 if (s->flags & SLAB_STORE_USER)
1127 set_track(s, object, TRACK_ALLOC, addr);
1128 trace(s, page, object, 1);
1129 init_object(s, object, SLUB_RED_ACTIVE);
1130 return 1;
1131
1132bad:
1133 if (PageSlab(page)) {
1134 /*
1135 * If this is a slab page then lets do the best we can
1136 * to avoid issues in the future. Marking all objects
1137 * as used avoids touching the remaining objects.
1138 */
1139 slab_fix(s, "Marking all objects used");
1140 page->inuse = page->objects;
1141 page->freelist = NULL;
1142 }
1143 return 0;
1144}
1145
1146static noinline int free_debug_processing(struct kmem_cache *s,
1147 struct page *page, void *object, unsigned long addr)
1148{
1149 unsigned long flags;
1150 int rc = 0;
1151
1152 local_irq_save(flags);
1153 slab_lock(page);
1154
1155 if (!check_slab(s, page))
1156 goto fail;
1157
1158 if (!check_valid_pointer(s, page, object)) {
1159 slab_err(s, page, "Invalid object pointer 0x%p", object);
1160 goto fail;
1161 }
1162
1163 if (on_freelist(s, page, object)) {
1164 object_err(s, page, object, "Object already free");
1165 goto fail;
1166 }
1167
1168 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1169 goto out;
1170
1171 if (unlikely(s != page->slab)) {
1172 if (!PageSlab(page)) {
1173 slab_err(s, page, "Attempt to free object(0x%p) "
1174 "outside of slab", object);
1175 } else if (!page->slab) {
1176 printk(KERN_ERR
1177 "SLUB <none>: no slab for object 0x%p.\n",
1178 object);
1179 dump_stack();
1180 } else
1181 object_err(s, page, object,
1182 "page slab pointer corrupt.");
1183 goto fail;
1184 }
1185
1186 if (s->flags & SLAB_STORE_USER)
1187 set_track(s, object, TRACK_FREE, addr);
1188 trace(s, page, object, 0);
1189 init_object(s, object, SLUB_RED_INACTIVE);
1190 rc = 1;
1191out:
1192 slab_unlock(page);
1193 local_irq_restore(flags);
1194 return rc;
1195
1196fail:
1197 slab_fix(s, "Object at 0x%p not freed", object);
1198 goto out;
1199}
1200
1201static int __init setup_slub_debug(char *str)
1202{
1203 slub_debug = DEBUG_DEFAULT_FLAGS;
1204 if (*str++ != '=' || !*str)
1205 /*
1206 * No options specified. Switch on full debugging.
1207 */
1208 goto out;
1209
1210 if (*str == ',')
1211 /*
1212 * No options but restriction on slabs. This means full
1213 * debugging for slabs matching a pattern.
1214 */
1215 goto check_slabs;
1216
1217 if (tolower(*str) == 'o') {
1218 /*
1219 * Avoid enabling debugging on caches if its minimum order
1220 * would increase as a result.
1221 */
1222 disable_higher_order_debug = 1;
1223 goto out;
1224 }
1225
1226 slub_debug = 0;
1227 if (*str == '-')
1228 /*
1229 * Switch off all debugging measures.
1230 */
1231 goto out;
1232
1233 /*
1234 * Determine which debug features should be switched on
1235 */
1236 for (; *str && *str != ','; str++) {
1237 switch (tolower(*str)) {
1238 case 'f':
1239 slub_debug |= SLAB_DEBUG_FREE;
1240 break;
1241 case 'z':
1242 slub_debug |= SLAB_RED_ZONE;
1243 break;
1244 case 'p':
1245 slub_debug |= SLAB_POISON;
1246 break;
1247 case 'u':
1248 slub_debug |= SLAB_STORE_USER;
1249 break;
1250 case 't':
1251 slub_debug |= SLAB_TRACE;
1252 break;
1253 case 'a':
1254 slub_debug |= SLAB_FAILSLAB;
1255 break;
1256 default:
1257 printk(KERN_ERR "slub_debug option '%c' "
1258 "unknown. skipped\n", *str);
1259 }
1260 }
1261
1262check_slabs:
1263 if (*str == ',')
1264 slub_debug_slabs = str + 1;
1265out:
1266 return 1;
1267}
1268
1269__setup("slub_debug", setup_slub_debug);
1270
1271static unsigned long kmem_cache_flags(unsigned long objsize,
1272 unsigned long flags, const char *name,
1273 void (*ctor)(void *))
1274{
1275 /*
1276 * Enable debugging if selected on the kernel commandline.
1277 */
1278 if (slub_debug && (!slub_debug_slabs ||
1279 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1280 flags |= slub_debug;
1281
1282 return flags;
1283}
1284#else
1285static inline void setup_object_debug(struct kmem_cache *s,
1286 struct page *page, void *object) {}
1287
1288static inline int alloc_debug_processing(struct kmem_cache *s,
1289 struct page *page, void *object, unsigned long addr) { return 0; }
1290
1291static inline int free_debug_processing(struct kmem_cache *s,
1292 struct page *page, void *object, unsigned long addr) { return 0; }
1293
1294static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1295 { return 1; }
1296static inline int check_object(struct kmem_cache *s, struct page *page,
1297 void *object, u8 val) { return 1; }
1298static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1299 struct page *page) {}
1300static inline void remove_full(struct kmem_cache *s, struct page *page) {}
1301static inline unsigned long kmem_cache_flags(unsigned long objsize,
1302 unsigned long flags, const char *name,
1303 void (*ctor)(void *))
1304{
1305 return flags;
1306}
1307#define slub_debug 0
1308
1309#define disable_higher_order_debug 0
1310
1311static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1312 { return 0; }
1313static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1314 { return 0; }
1315static inline void inc_slabs_node(struct kmem_cache *s, int node,
1316 int objects) {}
1317static inline void dec_slabs_node(struct kmem_cache *s, int node,
1318 int objects) {}
1319
1320static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1321 { return 0; }
1322
1323static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1324 void *object) {}
1325
1326static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1327
1328#endif /* CONFIG_SLUB_DEBUG */
1329
1330/*
1331 * Slab allocation and freeing
1332 */
1333static inline struct page *alloc_slab_page(gfp_t flags, int node,
1334 struct kmem_cache_order_objects oo)
1335{
1336 int order = oo_order(oo);
1337
1338 flags |= __GFP_NOTRACK;
1339
1340 if (node == NUMA_NO_NODE)
1341 return alloc_pages(flags, order);
1342 else
1343 return alloc_pages_exact_node(node, flags, order);
1344}
1345
1346static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1347{
1348 struct page *page;
1349 struct kmem_cache_order_objects oo = s->oo;
1350 gfp_t alloc_gfp;
1351
1352 flags &= gfp_allowed_mask;
1353
1354 if (flags & __GFP_WAIT)
1355 local_irq_enable();
1356
1357 flags |= s->allocflags;
1358
1359 /*
1360 * Let the initial higher-order allocation fail under memory pressure
1361 * so we fall-back to the minimum order allocation.
1362 */
1363 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1364
1365 page = alloc_slab_page(alloc_gfp, node, oo);
1366 if (unlikely(!page)) {
1367 oo = s->min;
1368 /*
1369 * Allocation may have failed due to fragmentation.
1370 * Try a lower order alloc if possible
1371 */
1372 page = alloc_slab_page(flags, node, oo);
1373
1374 if (page)
1375 stat(s, ORDER_FALLBACK);
1376 }
1377
1378 if (flags & __GFP_WAIT)
1379 local_irq_disable();
1380
1381 if (!page)
1382 return NULL;
1383
1384 if (kmemcheck_enabled
1385 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1386 int pages = 1 << oo_order(oo);
1387
1388 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1389
1390 /*
1391 * Objects from caches that have a constructor don't get
1392 * cleared when they're allocated, so we need to do it here.
1393 */
1394 if (s->ctor)
1395 kmemcheck_mark_uninitialized_pages(page, pages);
1396 else
1397 kmemcheck_mark_unallocated_pages(page, pages);
1398 }
1399
1400 page->objects = oo_objects(oo);
1401 mod_zone_page_state(page_zone(page),
1402 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1403 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1404 1 << oo_order(oo));
1405
1406 return page;
1407}
1408
1409static void setup_object(struct kmem_cache *s, struct page *page,
1410 void *object)
1411{
1412 setup_object_debug(s, page, object);
1413 if (unlikely(s->ctor))
1414 s->ctor(object);
1415}
1416
1417static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1418{
1419 struct page *page;
1420 void *start;
1421 void *last;
1422 void *p;
1423
1424 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1425
1426 page = allocate_slab(s,
1427 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1428 if (!page)
1429 goto out;
1430
1431 inc_slabs_node(s, page_to_nid(page), page->objects);
1432 page->slab = s;
1433 page->flags |= 1 << PG_slab;
1434
1435 start = page_address(page);
1436
1437 if (unlikely(s->flags & SLAB_POISON))
1438 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1439
1440 last = start;
1441 for_each_object(p, s, start, page->objects) {
1442 setup_object(s, page, last);
1443 set_freepointer(s, last, p);
1444 last = p;
1445 }
1446 setup_object(s, page, last);
1447 set_freepointer(s, last, NULL);
1448
1449 page->freelist = start;
1450 page->inuse = 0;
1451 page->frozen = 1;
1452out:
1453 return page;
1454}
1455
1456static void __free_slab(struct kmem_cache *s, struct page *page)
1457{
1458 int order = compound_order(page);
1459 int pages = 1 << order;
1460
1461 if (kmem_cache_debug(s)) {
1462 void *p;
1463
1464 slab_pad_check(s, page);
1465 for_each_object(p, s, page_address(page),
1466 page->objects)
1467 check_object(s, page, p, SLUB_RED_INACTIVE);
1468 }
1469
1470 kmemcheck_free_shadow(page, compound_order(page));
1471
1472 mod_zone_page_state(page_zone(page),
1473 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1474 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1475 -pages);
1476
1477 __ClearPageSlab(page);
1478 reset_page_mapcount(page);
1479 if (current->reclaim_state)
1480 current->reclaim_state->reclaimed_slab += pages;
1481 __free_pages(page, order);
1482}
1483
1484#define need_reserve_slab_rcu \
1485 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1486
1487static void rcu_free_slab(struct rcu_head *h)
1488{
1489 struct page *page;
1490
1491 if (need_reserve_slab_rcu)
1492 page = virt_to_head_page(h);
1493 else
1494 page = container_of((struct list_head *)h, struct page, lru);
1495
1496 __free_slab(page->slab, page);
1497}
1498
1499static void free_slab(struct kmem_cache *s, struct page *page)
1500{
1501 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1502 struct rcu_head *head;
1503
1504 if (need_reserve_slab_rcu) {
1505 int order = compound_order(page);
1506 int offset = (PAGE_SIZE << order) - s->reserved;
1507
1508 VM_BUG_ON(s->reserved != sizeof(*head));
1509 head = page_address(page) + offset;
1510 } else {
1511 /*
1512 * RCU free overloads the RCU head over the LRU
1513 */
1514 head = (void *)&page->lru;
1515 }
1516
1517 call_rcu(head, rcu_free_slab);
1518 } else
1519 __free_slab(s, page);
1520}
1521
1522static void discard_slab(struct kmem_cache *s, struct page *page)
1523{
1524 dec_slabs_node(s, page_to_nid(page), page->objects);
1525 free_slab(s, page);
1526}
1527
1528/*
1529 * Management of partially allocated slabs.
1530 *
1531 * list_lock must be held.
1532 */
1533static inline void add_partial(struct kmem_cache_node *n,
1534 struct page *page, int tail)
1535{
1536 n->nr_partial++;
1537 if (tail)
1538 list_add_tail(&page->lru, &n->partial);
1539 else
1540 list_add(&page->lru, &n->partial);
1541}
1542
1543/*
1544 * list_lock must be held.
1545 */
1546static inline void remove_partial(struct kmem_cache_node *n,
1547 struct page *page)
1548{
1549 list_del(&page->lru);
1550 n->nr_partial--;
1551}
1552
1553/*
1554 * Lock slab, remove from the partial list and put the object into the
1555 * per cpu freelist.
1556 *
1557 * Must hold list_lock.
1558 */
1559static inline int acquire_slab(struct kmem_cache *s,
1560 struct kmem_cache_node *n, struct page *page)
1561{
1562 void *freelist;
1563 unsigned long counters;
1564 struct page new;
1565
1566 /*
1567 * Zap the freelist and set the frozen bit.
1568 * The old freelist is the list of objects for the
1569 * per cpu allocation list.
1570 */
1571 do {
1572 freelist = page->freelist;
1573 counters = page->counters;
1574 new.counters = counters;
1575 new.inuse = page->objects;
1576
1577 VM_BUG_ON(new.frozen);
1578 new.frozen = 1;
1579
1580 } while (!__cmpxchg_double_slab(s, page,
1581 freelist, counters,
1582 NULL, new.counters,
1583 "lock and freeze"));
1584
1585 remove_partial(n, page);
1586
1587 if (freelist) {
1588 /* Populate the per cpu freelist */
1589 this_cpu_write(s->cpu_slab->freelist, freelist);
1590 this_cpu_write(s->cpu_slab->page, page);
1591 this_cpu_write(s->cpu_slab->node, page_to_nid(page));
1592 return 1;
1593 } else {
1594 /*
1595 * Slab page came from the wrong list. No object to allocate
1596 * from. Put it onto the correct list and continue partial
1597 * scan.
1598 */
1599 printk(KERN_ERR "SLUB: %s : Page without available objects on"
1600 " partial list\n", s->name);
1601 return 0;
1602 }
1603}
1604
1605/*
1606 * Try to allocate a partial slab from a specific node.
1607 */
1608static struct page *get_partial_node(struct kmem_cache *s,
1609 struct kmem_cache_node *n)
1610{
1611 struct page *page;
1612
1613 /*
1614 * Racy check. If we mistakenly see no partial slabs then we
1615 * just allocate an empty slab. If we mistakenly try to get a
1616 * partial slab and there is none available then get_partials()
1617 * will return NULL.
1618 */
1619 if (!n || !n->nr_partial)
1620 return NULL;
1621
1622 spin_lock(&n->list_lock);
1623 list_for_each_entry(page, &n->partial, lru)
1624 if (acquire_slab(s, n, page))
1625 goto out;
1626 page = NULL;
1627out:
1628 spin_unlock(&n->list_lock);
1629 return page;
1630}
1631
1632/*
1633 * Get a page from somewhere. Search in increasing NUMA distances.
1634 */
1635static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1636{
1637#ifdef CONFIG_NUMA
1638 struct zonelist *zonelist;
1639 struct zoneref *z;
1640 struct zone *zone;
1641 enum zone_type high_zoneidx = gfp_zone(flags);
1642 struct page *page;
1643
1644 /*
1645 * The defrag ratio allows a configuration of the tradeoffs between
1646 * inter node defragmentation and node local allocations. A lower
1647 * defrag_ratio increases the tendency to do local allocations
1648 * instead of attempting to obtain partial slabs from other nodes.
1649 *
1650 * If the defrag_ratio is set to 0 then kmalloc() always
1651 * returns node local objects. If the ratio is higher then kmalloc()
1652 * may return off node objects because partial slabs are obtained
1653 * from other nodes and filled up.
1654 *
1655 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1656 * defrag_ratio = 1000) then every (well almost) allocation will
1657 * first attempt to defrag slab caches on other nodes. This means
1658 * scanning over all nodes to look for partial slabs which may be
1659 * expensive if we do it every time we are trying to find a slab
1660 * with available objects.
1661 */
1662 if (!s->remote_node_defrag_ratio ||
1663 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1664 return NULL;
1665
1666 get_mems_allowed();
1667 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1668 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1669 struct kmem_cache_node *n;
1670
1671 n = get_node(s, zone_to_nid(zone));
1672
1673 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1674 n->nr_partial > s->min_partial) {
1675 page = get_partial_node(s, n);
1676 if (page) {
1677 put_mems_allowed();
1678 return page;
1679 }
1680 }
1681 }
1682 put_mems_allowed();
1683#endif
1684 return NULL;
1685}
1686
1687/*
1688 * Get a partial page, lock it and return it.
1689 */
1690static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1691{
1692 struct page *page;
1693 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1694
1695 page = get_partial_node(s, get_node(s, searchnode));
1696 if (page || node != NUMA_NO_NODE)
1697 return page;
1698
1699 return get_any_partial(s, flags);
1700}
1701
1702#ifdef CONFIG_PREEMPT
1703/*
1704 * Calculate the next globally unique transaction for disambiguiation
1705 * during cmpxchg. The transactions start with the cpu number and are then
1706 * incremented by CONFIG_NR_CPUS.
1707 */
1708#define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1709#else
1710/*
1711 * No preemption supported therefore also no need to check for
1712 * different cpus.
1713 */
1714#define TID_STEP 1
1715#endif
1716
1717static inline unsigned long next_tid(unsigned long tid)
1718{
1719 return tid + TID_STEP;
1720}
1721
1722static inline unsigned int tid_to_cpu(unsigned long tid)
1723{
1724 return tid % TID_STEP;
1725}
1726
1727static inline unsigned long tid_to_event(unsigned long tid)
1728{
1729 return tid / TID_STEP;
1730}
1731
1732static inline unsigned int init_tid(int cpu)
1733{
1734 return cpu;
1735}
1736
1737static inline void note_cmpxchg_failure(const char *n,
1738 const struct kmem_cache *s, unsigned long tid)
1739{
1740#ifdef SLUB_DEBUG_CMPXCHG
1741 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1742
1743 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1744
1745#ifdef CONFIG_PREEMPT
1746 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1747 printk("due to cpu change %d -> %d\n",
1748 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1749 else
1750#endif
1751 if (tid_to_event(tid) != tid_to_event(actual_tid))
1752 printk("due to cpu running other code. Event %ld->%ld\n",
1753 tid_to_event(tid), tid_to_event(actual_tid));
1754 else
1755 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1756 actual_tid, tid, next_tid(tid));
1757#endif
1758 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1759}
1760
1761void init_kmem_cache_cpus(struct kmem_cache *s)
1762{
1763 int cpu;
1764
1765 for_each_possible_cpu(cpu)
1766 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1767}
1768/*
1769 * Remove the cpu slab
1770 */
1771
1772/*
1773 * Remove the cpu slab
1774 */
1775static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1776{
1777 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1778 struct page *page = c->page;
1779 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1780 int lock = 0;
1781 enum slab_modes l = M_NONE, m = M_NONE;
1782 void *freelist;
1783 void *nextfree;
1784 int tail = 0;
1785 struct page new;
1786 struct page old;
1787
1788 if (page->freelist) {
1789 stat(s, DEACTIVATE_REMOTE_FREES);
1790 tail = 1;
1791 }
1792
1793 c->tid = next_tid(c->tid);
1794 c->page = NULL;
1795 freelist = c->freelist;
1796 c->freelist = NULL;
1797
1798 /*
1799 * Stage one: Free all available per cpu objects back
1800 * to the page freelist while it is still frozen. Leave the
1801 * last one.
1802 *
1803 * There is no need to take the list->lock because the page
1804 * is still frozen.
1805 */
1806 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1807 void *prior;
1808 unsigned long counters;
1809
1810 do {
1811 prior = page->freelist;
1812 counters = page->counters;
1813 set_freepointer(s, freelist, prior);
1814 new.counters = counters;
1815 new.inuse--;
1816 VM_BUG_ON(!new.frozen);
1817
1818 } while (!__cmpxchg_double_slab(s, page,
1819 prior, counters,
1820 freelist, new.counters,
1821 "drain percpu freelist"));
1822
1823 freelist = nextfree;
1824 }
1825
1826 /*
1827 * Stage two: Ensure that the page is unfrozen while the
1828 * list presence reflects the actual number of objects
1829 * during unfreeze.
1830 *
1831 * We setup the list membership and then perform a cmpxchg
1832 * with the count. If there is a mismatch then the page
1833 * is not unfrozen but the page is on the wrong list.
1834 *
1835 * Then we restart the process which may have to remove
1836 * the page from the list that we just put it on again
1837 * because the number of objects in the slab may have
1838 * changed.
1839 */
1840redo:
1841
1842 old.freelist = page->freelist;
1843 old.counters = page->counters;
1844 VM_BUG_ON(!old.frozen);
1845
1846 /* Determine target state of the slab */
1847 new.counters = old.counters;
1848 if (freelist) {
1849 new.inuse--;
1850 set_freepointer(s, freelist, old.freelist);
1851 new.freelist = freelist;
1852 } else
1853 new.freelist = old.freelist;
1854
1855 new.frozen = 0;
1856
1857 if (!new.inuse && n->nr_partial > s->min_partial)
1858 m = M_FREE;
1859 else if (new.freelist) {
1860 m = M_PARTIAL;
1861 if (!lock) {
1862 lock = 1;
1863 /*
1864 * Taking the spinlock removes the possiblity
1865 * that acquire_slab() will see a slab page that
1866 * is frozen
1867 */
1868 spin_lock(&n->list_lock);
1869 }
1870 } else {
1871 m = M_FULL;
1872 if (kmem_cache_debug(s) && !lock) {
1873 lock = 1;
1874 /*
1875 * This also ensures that the scanning of full
1876 * slabs from diagnostic functions will not see
1877 * any frozen slabs.
1878 */
1879 spin_lock(&n->list_lock);
1880 }
1881 }
1882
1883 if (l != m) {
1884
1885 if (l == M_PARTIAL)
1886
1887 remove_partial(n, page);
1888
1889 else if (l == M_FULL)
1890
1891 remove_full(s, page);
1892
1893 if (m == M_PARTIAL) {
1894
1895 add_partial(n, page, tail);
1896 stat(s, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1897
1898 } else if (m == M_FULL) {
1899
1900 stat(s, DEACTIVATE_FULL);
1901 add_full(s, n, page);
1902
1903 }
1904 }
1905
1906 l = m;
1907 if (!__cmpxchg_double_slab(s, page,
1908 old.freelist, old.counters,
1909 new.freelist, new.counters,
1910 "unfreezing slab"))
1911 goto redo;
1912
1913 if (lock)
1914 spin_unlock(&n->list_lock);
1915
1916 if (m == M_FREE) {
1917 stat(s, DEACTIVATE_EMPTY);
1918 discard_slab(s, page);
1919 stat(s, FREE_SLAB);
1920 }
1921}
1922
1923static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1924{
1925 stat(s, CPUSLAB_FLUSH);
1926 deactivate_slab(s, c);
1927}
1928
1929/*
1930 * Flush cpu slab.
1931 *
1932 * Called from IPI handler with interrupts disabled.
1933 */
1934static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1935{
1936 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
1937
1938 if (likely(c && c->page))
1939 flush_slab(s, c);
1940}
1941
1942static void flush_cpu_slab(void *d)
1943{
1944 struct kmem_cache *s = d;
1945
1946 __flush_cpu_slab(s, smp_processor_id());
1947}
1948
1949static void flush_all(struct kmem_cache *s)
1950{
1951 on_each_cpu(flush_cpu_slab, s, 1);
1952}
1953
1954/*
1955 * Check if the objects in a per cpu structure fit numa
1956 * locality expectations.
1957 */
1958static inline int node_match(struct kmem_cache_cpu *c, int node)
1959{
1960#ifdef CONFIG_NUMA
1961 if (node != NUMA_NO_NODE && c->node != node)
1962 return 0;
1963#endif
1964 return 1;
1965}
1966
1967static int count_free(struct page *page)
1968{
1969 return page->objects - page->inuse;
1970}
1971
1972static unsigned long count_partial(struct kmem_cache_node *n,
1973 int (*get_count)(struct page *))
1974{
1975 unsigned long flags;
1976 unsigned long x = 0;
1977 struct page *page;
1978
1979 spin_lock_irqsave(&n->list_lock, flags);
1980 list_for_each_entry(page, &n->partial, lru)
1981 x += get_count(page);
1982 spin_unlock_irqrestore(&n->list_lock, flags);
1983 return x;
1984}
1985
1986static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
1987{
1988#ifdef CONFIG_SLUB_DEBUG
1989 return atomic_long_read(&n->total_objects);
1990#else
1991 return 0;
1992#endif
1993}
1994
1995static noinline void
1996slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
1997{
1998 int node;
1999
2000 printk(KERN_WARNING
2001 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2002 nid, gfpflags);
2003 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
2004 "default order: %d, min order: %d\n", s->name, s->objsize,
2005 s->size, oo_order(s->oo), oo_order(s->min));
2006
2007 if (oo_order(s->min) > get_order(s->objsize))
2008 printk(KERN_WARNING " %s debugging increased min order, use "
2009 "slub_debug=O to disable.\n", s->name);
2010
2011 for_each_online_node(node) {
2012 struct kmem_cache_node *n = get_node(s, node);
2013 unsigned long nr_slabs;
2014 unsigned long nr_objs;
2015 unsigned long nr_free;
2016
2017 if (!n)
2018 continue;
2019
2020 nr_free = count_partial(n, count_free);
2021 nr_slabs = node_nr_slabs(n);
2022 nr_objs = node_nr_objs(n);
2023
2024 printk(KERN_WARNING
2025 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2026 node, nr_slabs, nr_objs, nr_free);
2027 }
2028}
2029
2030/*
2031 * Slow path. The lockless freelist is empty or we need to perform
2032 * debugging duties.
2033 *
2034 * Interrupts are disabled.
2035 *
2036 * Processing is still very fast if new objects have been freed to the
2037 * regular freelist. In that case we simply take over the regular freelist
2038 * as the lockless freelist and zap the regular freelist.
2039 *
2040 * If that is not working then we fall back to the partial lists. We take the
2041 * first element of the freelist as the object to allocate now and move the
2042 * rest of the freelist to the lockless freelist.
2043 *
2044 * And if we were unable to get a new slab from the partial slab lists then
2045 * we need to allocate a new slab. This is the slowest path since it involves
2046 * a call to the page allocator and the setup of a new slab.
2047 */
2048static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2049 unsigned long addr, struct kmem_cache_cpu *c)
2050{
2051 void **object;
2052 struct page *page;
2053 unsigned long flags;
2054 struct page new;
2055 unsigned long counters;
2056
2057 local_irq_save(flags);
2058#ifdef CONFIG_PREEMPT
2059 /*
2060 * We may have been preempted and rescheduled on a different
2061 * cpu before disabling interrupts. Need to reload cpu area
2062 * pointer.
2063 */
2064 c = this_cpu_ptr(s->cpu_slab);
2065#endif
2066
2067 /* We handle __GFP_ZERO in the caller */
2068 gfpflags &= ~__GFP_ZERO;
2069
2070 page = c->page;
2071 if (!page)
2072 goto new_slab;
2073
2074 if (unlikely(!node_match(c, node))) {
2075 stat(s, ALLOC_NODE_MISMATCH);
2076 deactivate_slab(s, c);
2077 goto new_slab;
2078 }
2079
2080 stat(s, ALLOC_SLOWPATH);
2081
2082 do {
2083 object = page->freelist;
2084 counters = page->counters;
2085 new.counters = counters;
2086 VM_BUG_ON(!new.frozen);
2087
2088 /*
2089 * If there is no object left then we use this loop to
2090 * deactivate the slab which is simple since no objects
2091 * are left in the slab and therefore we do not need to
2092 * put the page back onto the partial list.
2093 *
2094 * If there are objects left then we retrieve them
2095 * and use them to refill the per cpu queue.
2096 */
2097
2098 new.inuse = page->objects;
2099 new.frozen = object != NULL;
2100
2101 } while (!__cmpxchg_double_slab(s, page,
2102 object, counters,
2103 NULL, new.counters,
2104 "__slab_alloc"));
2105
2106 if (unlikely(!object)) {
2107 c->page = NULL;
2108 stat(s, DEACTIVATE_BYPASS);
2109 goto new_slab;
2110 }
2111
2112 stat(s, ALLOC_REFILL);
2113
2114load_freelist:
2115 VM_BUG_ON(!page->frozen);
2116 c->freelist = get_freepointer(s, object);
2117 c->tid = next_tid(c->tid);
2118 local_irq_restore(flags);
2119 return object;
2120
2121new_slab:
2122 page = get_partial(s, gfpflags, node);
2123 if (page) {
2124 stat(s, ALLOC_FROM_PARTIAL);
2125 object = c->freelist;
2126
2127 if (kmem_cache_debug(s))
2128 goto debug;
2129 goto load_freelist;
2130 }
2131
2132 page = new_slab(s, gfpflags, node);
2133
2134 if (page) {
2135 c = __this_cpu_ptr(s->cpu_slab);
2136 if (c->page)
2137 flush_slab(s, c);
2138
2139 /*
2140 * No other reference to the page yet so we can
2141 * muck around with it freely without cmpxchg
2142 */
2143 object = page->freelist;
2144 page->freelist = NULL;
2145 page->inuse = page->objects;
2146
2147 stat(s, ALLOC_SLAB);
2148 c->node = page_to_nid(page);
2149 c->page = page;
2150
2151 if (kmem_cache_debug(s))
2152 goto debug;
2153 goto load_freelist;
2154 }
2155 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2156 slab_out_of_memory(s, gfpflags, node);
2157 local_irq_restore(flags);
2158 return NULL;
2159
2160debug:
2161 if (!object || !alloc_debug_processing(s, page, object, addr))
2162 goto new_slab;
2163
2164 c->freelist = get_freepointer(s, object);
2165 deactivate_slab(s, c);
2166 c->page = NULL;
2167 c->node = NUMA_NO_NODE;
2168 local_irq_restore(flags);
2169 return object;
2170}
2171
2172/*
2173 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2174 * have the fastpath folded into their functions. So no function call
2175 * overhead for requests that can be satisfied on the fastpath.
2176 *
2177 * The fastpath works by first checking if the lockless freelist can be used.
2178 * If not then __slab_alloc is called for slow processing.
2179 *
2180 * Otherwise we can simply pick the next object from the lockless free list.
2181 */
2182static __always_inline void *slab_alloc(struct kmem_cache *s,
2183 gfp_t gfpflags, int node, unsigned long addr)
2184{
2185 void **object;
2186 struct kmem_cache_cpu *c;
2187 unsigned long tid;
2188
2189 if (slab_pre_alloc_hook(s, gfpflags))
2190 return NULL;
2191
2192redo:
2193
2194 /*
2195 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2196 * enabled. We may switch back and forth between cpus while
2197 * reading from one cpu area. That does not matter as long
2198 * as we end up on the original cpu again when doing the cmpxchg.
2199 */
2200 c = __this_cpu_ptr(s->cpu_slab);
2201
2202 /*
2203 * The transaction ids are globally unique per cpu and per operation on
2204 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2205 * occurs on the right processor and that there was no operation on the
2206 * linked list in between.
2207 */
2208 tid = c->tid;
2209 barrier();
2210
2211 object = c->freelist;
2212 if (unlikely(!object || !node_match(c, node)))
2213
2214 object = __slab_alloc(s, gfpflags, node, addr, c);
2215
2216 else {
2217 /*
2218 * The cmpxchg will only match if there was no additional
2219 * operation and if we are on the right processor.
2220 *
2221 * The cmpxchg does the following atomically (without lock semantics!)
2222 * 1. Relocate first pointer to the current per cpu area.
2223 * 2. Verify that tid and freelist have not been changed
2224 * 3. If they were not changed replace tid and freelist
2225 *
2226 * Since this is without lock semantics the protection is only against
2227 * code executing on this cpu *not* from access by other cpus.
2228 */
2229 if (unlikely(!irqsafe_cpu_cmpxchg_double(
2230 s->cpu_slab->freelist, s->cpu_slab->tid,
2231 object, tid,
2232 get_freepointer_safe(s, object), next_tid(tid)))) {
2233
2234 note_cmpxchg_failure("slab_alloc", s, tid);
2235 goto redo;
2236 }
2237 stat(s, ALLOC_FASTPATH);
2238 }
2239
2240 if (unlikely(gfpflags & __GFP_ZERO) && object)
2241 memset(object, 0, s->objsize);
2242
2243 slab_post_alloc_hook(s, gfpflags, object);
2244
2245 return object;
2246}
2247
2248void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2249{
2250 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2251
2252 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
2253
2254 return ret;
2255}
2256EXPORT_SYMBOL(kmem_cache_alloc);
2257
2258#ifdef CONFIG_TRACING
2259void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2260{
2261 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2262 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2263 return ret;
2264}
2265EXPORT_SYMBOL(kmem_cache_alloc_trace);
2266
2267void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2268{
2269 void *ret = kmalloc_order(size, flags, order);
2270 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2271 return ret;
2272}
2273EXPORT_SYMBOL(kmalloc_order_trace);
2274#endif
2275
2276#ifdef CONFIG_NUMA
2277void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2278{
2279 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2280
2281 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2282 s->objsize, s->size, gfpflags, node);
2283
2284 return ret;
2285}
2286EXPORT_SYMBOL(kmem_cache_alloc_node);
2287
2288#ifdef CONFIG_TRACING
2289void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2290 gfp_t gfpflags,
2291 int node, size_t size)
2292{
2293 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2294
2295 trace_kmalloc_node(_RET_IP_, ret,
2296 size, s->size, gfpflags, node);
2297 return ret;
2298}
2299EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2300#endif
2301#endif
2302
2303/*
2304 * Slow patch handling. This may still be called frequently since objects
2305 * have a longer lifetime than the cpu slabs in most processing loads.
2306 *
2307 * So we still attempt to reduce cache line usage. Just take the slab
2308 * lock and free the item. If there is no additional partial page
2309 * handling required then we can return immediately.
2310 */
2311static void __slab_free(struct kmem_cache *s, struct page *page,
2312 void *x, unsigned long addr)
2313{
2314 void *prior;
2315 void **object = (void *)x;
2316 int was_frozen;
2317 int inuse;
2318 struct page new;
2319 unsigned long counters;
2320 struct kmem_cache_node *n = NULL;
2321 unsigned long uninitialized_var(flags);
2322
2323 stat(s, FREE_SLOWPATH);
2324
2325 if (kmem_cache_debug(s) && !free_debug_processing(s, page, x, addr))
2326 return;
2327
2328 do {
2329 prior = page->freelist;
2330 counters = page->counters;
2331 set_freepointer(s, object, prior);
2332 new.counters = counters;
2333 was_frozen = new.frozen;
2334 new.inuse--;
2335 if ((!new.inuse || !prior) && !was_frozen && !n) {
2336 n = get_node(s, page_to_nid(page));
2337 /*
2338 * Speculatively acquire the list_lock.
2339 * If the cmpxchg does not succeed then we may
2340 * drop the list_lock without any processing.
2341 *
2342 * Otherwise the list_lock will synchronize with
2343 * other processors updating the list of slabs.
2344 */
2345 spin_lock_irqsave(&n->list_lock, flags);
2346 }
2347 inuse = new.inuse;
2348
2349 } while (!cmpxchg_double_slab(s, page,
2350 prior, counters,
2351 object, new.counters,
2352 "__slab_free"));
2353
2354 if (likely(!n)) {
2355 /*
2356 * The list lock was not taken therefore no list
2357 * activity can be necessary.
2358 */
2359 if (was_frozen)
2360 stat(s, FREE_FROZEN);
2361 return;
2362 }
2363
2364 /*
2365 * was_frozen may have been set after we acquired the list_lock in
2366 * an earlier loop. So we need to check it here again.
2367 */
2368 if (was_frozen)
2369 stat(s, FREE_FROZEN);
2370 else {
2371 if (unlikely(!inuse && n->nr_partial > s->min_partial))
2372 goto slab_empty;
2373
2374 /*
2375 * Objects left in the slab. If it was not on the partial list before
2376 * then add it.
2377 */
2378 if (unlikely(!prior)) {
2379 remove_full(s, page);
2380 add_partial(n, page, 1);
2381 stat(s, FREE_ADD_PARTIAL);
2382 }
2383 }
2384 spin_unlock_irqrestore(&n->list_lock, flags);
2385 return;
2386
2387slab_empty:
2388 if (prior) {
2389 /*
2390 * Slab on the partial list.
2391 */
2392 remove_partial(n, page);
2393 stat(s, FREE_REMOVE_PARTIAL);
2394 } else
2395 /* Slab must be on the full list */
2396 remove_full(s, page);
2397
2398 spin_unlock_irqrestore(&n->list_lock, flags);
2399 stat(s, FREE_SLAB);
2400 discard_slab(s, page);
2401}
2402
2403/*
2404 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2405 * can perform fastpath freeing without additional function calls.
2406 *
2407 * The fastpath is only possible if we are freeing to the current cpu slab
2408 * of this processor. This typically the case if we have just allocated
2409 * the item before.
2410 *
2411 * If fastpath is not possible then fall back to __slab_free where we deal
2412 * with all sorts of special processing.
2413 */
2414static __always_inline void slab_free(struct kmem_cache *s,
2415 struct page *page, void *x, unsigned long addr)
2416{
2417 void **object = (void *)x;
2418 struct kmem_cache_cpu *c;
2419 unsigned long tid;
2420
2421 slab_free_hook(s, x);
2422
2423redo:
2424
2425 /*
2426 * Determine the currently cpus per cpu slab.
2427 * The cpu may change afterward. However that does not matter since
2428 * data is retrieved via this pointer. If we are on the same cpu
2429 * during the cmpxchg then the free will succedd.
2430 */
2431 c = __this_cpu_ptr(s->cpu_slab);
2432
2433 tid = c->tid;
2434 barrier();
2435
2436 if (likely(page == c->page)) {
2437 set_freepointer(s, object, c->freelist);
2438
2439 if (unlikely(!irqsafe_cpu_cmpxchg_double(
2440 s->cpu_slab->freelist, s->cpu_slab->tid,
2441 c->freelist, tid,
2442 object, next_tid(tid)))) {
2443
2444 note_cmpxchg_failure("slab_free", s, tid);
2445 goto redo;
2446 }
2447 stat(s, FREE_FASTPATH);
2448 } else
2449 __slab_free(s, page, x, addr);
2450
2451}
2452
2453void kmem_cache_free(struct kmem_cache *s, void *x)
2454{
2455 struct page *page;
2456
2457 page = virt_to_head_page(x);
2458
2459 slab_free(s, page, x, _RET_IP_);
2460
2461 trace_kmem_cache_free(_RET_IP_, x);
2462}
2463EXPORT_SYMBOL(kmem_cache_free);
2464
2465/*
2466 * Object placement in a slab is made very easy because we always start at
2467 * offset 0. If we tune the size of the object to the alignment then we can
2468 * get the required alignment by putting one properly sized object after
2469 * another.
2470 *
2471 * Notice that the allocation order determines the sizes of the per cpu
2472 * caches. Each processor has always one slab available for allocations.
2473 * Increasing the allocation order reduces the number of times that slabs
2474 * must be moved on and off the partial lists and is therefore a factor in
2475 * locking overhead.
2476 */
2477
2478/*
2479 * Mininum / Maximum order of slab pages. This influences locking overhead
2480 * and slab fragmentation. A higher order reduces the number of partial slabs
2481 * and increases the number of allocations possible without having to
2482 * take the list_lock.
2483 */
2484static int slub_min_order;
2485static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2486static int slub_min_objects;
2487
2488/*
2489 * Merge control. If this is set then no merging of slab caches will occur.
2490 * (Could be removed. This was introduced to pacify the merge skeptics.)
2491 */
2492static int slub_nomerge;
2493
2494/*
2495 * Calculate the order of allocation given an slab object size.
2496 *
2497 * The order of allocation has significant impact on performance and other
2498 * system components. Generally order 0 allocations should be preferred since
2499 * order 0 does not cause fragmentation in the page allocator. Larger objects
2500 * be problematic to put into order 0 slabs because there may be too much
2501 * unused space left. We go to a higher order if more than 1/16th of the slab
2502 * would be wasted.
2503 *
2504 * In order to reach satisfactory performance we must ensure that a minimum
2505 * number of objects is in one slab. Otherwise we may generate too much
2506 * activity on the partial lists which requires taking the list_lock. This is
2507 * less a concern for large slabs though which are rarely used.
2508 *
2509 * slub_max_order specifies the order where we begin to stop considering the
2510 * number of objects in a slab as critical. If we reach slub_max_order then
2511 * we try to keep the page order as low as possible. So we accept more waste
2512 * of space in favor of a small page order.
2513 *
2514 * Higher order allocations also allow the placement of more objects in a
2515 * slab and thereby reduce object handling overhead. If the user has
2516 * requested a higher mininum order then we start with that one instead of
2517 * the smallest order which will fit the object.
2518 */
2519static inline int slab_order(int size, int min_objects,
2520 int max_order, int fract_leftover, int reserved)
2521{
2522 int order;
2523 int rem;
2524 int min_order = slub_min_order;
2525
2526 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2527 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2528
2529 for (order = max(min_order,
2530 fls(min_objects * size - 1) - PAGE_SHIFT);
2531 order <= max_order; order++) {
2532
2533 unsigned long slab_size = PAGE_SIZE << order;
2534
2535 if (slab_size < min_objects * size + reserved)
2536 continue;
2537
2538 rem = (slab_size - reserved) % size;
2539
2540 if (rem <= slab_size / fract_leftover)
2541 break;
2542
2543 }
2544
2545 return order;
2546}
2547
2548static inline int calculate_order(int size, int reserved)
2549{
2550 int order;
2551 int min_objects;
2552 int fraction;
2553 int max_objects;
2554
2555 /*
2556 * Attempt to find best configuration for a slab. This
2557 * works by first attempting to generate a layout with
2558 * the best configuration and backing off gradually.
2559 *
2560 * First we reduce the acceptable waste in a slab. Then
2561 * we reduce the minimum objects required in a slab.
2562 */
2563 min_objects = slub_min_objects;
2564 if (!min_objects)
2565 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2566 max_objects = order_objects(slub_max_order, size, reserved);
2567 min_objects = min(min_objects, max_objects);
2568
2569 while (min_objects > 1) {
2570 fraction = 16;
2571 while (fraction >= 4) {
2572 order = slab_order(size, min_objects,
2573 slub_max_order, fraction, reserved);
2574 if (order <= slub_max_order)
2575 return order;
2576 fraction /= 2;
2577 }
2578 min_objects--;
2579 }
2580
2581 /*
2582 * We were unable to place multiple objects in a slab. Now
2583 * lets see if we can place a single object there.
2584 */
2585 order = slab_order(size, 1, slub_max_order, 1, reserved);
2586 if (order <= slub_max_order)
2587 return order;
2588
2589 /*
2590 * Doh this slab cannot be placed using slub_max_order.
2591 */
2592 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2593 if (order < MAX_ORDER)
2594 return order;
2595 return -ENOSYS;
2596}
2597
2598/*
2599 * Figure out what the alignment of the objects will be.
2600 */
2601static unsigned long calculate_alignment(unsigned long flags,
2602 unsigned long align, unsigned long size)
2603{
2604 /*
2605 * If the user wants hardware cache aligned objects then follow that
2606 * suggestion if the object is sufficiently large.
2607 *
2608 * The hardware cache alignment cannot override the specified
2609 * alignment though. If that is greater then use it.
2610 */
2611 if (flags & SLAB_HWCACHE_ALIGN) {
2612 unsigned long ralign = cache_line_size();
2613 while (size <= ralign / 2)
2614 ralign /= 2;
2615 align = max(align, ralign);
2616 }
2617
2618 if (align < ARCH_SLAB_MINALIGN)
2619 align = ARCH_SLAB_MINALIGN;
2620
2621 return ALIGN(align, sizeof(void *));
2622}
2623
2624static void
2625init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
2626{
2627 n->nr_partial = 0;
2628 spin_lock_init(&n->list_lock);
2629 INIT_LIST_HEAD(&n->partial);
2630#ifdef CONFIG_SLUB_DEBUG
2631 atomic_long_set(&n->nr_slabs, 0);
2632 atomic_long_set(&n->total_objects, 0);
2633 INIT_LIST_HEAD(&n->full);
2634#endif
2635}
2636
2637static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2638{
2639 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2640 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2641
2642 /*
2643 * Must align to double word boundary for the double cmpxchg
2644 * instructions to work; see __pcpu_double_call_return_bool().
2645 */
2646 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2647 2 * sizeof(void *));
2648
2649 if (!s->cpu_slab)
2650 return 0;
2651
2652 init_kmem_cache_cpus(s);
2653
2654 return 1;
2655}
2656
2657static struct kmem_cache *kmem_cache_node;
2658
2659/*
2660 * No kmalloc_node yet so do it by hand. We know that this is the first
2661 * slab on the node for this slabcache. There are no concurrent accesses
2662 * possible.
2663 *
2664 * Note that this function only works on the kmalloc_node_cache
2665 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2666 * memory on a fresh node that has no slab structures yet.
2667 */
2668static void early_kmem_cache_node_alloc(int node)
2669{
2670 struct page *page;
2671 struct kmem_cache_node *n;
2672
2673 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2674
2675 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2676
2677 BUG_ON(!page);
2678 if (page_to_nid(page) != node) {
2679 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2680 "node %d\n", node);
2681 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2682 "in order to be able to continue\n");
2683 }
2684
2685 n = page->freelist;
2686 BUG_ON(!n);
2687 page->freelist = get_freepointer(kmem_cache_node, n);
2688 page->inuse++;
2689 page->frozen = 0;
2690 kmem_cache_node->node[node] = n;
2691#ifdef CONFIG_SLUB_DEBUG
2692 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2693 init_tracking(kmem_cache_node, n);
2694#endif
2695 init_kmem_cache_node(n, kmem_cache_node);
2696 inc_slabs_node(kmem_cache_node, node, page->objects);
2697
2698 add_partial(n, page, 0);
2699}
2700
2701static void free_kmem_cache_nodes(struct kmem_cache *s)
2702{
2703 int node;
2704
2705 for_each_node_state(node, N_NORMAL_MEMORY) {
2706 struct kmem_cache_node *n = s->node[node];
2707
2708 if (n)
2709 kmem_cache_free(kmem_cache_node, n);
2710
2711 s->node[node] = NULL;
2712 }
2713}
2714
2715static int init_kmem_cache_nodes(struct kmem_cache *s)
2716{
2717 int node;
2718
2719 for_each_node_state(node, N_NORMAL_MEMORY) {
2720 struct kmem_cache_node *n;
2721
2722 if (slab_state == DOWN) {
2723 early_kmem_cache_node_alloc(node);
2724 continue;
2725 }
2726 n = kmem_cache_alloc_node(kmem_cache_node,
2727 GFP_KERNEL, node);
2728
2729 if (!n) {
2730 free_kmem_cache_nodes(s);
2731 return 0;
2732 }
2733
2734 s->node[node] = n;
2735 init_kmem_cache_node(n, s);
2736 }
2737 return 1;
2738}
2739
2740static void set_min_partial(struct kmem_cache *s, unsigned long min)
2741{
2742 if (min < MIN_PARTIAL)
2743 min = MIN_PARTIAL;
2744 else if (min > MAX_PARTIAL)
2745 min = MAX_PARTIAL;
2746 s->min_partial = min;
2747}
2748
2749/*
2750 * calculate_sizes() determines the order and the distribution of data within
2751 * a slab object.
2752 */
2753static int calculate_sizes(struct kmem_cache *s, int forced_order)
2754{
2755 unsigned long flags = s->flags;
2756 unsigned long size = s->objsize;
2757 unsigned long align = s->align;
2758 int order;
2759
2760 /*
2761 * Round up object size to the next word boundary. We can only
2762 * place the free pointer at word boundaries and this determines
2763 * the possible location of the free pointer.
2764 */
2765 size = ALIGN(size, sizeof(void *));
2766
2767#ifdef CONFIG_SLUB_DEBUG
2768 /*
2769 * Determine if we can poison the object itself. If the user of
2770 * the slab may touch the object after free or before allocation
2771 * then we should never poison the object itself.
2772 */
2773 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2774 !s->ctor)
2775 s->flags |= __OBJECT_POISON;
2776 else
2777 s->flags &= ~__OBJECT_POISON;
2778
2779
2780 /*
2781 * If we are Redzoning then check if there is some space between the
2782 * end of the object and the free pointer. If not then add an
2783 * additional word to have some bytes to store Redzone information.
2784 */
2785 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2786 size += sizeof(void *);
2787#endif
2788
2789 /*
2790 * With that we have determined the number of bytes in actual use
2791 * by the object. This is the potential offset to the free pointer.
2792 */
2793 s->inuse = size;
2794
2795 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2796 s->ctor)) {
2797 /*
2798 * Relocate free pointer after the object if it is not
2799 * permitted to overwrite the first word of the object on
2800 * kmem_cache_free.
2801 *
2802 * This is the case if we do RCU, have a constructor or
2803 * destructor or are poisoning the objects.
2804 */
2805 s->offset = size;
2806 size += sizeof(void *);
2807 }
2808
2809#ifdef CONFIG_SLUB_DEBUG
2810 if (flags & SLAB_STORE_USER)
2811 /*
2812 * Need to store information about allocs and frees after
2813 * the object.
2814 */
2815 size += 2 * sizeof(struct track);
2816
2817 if (flags & SLAB_RED_ZONE)
2818 /*
2819 * Add some empty padding so that we can catch
2820 * overwrites from earlier objects rather than let
2821 * tracking information or the free pointer be
2822 * corrupted if a user writes before the start
2823 * of the object.
2824 */
2825 size += sizeof(void *);
2826#endif
2827
2828 /*
2829 * Determine the alignment based on various parameters that the
2830 * user specified and the dynamic determination of cache line size
2831 * on bootup.
2832 */
2833 align = calculate_alignment(flags, align, s->objsize);
2834 s->align = align;
2835
2836 /*
2837 * SLUB stores one object immediately after another beginning from
2838 * offset 0. In order to align the objects we have to simply size
2839 * each object to conform to the alignment.
2840 */
2841 size = ALIGN(size, align);
2842 s->size = size;
2843 if (forced_order >= 0)
2844 order = forced_order;
2845 else
2846 order = calculate_order(size, s->reserved);
2847
2848 if (order < 0)
2849 return 0;
2850
2851 s->allocflags = 0;
2852 if (order)
2853 s->allocflags |= __GFP_COMP;
2854
2855 if (s->flags & SLAB_CACHE_DMA)
2856 s->allocflags |= SLUB_DMA;
2857
2858 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2859 s->allocflags |= __GFP_RECLAIMABLE;
2860
2861 /*
2862 * Determine the number of objects per slab
2863 */
2864 s->oo = oo_make(order, size, s->reserved);
2865 s->min = oo_make(get_order(size), size, s->reserved);
2866 if (oo_objects(s->oo) > oo_objects(s->max))
2867 s->max = s->oo;
2868
2869 return !!oo_objects(s->oo);
2870
2871}
2872
2873static int kmem_cache_open(struct kmem_cache *s,
2874 const char *name, size_t size,
2875 size_t align, unsigned long flags,
2876 void (*ctor)(void *))
2877{
2878 memset(s, 0, kmem_size);
2879 s->name = name;
2880 s->ctor = ctor;
2881 s->objsize = size;
2882 s->align = align;
2883 s->flags = kmem_cache_flags(size, flags, name, ctor);
2884 s->reserved = 0;
2885
2886 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
2887 s->reserved = sizeof(struct rcu_head);
2888
2889 if (!calculate_sizes(s, -1))
2890 goto error;
2891 if (disable_higher_order_debug) {
2892 /*
2893 * Disable debugging flags that store metadata if the min slab
2894 * order increased.
2895 */
2896 if (get_order(s->size) > get_order(s->objsize)) {
2897 s->flags &= ~DEBUG_METADATA_FLAGS;
2898 s->offset = 0;
2899 if (!calculate_sizes(s, -1))
2900 goto error;
2901 }
2902 }
2903
2904#ifdef CONFIG_CMPXCHG_DOUBLE
2905 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
2906 /* Enable fast mode */
2907 s->flags |= __CMPXCHG_DOUBLE;
2908#endif
2909
2910 /*
2911 * The larger the object size is, the more pages we want on the partial
2912 * list to avoid pounding the page allocator excessively.
2913 */
2914 set_min_partial(s, ilog2(s->size));
2915 s->refcount = 1;
2916#ifdef CONFIG_NUMA
2917 s->remote_node_defrag_ratio = 1000;
2918#endif
2919 if (!init_kmem_cache_nodes(s))
2920 goto error;
2921
2922 if (alloc_kmem_cache_cpus(s))
2923 return 1;
2924
2925 free_kmem_cache_nodes(s);
2926error:
2927 if (flags & SLAB_PANIC)
2928 panic("Cannot create slab %s size=%lu realsize=%u "
2929 "order=%u offset=%u flags=%lx\n",
2930 s->name, (unsigned long)size, s->size, oo_order(s->oo),
2931 s->offset, flags);
2932 return 0;
2933}
2934
2935/*
2936 * Determine the size of a slab object
2937 */
2938unsigned int kmem_cache_size(struct kmem_cache *s)
2939{
2940 return s->objsize;
2941}
2942EXPORT_SYMBOL(kmem_cache_size);
2943
2944static void list_slab_objects(struct kmem_cache *s, struct page *page,
2945 const char *text)
2946{
2947#ifdef CONFIG_SLUB_DEBUG
2948 void *addr = page_address(page);
2949 void *p;
2950 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
2951 sizeof(long), GFP_ATOMIC);
2952 if (!map)
2953 return;
2954 slab_err(s, page, "%s", text);
2955 slab_lock(page);
2956
2957 get_map(s, page, map);
2958 for_each_object(p, s, addr, page->objects) {
2959
2960 if (!test_bit(slab_index(p, s, addr), map)) {
2961 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2962 p, p - addr);
2963 print_tracking(s, p);
2964 }
2965 }
2966 slab_unlock(page);
2967 kfree(map);
2968#endif
2969}
2970
2971/*
2972 * Attempt to free all partial slabs on a node.
2973 */
2974static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2975{
2976 unsigned long flags;
2977 struct page *page, *h;
2978
2979 spin_lock_irqsave(&n->list_lock, flags);
2980 list_for_each_entry_safe(page, h, &n->partial, lru) {
2981 if (!page->inuse) {
2982 remove_partial(n, page);
2983 discard_slab(s, page);
2984 } else {
2985 list_slab_objects(s, page,
2986 "Objects remaining on kmem_cache_close()");
2987 }
2988 }
2989 spin_unlock_irqrestore(&n->list_lock, flags);
2990}
2991
2992/*
2993 * Release all resources used by a slab cache.
2994 */
2995static inline int kmem_cache_close(struct kmem_cache *s)
2996{
2997 int node;
2998
2999 flush_all(s);
3000 free_percpu(s->cpu_slab);
3001 /* Attempt to free all objects */
3002 for_each_node_state(node, N_NORMAL_MEMORY) {
3003 struct kmem_cache_node *n = get_node(s, node);
3004
3005 free_partial(s, n);
3006 if (n->nr_partial || slabs_node(s, node))
3007 return 1;
3008 }
3009 free_kmem_cache_nodes(s);
3010 return 0;
3011}
3012
3013/*
3014 * Close a cache and release the kmem_cache structure
3015 * (must be used for caches created using kmem_cache_create)
3016 */
3017void kmem_cache_destroy(struct kmem_cache *s)
3018{
3019 down_write(&slub_lock);
3020 s->refcount--;
3021 if (!s->refcount) {
3022 list_del(&s->list);
3023 if (kmem_cache_close(s)) {
3024 printk(KERN_ERR "SLUB %s: %s called for cache that "
3025 "still has objects.\n", s->name, __func__);
3026 dump_stack();
3027 }
3028 if (s->flags & SLAB_DESTROY_BY_RCU)
3029 rcu_barrier();
3030 sysfs_slab_remove(s);
3031 }
3032 up_write(&slub_lock);
3033}
3034EXPORT_SYMBOL(kmem_cache_destroy);
3035
3036/********************************************************************
3037 * Kmalloc subsystem
3038 *******************************************************************/
3039
3040struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
3041EXPORT_SYMBOL(kmalloc_caches);
3042
3043static struct kmem_cache *kmem_cache;
3044
3045#ifdef CONFIG_ZONE_DMA
3046static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
3047#endif
3048
3049static int __init setup_slub_min_order(char *str)
3050{
3051 get_option(&str, &slub_min_order);
3052
3053 return 1;
3054}
3055
3056__setup("slub_min_order=", setup_slub_min_order);
3057
3058static int __init setup_slub_max_order(char *str)
3059{
3060 get_option(&str, &slub_max_order);
3061 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3062
3063 return 1;
3064}
3065
3066__setup("slub_max_order=", setup_slub_max_order);
3067
3068static int __init setup_slub_min_objects(char *str)
3069{
3070 get_option(&str, &slub_min_objects);
3071
3072 return 1;
3073}
3074
3075__setup("slub_min_objects=", setup_slub_min_objects);
3076
3077static int __init setup_slub_nomerge(char *str)
3078{
3079 slub_nomerge = 1;
3080 return 1;
3081}
3082
3083__setup("slub_nomerge", setup_slub_nomerge);
3084
3085static struct kmem_cache *__init create_kmalloc_cache(const char *name,
3086 int size, unsigned int flags)
3087{
3088 struct kmem_cache *s;
3089
3090 s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3091
3092 /*
3093 * This function is called with IRQs disabled during early-boot on
3094 * single CPU so there's no need to take slub_lock here.
3095 */
3096 if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN,
3097 flags, NULL))
3098 goto panic;
3099
3100 list_add(&s->list, &slab_caches);
3101 return s;
3102
3103panic:
3104 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
3105 return NULL;
3106}
3107
3108/*
3109 * Conversion table for small slabs sizes / 8 to the index in the
3110 * kmalloc array. This is necessary for slabs < 192 since we have non power
3111 * of two cache sizes there. The size of larger slabs can be determined using
3112 * fls.
3113 */
3114static s8 size_index[24] = {
3115 3, /* 8 */
3116 4, /* 16 */
3117 5, /* 24 */
3118 5, /* 32 */
3119 6, /* 40 */
3120 6, /* 48 */
3121 6, /* 56 */
3122 6, /* 64 */
3123 1, /* 72 */
3124 1, /* 80 */
3125 1, /* 88 */
3126 1, /* 96 */
3127 7, /* 104 */
3128 7, /* 112 */
3129 7, /* 120 */
3130 7, /* 128 */
3131 2, /* 136 */
3132 2, /* 144 */
3133 2, /* 152 */
3134 2, /* 160 */
3135 2, /* 168 */
3136 2, /* 176 */
3137 2, /* 184 */
3138 2 /* 192 */
3139};
3140
3141static inline int size_index_elem(size_t bytes)
3142{
3143 return (bytes - 1) / 8;
3144}
3145
3146static struct kmem_cache *get_slab(size_t size, gfp_t flags)
3147{
3148 int index;
3149
3150 if (size <= 192) {
3151 if (!size)
3152 return ZERO_SIZE_PTR;
3153
3154 index = size_index[size_index_elem(size)];
3155 } else
3156 index = fls(size - 1);
3157
3158#ifdef CONFIG_ZONE_DMA
3159 if (unlikely((flags & SLUB_DMA)))
3160 return kmalloc_dma_caches[index];
3161
3162#endif
3163 return kmalloc_caches[index];
3164}
3165
3166void *__kmalloc(size_t size, gfp_t flags)
3167{
3168 struct kmem_cache *s;
3169 void *ret;
3170
3171 if (unlikely(size > SLUB_MAX_SIZE))
3172 return kmalloc_large(size, flags);
3173
3174 s = get_slab(size, flags);
3175
3176 if (unlikely(ZERO_OR_NULL_PTR(s)))
3177 return s;
3178
3179 ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_);
3180
3181 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3182
3183 return ret;
3184}
3185EXPORT_SYMBOL(__kmalloc);
3186
3187#ifdef CONFIG_NUMA
3188static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3189{
3190 struct page *page;
3191 void *ptr = NULL;
3192
3193 flags |= __GFP_COMP | __GFP_NOTRACK;
3194 page = alloc_pages_node(node, flags, get_order(size));
3195 if (page)
3196 ptr = page_address(page);
3197
3198 kmemleak_alloc(ptr, size, 1, flags);
3199 return ptr;
3200}
3201
3202void *__kmalloc_node(size_t size, gfp_t flags, int node)
3203{
3204 struct kmem_cache *s;
3205 void *ret;
3206
3207 if (unlikely(size > SLUB_MAX_SIZE)) {
3208 ret = kmalloc_large_node(size, flags, node);
3209
3210 trace_kmalloc_node(_RET_IP_, ret,
3211 size, PAGE_SIZE << get_order(size),
3212 flags, node);
3213
3214 return ret;
3215 }
3216
3217 s = get_slab(size, flags);
3218
3219 if (unlikely(ZERO_OR_NULL_PTR(s)))
3220 return s;
3221
3222 ret = slab_alloc(s, flags, node, _RET_IP_);
3223
3224 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3225
3226 return ret;
3227}
3228EXPORT_SYMBOL(__kmalloc_node);
3229#endif
3230
3231size_t ksize(const void *object)
3232{
3233 struct page *page;
3234
3235 if (unlikely(object == ZERO_SIZE_PTR))
3236 return 0;
3237
3238 page = virt_to_head_page(object);
3239
3240 if (unlikely(!PageSlab(page))) {
3241 WARN_ON(!PageCompound(page));
3242 return PAGE_SIZE << compound_order(page);
3243 }
3244
3245 return slab_ksize(page->slab);
3246}
3247EXPORT_SYMBOL(ksize);
3248
3249#ifdef CONFIG_SLUB_DEBUG
3250bool verify_mem_not_deleted(const void *x)
3251{
3252 struct page *page;
3253 void *object = (void *)x;
3254 unsigned long flags;
3255 bool rv;
3256
3257 if (unlikely(ZERO_OR_NULL_PTR(x)))
3258 return false;
3259
3260 local_irq_save(flags);
3261
3262 page = virt_to_head_page(x);
3263 if (unlikely(!PageSlab(page))) {
3264 /* maybe it was from stack? */
3265 rv = true;
3266 goto out_unlock;
3267 }
3268
3269 slab_lock(page);
3270 if (on_freelist(page->slab, page, object)) {
3271 object_err(page->slab, page, object, "Object is on free-list");
3272 rv = false;
3273 } else {
3274 rv = true;
3275 }
3276 slab_unlock(page);
3277
3278out_unlock:
3279 local_irq_restore(flags);
3280 return rv;
3281}
3282EXPORT_SYMBOL(verify_mem_not_deleted);
3283#endif
3284
3285void kfree(const void *x)
3286{
3287 struct page *page;
3288 void *object = (void *)x;
3289
3290 trace_kfree(_RET_IP_, x);
3291
3292 if (unlikely(ZERO_OR_NULL_PTR(x)))
3293 return;
3294
3295 page = virt_to_head_page(x);
3296 if (unlikely(!PageSlab(page))) {
3297 BUG_ON(!PageCompound(page));
3298 kmemleak_free(x);
3299 put_page(page);
3300 return;
3301 }
3302 slab_free(page->slab, page, object, _RET_IP_);
3303}
3304EXPORT_SYMBOL(kfree);
3305
3306/*
3307 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3308 * the remaining slabs by the number of items in use. The slabs with the
3309 * most items in use come first. New allocations will then fill those up
3310 * and thus they can be removed from the partial lists.
3311 *
3312 * The slabs with the least items are placed last. This results in them
3313 * being allocated from last increasing the chance that the last objects
3314 * are freed in them.
3315 */
3316int kmem_cache_shrink(struct kmem_cache *s)
3317{
3318 int node;
3319 int i;
3320 struct kmem_cache_node *n;
3321 struct page *page;
3322 struct page *t;
3323 int objects = oo_objects(s->max);
3324 struct list_head *slabs_by_inuse =
3325 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3326 unsigned long flags;
3327
3328 if (!slabs_by_inuse)
3329 return -ENOMEM;
3330
3331 flush_all(s);
3332 for_each_node_state(node, N_NORMAL_MEMORY) {
3333 n = get_node(s, node);
3334
3335 if (!n->nr_partial)
3336 continue;
3337
3338 for (i = 0; i < objects; i++)
3339 INIT_LIST_HEAD(slabs_by_inuse + i);
3340
3341 spin_lock_irqsave(&n->list_lock, flags);
3342
3343 /*
3344 * Build lists indexed by the items in use in each slab.
3345 *
3346 * Note that concurrent frees may occur while we hold the
3347 * list_lock. page->inuse here is the upper limit.
3348 */
3349 list_for_each_entry_safe(page, t, &n->partial, lru) {
3350 if (!page->inuse) {
3351 remove_partial(n, page);
3352 discard_slab(s, page);
3353 } else {
3354 list_move(&page->lru,
3355 slabs_by_inuse + page->inuse);
3356 }
3357 }
3358
3359 /*
3360 * Rebuild the partial list with the slabs filled up most
3361 * first and the least used slabs at the end.
3362 */
3363 for (i = objects - 1; i >= 0; i--)
3364 list_splice(slabs_by_inuse + i, n->partial.prev);
3365
3366 spin_unlock_irqrestore(&n->list_lock, flags);
3367 }
3368
3369 kfree(slabs_by_inuse);
3370 return 0;
3371}
3372EXPORT_SYMBOL(kmem_cache_shrink);
3373
3374#if defined(CONFIG_MEMORY_HOTPLUG)
3375static int slab_mem_going_offline_callback(void *arg)
3376{
3377 struct kmem_cache *s;
3378
3379 down_read(&slub_lock);
3380 list_for_each_entry(s, &slab_caches, list)
3381 kmem_cache_shrink(s);
3382 up_read(&slub_lock);
3383
3384 return 0;
3385}
3386
3387static void slab_mem_offline_callback(void *arg)
3388{
3389 struct kmem_cache_node *n;
3390 struct kmem_cache *s;
3391 struct memory_notify *marg = arg;
3392 int offline_node;
3393
3394 offline_node = marg->status_change_nid;
3395
3396 /*
3397 * If the node still has available memory. we need kmem_cache_node
3398 * for it yet.
3399 */
3400 if (offline_node < 0)
3401 return;
3402
3403 down_read(&slub_lock);
3404 list_for_each_entry(s, &slab_caches, list) {
3405 n = get_node(s, offline_node);
3406 if (n) {
3407 /*
3408 * if n->nr_slabs > 0, slabs still exist on the node
3409 * that is going down. We were unable to free them,
3410 * and offline_pages() function shouldn't call this
3411 * callback. So, we must fail.
3412 */
3413 BUG_ON(slabs_node(s, offline_node));
3414
3415 s->node[offline_node] = NULL;
3416 kmem_cache_free(kmem_cache_node, n);
3417 }
3418 }
3419 up_read(&slub_lock);
3420}
3421
3422static int slab_mem_going_online_callback(void *arg)
3423{
3424 struct kmem_cache_node *n;
3425 struct kmem_cache *s;
3426 struct memory_notify *marg = arg;
3427 int nid = marg->status_change_nid;
3428 int ret = 0;
3429
3430 /*
3431 * If the node's memory is already available, then kmem_cache_node is
3432 * already created. Nothing to do.
3433 */
3434 if (nid < 0)
3435 return 0;
3436
3437 /*
3438 * We are bringing a node online. No memory is available yet. We must
3439 * allocate a kmem_cache_node structure in order to bring the node
3440 * online.
3441 */
3442 down_read(&slub_lock);
3443 list_for_each_entry(s, &slab_caches, list) {
3444 /*
3445 * XXX: kmem_cache_alloc_node will fallback to other nodes
3446 * since memory is not yet available from the node that
3447 * is brought up.
3448 */
3449 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3450 if (!n) {
3451 ret = -ENOMEM;
3452 goto out;
3453 }
3454 init_kmem_cache_node(n, s);
3455 s->node[nid] = n;
3456 }
3457out:
3458 up_read(&slub_lock);
3459 return ret;
3460}
3461
3462static int slab_memory_callback(struct notifier_block *self,
3463 unsigned long action, void *arg)
3464{
3465 int ret = 0;
3466
3467 switch (action) {
3468 case MEM_GOING_ONLINE:
3469 ret = slab_mem_going_online_callback(arg);
3470 break;
3471 case MEM_GOING_OFFLINE:
3472 ret = slab_mem_going_offline_callback(arg);
3473 break;
3474 case MEM_OFFLINE:
3475 case MEM_CANCEL_ONLINE:
3476 slab_mem_offline_callback(arg);
3477 break;
3478 case MEM_ONLINE:
3479 case MEM_CANCEL_OFFLINE:
3480 break;
3481 }
3482 if (ret)
3483 ret = notifier_from_errno(ret);
3484 else
3485 ret = NOTIFY_OK;
3486 return ret;
3487}
3488
3489#endif /* CONFIG_MEMORY_HOTPLUG */
3490
3491/********************************************************************
3492 * Basic setup of slabs
3493 *******************************************************************/
3494
3495/*
3496 * Used for early kmem_cache structures that were allocated using
3497 * the page allocator
3498 */
3499
3500static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s)
3501{
3502 int node;
3503
3504 list_add(&s->list, &slab_caches);
3505 s->refcount = -1;
3506
3507 for_each_node_state(node, N_NORMAL_MEMORY) {
3508 struct kmem_cache_node *n = get_node(s, node);
3509 struct page *p;
3510
3511 if (n) {
3512 list_for_each_entry(p, &n->partial, lru)
3513 p->slab = s;
3514
3515#ifdef CONFIG_SLUB_DEBUG
3516 list_for_each_entry(p, &n->full, lru)
3517 p->slab = s;
3518#endif
3519 }
3520 }
3521}
3522
3523void __init kmem_cache_init(void)
3524{
3525 int i;
3526 int caches = 0;
3527 struct kmem_cache *temp_kmem_cache;
3528 int order;
3529 struct kmem_cache *temp_kmem_cache_node;
3530 unsigned long kmalloc_size;
3531
3532 kmem_size = offsetof(struct kmem_cache, node) +
3533 nr_node_ids * sizeof(struct kmem_cache_node *);
3534
3535 /* Allocate two kmem_caches from the page allocator */
3536 kmalloc_size = ALIGN(kmem_size, cache_line_size());
3537 order = get_order(2 * kmalloc_size);
3538 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order);
3539
3540 /*
3541 * Must first have the slab cache available for the allocations of the
3542 * struct kmem_cache_node's. There is special bootstrap code in
3543 * kmem_cache_open for slab_state == DOWN.
3544 */
3545 kmem_cache_node = (void *)kmem_cache + kmalloc_size;
3546
3547 kmem_cache_open(kmem_cache_node, "kmem_cache_node",
3548 sizeof(struct kmem_cache_node),
3549 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3550
3551 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3552
3553 /* Able to allocate the per node structures */
3554 slab_state = PARTIAL;
3555
3556 temp_kmem_cache = kmem_cache;
3557 kmem_cache_open(kmem_cache, "kmem_cache", kmem_size,
3558 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3559 kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3560 memcpy(kmem_cache, temp_kmem_cache, kmem_size);
3561
3562 /*
3563 * Allocate kmem_cache_node properly from the kmem_cache slab.
3564 * kmem_cache_node is separately allocated so no need to
3565 * update any list pointers.
3566 */
3567 temp_kmem_cache_node = kmem_cache_node;
3568
3569 kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3570 memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size);
3571
3572 kmem_cache_bootstrap_fixup(kmem_cache_node);
3573
3574 caches++;
3575 kmem_cache_bootstrap_fixup(kmem_cache);
3576 caches++;
3577 /* Free temporary boot structure */
3578 free_pages((unsigned long)temp_kmem_cache, order);
3579
3580 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3581
3582 /*
3583 * Patch up the size_index table if we have strange large alignment
3584 * requirements for the kmalloc array. This is only the case for
3585 * MIPS it seems. The standard arches will not generate any code here.
3586 *
3587 * Largest permitted alignment is 256 bytes due to the way we
3588 * handle the index determination for the smaller caches.
3589 *
3590 * Make sure that nothing crazy happens if someone starts tinkering
3591 * around with ARCH_KMALLOC_MINALIGN
3592 */
3593 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3594 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3595
3596 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3597 int elem = size_index_elem(i);
3598 if (elem >= ARRAY_SIZE(size_index))
3599 break;
3600 size_index[elem] = KMALLOC_SHIFT_LOW;
3601 }
3602
3603 if (KMALLOC_MIN_SIZE == 64) {
3604 /*
3605 * The 96 byte size cache is not used if the alignment
3606 * is 64 byte.
3607 */
3608 for (i = 64 + 8; i <= 96; i += 8)
3609 size_index[size_index_elem(i)] = 7;
3610 } else if (KMALLOC_MIN_SIZE == 128) {
3611 /*
3612 * The 192 byte sized cache is not used if the alignment
3613 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3614 * instead.
3615 */
3616 for (i = 128 + 8; i <= 192; i += 8)
3617 size_index[size_index_elem(i)] = 8;
3618 }
3619
3620 /* Caches that are not of the two-to-the-power-of size */
3621 if (KMALLOC_MIN_SIZE <= 32) {
3622 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3623 caches++;
3624 }
3625
3626 if (KMALLOC_MIN_SIZE <= 64) {
3627 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3628 caches++;
3629 }
3630
3631 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3632 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3633 caches++;
3634 }
3635
3636 slab_state = UP;
3637
3638 /* Provide the correct kmalloc names now that the caches are up */
3639 if (KMALLOC_MIN_SIZE <= 32) {
3640 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3641 BUG_ON(!kmalloc_caches[1]->name);
3642 }
3643
3644 if (KMALLOC_MIN_SIZE <= 64) {
3645 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3646 BUG_ON(!kmalloc_caches[2]->name);
3647 }
3648
3649 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3650 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3651
3652 BUG_ON(!s);
3653 kmalloc_caches[i]->name = s;
3654 }
3655
3656#ifdef CONFIG_SMP
3657 register_cpu_notifier(&slab_notifier);
3658#endif
3659
3660#ifdef CONFIG_ZONE_DMA
3661 for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3662 struct kmem_cache *s = kmalloc_caches[i];
3663
3664 if (s && s->size) {
3665 char *name = kasprintf(GFP_NOWAIT,
3666 "dma-kmalloc-%d", s->objsize);
3667
3668 BUG_ON(!name);
3669 kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3670 s->objsize, SLAB_CACHE_DMA);
3671 }
3672 }
3673#endif
3674 printk(KERN_INFO
3675 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3676 " CPUs=%d, Nodes=%d\n",
3677 caches, cache_line_size(),
3678 slub_min_order, slub_max_order, slub_min_objects,
3679 nr_cpu_ids, nr_node_ids);
3680}
3681
3682void __init kmem_cache_init_late(void)
3683{
3684}
3685
3686/*
3687 * Find a mergeable slab cache
3688 */
3689static int slab_unmergeable(struct kmem_cache *s)
3690{
3691 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3692 return 1;
3693
3694 if (s->ctor)
3695 return 1;
3696
3697 /*
3698 * We may have set a slab to be unmergeable during bootstrap.
3699 */
3700 if (s->refcount < 0)
3701 return 1;
3702
3703 return 0;
3704}
3705
3706static struct kmem_cache *find_mergeable(size_t size,
3707 size_t align, unsigned long flags, const char *name,
3708 void (*ctor)(void *))
3709{
3710 struct kmem_cache *s;
3711
3712 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3713 return NULL;
3714
3715 if (ctor)
3716 return NULL;
3717
3718 size = ALIGN(size, sizeof(void *));
3719 align = calculate_alignment(flags, align, size);
3720 size = ALIGN(size, align);
3721 flags = kmem_cache_flags(size, flags, name, NULL);
3722
3723 list_for_each_entry(s, &slab_caches, list) {
3724 if (slab_unmergeable(s))
3725 continue;
3726
3727 if (size > s->size)
3728 continue;
3729
3730 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3731 continue;
3732 /*
3733 * Check if alignment is compatible.
3734 * Courtesy of Adrian Drzewiecki
3735 */
3736 if ((s->size & ~(align - 1)) != s->size)
3737 continue;
3738
3739 if (s->size - size >= sizeof(void *))
3740 continue;
3741
3742 return s;
3743 }
3744 return NULL;
3745}
3746
3747struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3748 size_t align, unsigned long flags, void (*ctor)(void *))
3749{
3750 struct kmem_cache *s;
3751 char *n;
3752
3753 if (WARN_ON(!name))
3754 return NULL;
3755
3756 down_write(&slub_lock);
3757 s = find_mergeable(size, align, flags, name, ctor);
3758 if (s) {
3759 s->refcount++;
3760 /*
3761 * Adjust the object sizes so that we clear
3762 * the complete object on kzalloc.
3763 */
3764 s->objsize = max(s->objsize, (int)size);
3765 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3766
3767 if (sysfs_slab_alias(s, name)) {
3768 s->refcount--;
3769 goto err;
3770 }
3771 up_write(&slub_lock);
3772 return s;
3773 }
3774
3775 n = kstrdup(name, GFP_KERNEL);
3776 if (!n)
3777 goto err;
3778
3779 s = kmalloc(kmem_size, GFP_KERNEL);
3780 if (s) {
3781 if (kmem_cache_open(s, n,
3782 size, align, flags, ctor)) {
3783 list_add(&s->list, &slab_caches);
3784 if (sysfs_slab_add(s)) {
3785 list_del(&s->list);
3786 kfree(n);
3787 kfree(s);
3788 goto err;
3789 }
3790 up_write(&slub_lock);
3791 return s;
3792 }
3793 kfree(n);
3794 kfree(s);
3795 }
3796err:
3797 up_write(&slub_lock);
3798
3799 if (flags & SLAB_PANIC)
3800 panic("Cannot create slabcache %s\n", name);
3801 else
3802 s = NULL;
3803 return s;
3804}
3805EXPORT_SYMBOL(kmem_cache_create);
3806
3807#ifdef CONFIG_SMP
3808/*
3809 * Use the cpu notifier to insure that the cpu slabs are flushed when
3810 * necessary.
3811 */
3812static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3813 unsigned long action, void *hcpu)
3814{
3815 long cpu = (long)hcpu;
3816 struct kmem_cache *s;
3817 unsigned long flags;
3818
3819 switch (action) {
3820 case CPU_UP_CANCELED:
3821 case CPU_UP_CANCELED_FROZEN:
3822 case CPU_DEAD:
3823 case CPU_DEAD_FROZEN:
3824 down_read(&slub_lock);
3825 list_for_each_entry(s, &slab_caches, list) {
3826 local_irq_save(flags);
3827 __flush_cpu_slab(s, cpu);
3828 local_irq_restore(flags);
3829 }
3830 up_read(&slub_lock);
3831 break;
3832 default:
3833 break;
3834 }
3835 return NOTIFY_OK;
3836}
3837
3838static struct notifier_block __cpuinitdata slab_notifier = {
3839 .notifier_call = slab_cpuup_callback
3840};
3841
3842#endif
3843
3844void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3845{
3846 struct kmem_cache *s;
3847 void *ret;
3848
3849 if (unlikely(size > SLUB_MAX_SIZE))
3850 return kmalloc_large(size, gfpflags);
3851
3852 s = get_slab(size, gfpflags);
3853
3854 if (unlikely(ZERO_OR_NULL_PTR(s)))
3855 return s;
3856
3857 ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller);
3858
3859 /* Honor the call site pointer we received. */
3860 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3861
3862 return ret;
3863}
3864
3865#ifdef CONFIG_NUMA
3866void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3867 int node, unsigned long caller)
3868{
3869 struct kmem_cache *s;
3870 void *ret;
3871
3872 if (unlikely(size > SLUB_MAX_SIZE)) {
3873 ret = kmalloc_large_node(size, gfpflags, node);
3874
3875 trace_kmalloc_node(caller, ret,
3876 size, PAGE_SIZE << get_order(size),
3877 gfpflags, node);
3878
3879 return ret;
3880 }
3881
3882 s = get_slab(size, gfpflags);
3883
3884 if (unlikely(ZERO_OR_NULL_PTR(s)))
3885 return s;
3886
3887 ret = slab_alloc(s, gfpflags, node, caller);
3888
3889 /* Honor the call site pointer we received. */
3890 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3891
3892 return ret;
3893}
3894#endif
3895
3896#ifdef CONFIG_SYSFS
3897static int count_inuse(struct page *page)
3898{
3899 return page->inuse;
3900}
3901
3902static int count_total(struct page *page)
3903{
3904 return page->objects;
3905}
3906#endif
3907
3908#ifdef CONFIG_SLUB_DEBUG
3909static int validate_slab(struct kmem_cache *s, struct page *page,
3910 unsigned long *map)
3911{
3912 void *p;
3913 void *addr = page_address(page);
3914
3915 if (!check_slab(s, page) ||
3916 !on_freelist(s, page, NULL))
3917 return 0;
3918
3919 /* Now we know that a valid freelist exists */
3920 bitmap_zero(map, page->objects);
3921
3922 get_map(s, page, map);
3923 for_each_object(p, s, addr, page->objects) {
3924 if (test_bit(slab_index(p, s, addr), map))
3925 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3926 return 0;
3927 }
3928
3929 for_each_object(p, s, addr, page->objects)
3930 if (!test_bit(slab_index(p, s, addr), map))
3931 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3932 return 0;
3933 return 1;
3934}
3935
3936static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3937 unsigned long *map)
3938{
3939 slab_lock(page);
3940 validate_slab(s, page, map);
3941 slab_unlock(page);
3942}
3943
3944static int validate_slab_node(struct kmem_cache *s,
3945 struct kmem_cache_node *n, unsigned long *map)
3946{
3947 unsigned long count = 0;
3948 struct page *page;
3949 unsigned long flags;
3950
3951 spin_lock_irqsave(&n->list_lock, flags);
3952
3953 list_for_each_entry(page, &n->partial, lru) {
3954 validate_slab_slab(s, page, map);
3955 count++;
3956 }
3957 if (count != n->nr_partial)
3958 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3959 "counter=%ld\n", s->name, count, n->nr_partial);
3960
3961 if (!(s->flags & SLAB_STORE_USER))
3962 goto out;
3963
3964 list_for_each_entry(page, &n->full, lru) {
3965 validate_slab_slab(s, page, map);
3966 count++;
3967 }
3968 if (count != atomic_long_read(&n->nr_slabs))
3969 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3970 "counter=%ld\n", s->name, count,
3971 atomic_long_read(&n->nr_slabs));
3972
3973out:
3974 spin_unlock_irqrestore(&n->list_lock, flags);
3975 return count;
3976}
3977
3978static long validate_slab_cache(struct kmem_cache *s)
3979{
3980 int node;
3981 unsigned long count = 0;
3982 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3983 sizeof(unsigned long), GFP_KERNEL);
3984
3985 if (!map)
3986 return -ENOMEM;
3987
3988 flush_all(s);
3989 for_each_node_state(node, N_NORMAL_MEMORY) {
3990 struct kmem_cache_node *n = get_node(s, node);
3991
3992 count += validate_slab_node(s, n, map);
3993 }
3994 kfree(map);
3995 return count;
3996}
3997/*
3998 * Generate lists of code addresses where slabcache objects are allocated
3999 * and freed.
4000 */
4001
4002struct location {
4003 unsigned long count;
4004 unsigned long addr;
4005 long long sum_time;
4006 long min_time;
4007 long max_time;
4008 long min_pid;
4009 long max_pid;
4010 DECLARE_BITMAP(cpus, NR_CPUS);
4011 nodemask_t nodes;
4012};
4013
4014struct loc_track {
4015 unsigned long max;
4016 unsigned long count;
4017 struct location *loc;
4018};
4019
4020static void free_loc_track(struct loc_track *t)
4021{
4022 if (t->max)
4023 free_pages((unsigned long)t->loc,
4024 get_order(sizeof(struct location) * t->max));
4025}
4026
4027static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4028{
4029 struct location *l;
4030 int order;
4031
4032 order = get_order(sizeof(struct location) * max);
4033
4034 l = (void *)__get_free_pages(flags, order);
4035 if (!l)
4036 return 0;
4037
4038 if (t->count) {
4039 memcpy(l, t->loc, sizeof(struct location) * t->count);
4040 free_loc_track(t);
4041 }
4042 t->max = max;
4043 t->loc = l;
4044 return 1;
4045}
4046
4047static int add_location(struct loc_track *t, struct kmem_cache *s,
4048 const struct track *track)
4049{
4050 long start, end, pos;
4051 struct location *l;
4052 unsigned long caddr;
4053 unsigned long age = jiffies - track->when;
4054
4055 start = -1;
4056 end = t->count;
4057
4058 for ( ; ; ) {
4059 pos = start + (end - start + 1) / 2;
4060
4061 /*
4062 * There is nothing at "end". If we end up there
4063 * we need to add something to before end.
4064 */
4065 if (pos == end)
4066 break;
4067
4068 caddr = t->loc[pos].addr;
4069 if (track->addr == caddr) {
4070
4071 l = &t->loc[pos];
4072 l->count++;
4073 if (track->when) {
4074 l->sum_time += age;
4075 if (age < l->min_time)
4076 l->min_time = age;
4077 if (age > l->max_time)
4078 l->max_time = age;
4079
4080 if (track->pid < l->min_pid)
4081 l->min_pid = track->pid;
4082 if (track->pid > l->max_pid)
4083 l->max_pid = track->pid;
4084
4085 cpumask_set_cpu(track->cpu,
4086 to_cpumask(l->cpus));
4087 }
4088 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4089 return 1;
4090 }
4091
4092 if (track->addr < caddr)
4093 end = pos;
4094 else
4095 start = pos;
4096 }
4097
4098 /*
4099 * Not found. Insert new tracking element.
4100 */
4101 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4102 return 0;
4103
4104 l = t->loc + pos;
4105 if (pos < t->count)
4106 memmove(l + 1, l,
4107 (t->count - pos) * sizeof(struct location));
4108 t->count++;
4109 l->count = 1;
4110 l->addr = track->addr;
4111 l->sum_time = age;
4112 l->min_time = age;
4113 l->max_time = age;
4114 l->min_pid = track->pid;
4115 l->max_pid = track->pid;
4116 cpumask_clear(to_cpumask(l->cpus));
4117 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4118 nodes_clear(l->nodes);
4119 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4120 return 1;
4121}
4122
4123static void process_slab(struct loc_track *t, struct kmem_cache *s,
4124 struct page *page, enum track_item alloc,
4125 unsigned long *map)
4126{
4127 void *addr = page_address(page);
4128 void *p;
4129
4130 bitmap_zero(map, page->objects);
4131 get_map(s, page, map);
4132
4133 for_each_object(p, s, addr, page->objects)
4134 if (!test_bit(slab_index(p, s, addr), map))
4135 add_location(t, s, get_track(s, p, alloc));
4136}
4137
4138static int list_locations(struct kmem_cache *s, char *buf,
4139 enum track_item alloc)
4140{
4141 int len = 0;
4142 unsigned long i;
4143 struct loc_track t = { 0, 0, NULL };
4144 int node;
4145 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4146 sizeof(unsigned long), GFP_KERNEL);
4147
4148 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4149 GFP_TEMPORARY)) {
4150 kfree(map);
4151 return sprintf(buf, "Out of memory\n");
4152 }
4153 /* Push back cpu slabs */
4154 flush_all(s);
4155
4156 for_each_node_state(node, N_NORMAL_MEMORY) {
4157 struct kmem_cache_node *n = get_node(s, node);
4158 unsigned long flags;
4159 struct page *page;
4160
4161 if (!atomic_long_read(&n->nr_slabs))
4162 continue;
4163
4164 spin_lock_irqsave(&n->list_lock, flags);
4165 list_for_each_entry(page, &n->partial, lru)
4166 process_slab(&t, s, page, alloc, map);
4167 list_for_each_entry(page, &n->full, lru)
4168 process_slab(&t, s, page, alloc, map);
4169 spin_unlock_irqrestore(&n->list_lock, flags);
4170 }
4171
4172 for (i = 0; i < t.count; i++) {
4173 struct location *l = &t.loc[i];
4174
4175 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4176 break;
4177 len += sprintf(buf + len, "%7ld ", l->count);
4178
4179 if (l->addr)
4180 len += sprintf(buf + len, "%pS", (void *)l->addr);
4181 else
4182 len += sprintf(buf + len, "<not-available>");
4183
4184 if (l->sum_time != l->min_time) {
4185 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4186 l->min_time,
4187 (long)div_u64(l->sum_time, l->count),
4188 l->max_time);
4189 } else
4190 len += sprintf(buf + len, " age=%ld",
4191 l->min_time);
4192
4193 if (l->min_pid != l->max_pid)
4194 len += sprintf(buf + len, " pid=%ld-%ld",
4195 l->min_pid, l->max_pid);
4196 else
4197 len += sprintf(buf + len, " pid=%ld",
4198 l->min_pid);
4199
4200 if (num_online_cpus() > 1 &&
4201 !cpumask_empty(to_cpumask(l->cpus)) &&
4202 len < PAGE_SIZE - 60) {
4203 len += sprintf(buf + len, " cpus=");
4204 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4205 to_cpumask(l->cpus));
4206 }
4207
4208 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4209 len < PAGE_SIZE - 60) {
4210 len += sprintf(buf + len, " nodes=");
4211 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4212 l->nodes);
4213 }
4214
4215 len += sprintf(buf + len, "\n");
4216 }
4217
4218 free_loc_track(&t);
4219 kfree(map);
4220 if (!t.count)
4221 len += sprintf(buf, "No data\n");
4222 return len;
4223}
4224#endif
4225
4226#ifdef SLUB_RESILIENCY_TEST
4227static void resiliency_test(void)
4228{
4229 u8 *p;
4230
4231 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
4232
4233 printk(KERN_ERR "SLUB resiliency testing\n");
4234 printk(KERN_ERR "-----------------------\n");
4235 printk(KERN_ERR "A. Corruption after allocation\n");
4236
4237 p = kzalloc(16, GFP_KERNEL);
4238 p[16] = 0x12;
4239 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4240 " 0x12->0x%p\n\n", p + 16);
4241
4242 validate_slab_cache(kmalloc_caches[4]);
4243
4244 /* Hmmm... The next two are dangerous */
4245 p = kzalloc(32, GFP_KERNEL);
4246 p[32 + sizeof(void *)] = 0x34;
4247 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4248 " 0x34 -> -0x%p\n", p);
4249 printk(KERN_ERR
4250 "If allocated object is overwritten then not detectable\n\n");
4251
4252 validate_slab_cache(kmalloc_caches[5]);
4253 p = kzalloc(64, GFP_KERNEL);
4254 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4255 *p = 0x56;
4256 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4257 p);
4258 printk(KERN_ERR
4259 "If allocated object is overwritten then not detectable\n\n");
4260 validate_slab_cache(kmalloc_caches[6]);
4261
4262 printk(KERN_ERR "\nB. Corruption after free\n");
4263 p = kzalloc(128, GFP_KERNEL);
4264 kfree(p);
4265 *p = 0x78;
4266 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4267 validate_slab_cache(kmalloc_caches[7]);
4268
4269 p = kzalloc(256, GFP_KERNEL);
4270 kfree(p);
4271 p[50] = 0x9a;
4272 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4273 p);
4274 validate_slab_cache(kmalloc_caches[8]);
4275
4276 p = kzalloc(512, GFP_KERNEL);
4277 kfree(p);
4278 p[512] = 0xab;
4279 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4280 validate_slab_cache(kmalloc_caches[9]);
4281}
4282#else
4283#ifdef CONFIG_SYSFS
4284static void resiliency_test(void) {};
4285#endif
4286#endif
4287
4288#ifdef CONFIG_SYSFS
4289enum slab_stat_type {
4290 SL_ALL, /* All slabs */
4291 SL_PARTIAL, /* Only partially allocated slabs */
4292 SL_CPU, /* Only slabs used for cpu caches */
4293 SL_OBJECTS, /* Determine allocated objects not slabs */
4294 SL_TOTAL /* Determine object capacity not slabs */
4295};
4296
4297#define SO_ALL (1 << SL_ALL)
4298#define SO_PARTIAL (1 << SL_PARTIAL)
4299#define SO_CPU (1 << SL_CPU)
4300#define SO_OBJECTS (1 << SL_OBJECTS)
4301#define SO_TOTAL (1 << SL_TOTAL)
4302
4303static ssize_t show_slab_objects(struct kmem_cache *s,
4304 char *buf, unsigned long flags)
4305{
4306 unsigned long total = 0;
4307 int node;
4308 int x;
4309 unsigned long *nodes;
4310 unsigned long *per_cpu;
4311
4312 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4313 if (!nodes)
4314 return -ENOMEM;
4315 per_cpu = nodes + nr_node_ids;
4316
4317 if (flags & SO_CPU) {
4318 int cpu;
4319
4320 for_each_possible_cpu(cpu) {
4321 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4322
4323 if (!c || c->node < 0)
4324 continue;
4325
4326 if (c->page) {
4327 if (flags & SO_TOTAL)
4328 x = c->page->objects;
4329 else if (flags & SO_OBJECTS)
4330 x = c->page->inuse;
4331 else
4332 x = 1;
4333
4334 total += x;
4335 nodes[c->node] += x;
4336 }
4337 per_cpu[c->node]++;
4338 }
4339 }
4340
4341 lock_memory_hotplug();
4342#ifdef CONFIG_SLUB_DEBUG
4343 if (flags & SO_ALL) {
4344 for_each_node_state(node, N_NORMAL_MEMORY) {
4345 struct kmem_cache_node *n = get_node(s, node);
4346
4347 if (flags & SO_TOTAL)
4348 x = atomic_long_read(&n->total_objects);
4349 else if (flags & SO_OBJECTS)
4350 x = atomic_long_read(&n->total_objects) -
4351 count_partial(n, count_free);
4352
4353 else
4354 x = atomic_long_read(&n->nr_slabs);
4355 total += x;
4356 nodes[node] += x;
4357 }
4358
4359 } else
4360#endif
4361 if (flags & SO_PARTIAL) {
4362 for_each_node_state(node, N_NORMAL_MEMORY) {
4363 struct kmem_cache_node *n = get_node(s, node);
4364
4365 if (flags & SO_TOTAL)
4366 x = count_partial(n, count_total);
4367 else if (flags & SO_OBJECTS)
4368 x = count_partial(n, count_inuse);
4369 else
4370 x = n->nr_partial;
4371 total += x;
4372 nodes[node] += x;
4373 }
4374 }
4375 x = sprintf(buf, "%lu", total);
4376#ifdef CONFIG_NUMA
4377 for_each_node_state(node, N_NORMAL_MEMORY)
4378 if (nodes[node])
4379 x += sprintf(buf + x, " N%d=%lu",
4380 node, nodes[node]);
4381#endif
4382 unlock_memory_hotplug();
4383 kfree(nodes);
4384 return x + sprintf(buf + x, "\n");
4385}
4386
4387#ifdef CONFIG_SLUB_DEBUG
4388static int any_slab_objects(struct kmem_cache *s)
4389{
4390 int node;
4391
4392 for_each_online_node(node) {
4393 struct kmem_cache_node *n = get_node(s, node);
4394
4395 if (!n)
4396 continue;
4397
4398 if (atomic_long_read(&n->total_objects))
4399 return 1;
4400 }
4401 return 0;
4402}
4403#endif
4404
4405#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4406#define to_slab(n) container_of(n, struct kmem_cache, kobj)
4407
4408struct slab_attribute {
4409 struct attribute attr;
4410 ssize_t (*show)(struct kmem_cache *s, char *buf);
4411 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4412};
4413
4414#define SLAB_ATTR_RO(_name) \
4415 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
4416
4417#define SLAB_ATTR(_name) \
4418 static struct slab_attribute _name##_attr = \
4419 __ATTR(_name, 0644, _name##_show, _name##_store)
4420
4421static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4422{
4423 return sprintf(buf, "%d\n", s->size);
4424}
4425SLAB_ATTR_RO(slab_size);
4426
4427static ssize_t align_show(struct kmem_cache *s, char *buf)
4428{
4429 return sprintf(buf, "%d\n", s->align);
4430}
4431SLAB_ATTR_RO(align);
4432
4433static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4434{
4435 return sprintf(buf, "%d\n", s->objsize);
4436}
4437SLAB_ATTR_RO(object_size);
4438
4439static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4440{
4441 return sprintf(buf, "%d\n", oo_objects(s->oo));
4442}
4443SLAB_ATTR_RO(objs_per_slab);
4444
4445static ssize_t order_store(struct kmem_cache *s,
4446 const char *buf, size_t length)
4447{
4448 unsigned long order;
4449 int err;
4450
4451 err = strict_strtoul(buf, 10, &order);
4452 if (err)
4453 return err;
4454
4455 if (order > slub_max_order || order < slub_min_order)
4456 return -EINVAL;
4457
4458 calculate_sizes(s, order);
4459 return length;
4460}
4461
4462static ssize_t order_show(struct kmem_cache *s, char *buf)
4463{
4464 return sprintf(buf, "%d\n", oo_order(s->oo));
4465}
4466SLAB_ATTR(order);
4467
4468static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4469{
4470 return sprintf(buf, "%lu\n", s->min_partial);
4471}
4472
4473static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4474 size_t length)
4475{
4476 unsigned long min;
4477 int err;
4478
4479 err = strict_strtoul(buf, 10, &min);
4480 if (err)
4481 return err;
4482
4483 set_min_partial(s, min);
4484 return length;
4485}
4486SLAB_ATTR(min_partial);
4487
4488static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4489{
4490 if (!s->ctor)
4491 return 0;
4492 return sprintf(buf, "%pS\n", s->ctor);
4493}
4494SLAB_ATTR_RO(ctor);
4495
4496static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4497{
4498 return sprintf(buf, "%d\n", s->refcount - 1);
4499}
4500SLAB_ATTR_RO(aliases);
4501
4502static ssize_t partial_show(struct kmem_cache *s, char *buf)
4503{
4504 return show_slab_objects(s, buf, SO_PARTIAL);
4505}
4506SLAB_ATTR_RO(partial);
4507
4508static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4509{
4510 return show_slab_objects(s, buf, SO_CPU);
4511}
4512SLAB_ATTR_RO(cpu_slabs);
4513
4514static ssize_t objects_show(struct kmem_cache *s, char *buf)
4515{
4516 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4517}
4518SLAB_ATTR_RO(objects);
4519
4520static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4521{
4522 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4523}
4524SLAB_ATTR_RO(objects_partial);
4525
4526static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4527{
4528 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4529}
4530
4531static ssize_t reclaim_account_store(struct kmem_cache *s,
4532 const char *buf, size_t length)
4533{
4534 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4535 if (buf[0] == '1')
4536 s->flags |= SLAB_RECLAIM_ACCOUNT;
4537 return length;
4538}
4539SLAB_ATTR(reclaim_account);
4540
4541static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4542{
4543 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4544}
4545SLAB_ATTR_RO(hwcache_align);
4546
4547#ifdef CONFIG_ZONE_DMA
4548static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4549{
4550 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4551}
4552SLAB_ATTR_RO(cache_dma);
4553#endif
4554
4555static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4556{
4557 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4558}
4559SLAB_ATTR_RO(destroy_by_rcu);
4560
4561static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4562{
4563 return sprintf(buf, "%d\n", s->reserved);
4564}
4565SLAB_ATTR_RO(reserved);
4566
4567#ifdef CONFIG_SLUB_DEBUG
4568static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4569{
4570 return show_slab_objects(s, buf, SO_ALL);
4571}
4572SLAB_ATTR_RO(slabs);
4573
4574static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4575{
4576 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4577}
4578SLAB_ATTR_RO(total_objects);
4579
4580static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4581{
4582 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4583}
4584
4585static ssize_t sanity_checks_store(struct kmem_cache *s,
4586 const char *buf, size_t length)
4587{
4588 s->flags &= ~SLAB_DEBUG_FREE;
4589 if (buf[0] == '1') {
4590 s->flags &= ~__CMPXCHG_DOUBLE;
4591 s->flags |= SLAB_DEBUG_FREE;
4592 }
4593 return length;
4594}
4595SLAB_ATTR(sanity_checks);
4596
4597static ssize_t trace_show(struct kmem_cache *s, char *buf)
4598{
4599 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4600}
4601
4602static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4603 size_t length)
4604{
4605 s->flags &= ~SLAB_TRACE;
4606 if (buf[0] == '1') {
4607 s->flags &= ~__CMPXCHG_DOUBLE;
4608 s->flags |= SLAB_TRACE;
4609 }
4610 return length;
4611}
4612SLAB_ATTR(trace);
4613
4614static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4615{
4616 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4617}
4618
4619static ssize_t red_zone_store(struct kmem_cache *s,
4620 const char *buf, size_t length)
4621{
4622 if (any_slab_objects(s))
4623 return -EBUSY;
4624
4625 s->flags &= ~SLAB_RED_ZONE;
4626 if (buf[0] == '1') {
4627 s->flags &= ~__CMPXCHG_DOUBLE;
4628 s->flags |= SLAB_RED_ZONE;
4629 }
4630 calculate_sizes(s, -1);
4631 return length;
4632}
4633SLAB_ATTR(red_zone);
4634
4635static ssize_t poison_show(struct kmem_cache *s, char *buf)
4636{
4637 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4638}
4639
4640static ssize_t poison_store(struct kmem_cache *s,
4641 const char *buf, size_t length)
4642{
4643 if (any_slab_objects(s))
4644 return -EBUSY;
4645
4646 s->flags &= ~SLAB_POISON;
4647 if (buf[0] == '1') {
4648 s->flags &= ~__CMPXCHG_DOUBLE;
4649 s->flags |= SLAB_POISON;
4650 }
4651 calculate_sizes(s, -1);
4652 return length;
4653}
4654SLAB_ATTR(poison);
4655
4656static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4657{
4658 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4659}
4660
4661static ssize_t store_user_store(struct kmem_cache *s,
4662 const char *buf, size_t length)
4663{
4664 if (any_slab_objects(s))
4665 return -EBUSY;
4666
4667 s->flags &= ~SLAB_STORE_USER;
4668 if (buf[0] == '1') {
4669 s->flags &= ~__CMPXCHG_DOUBLE;
4670 s->flags |= SLAB_STORE_USER;
4671 }
4672 calculate_sizes(s, -1);
4673 return length;
4674}
4675SLAB_ATTR(store_user);
4676
4677static ssize_t validate_show(struct kmem_cache *s, char *buf)
4678{
4679 return 0;
4680}
4681
4682static ssize_t validate_store(struct kmem_cache *s,
4683 const char *buf, size_t length)
4684{
4685 int ret = -EINVAL;
4686
4687 if (buf[0] == '1') {
4688 ret = validate_slab_cache(s);
4689 if (ret >= 0)
4690 ret = length;
4691 }
4692 return ret;
4693}
4694SLAB_ATTR(validate);
4695
4696static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4697{
4698 if (!(s->flags & SLAB_STORE_USER))
4699 return -ENOSYS;
4700 return list_locations(s, buf, TRACK_ALLOC);
4701}
4702SLAB_ATTR_RO(alloc_calls);
4703
4704static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4705{
4706 if (!(s->flags & SLAB_STORE_USER))
4707 return -ENOSYS;
4708 return list_locations(s, buf, TRACK_FREE);
4709}
4710SLAB_ATTR_RO(free_calls);
4711#endif /* CONFIG_SLUB_DEBUG */
4712
4713#ifdef CONFIG_FAILSLAB
4714static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4715{
4716 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4717}
4718
4719static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4720 size_t length)
4721{
4722 s->flags &= ~SLAB_FAILSLAB;
4723 if (buf[0] == '1')
4724 s->flags |= SLAB_FAILSLAB;
4725 return length;
4726}
4727SLAB_ATTR(failslab);
4728#endif
4729
4730static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4731{
4732 return 0;
4733}
4734
4735static ssize_t shrink_store(struct kmem_cache *s,
4736 const char *buf, size_t length)
4737{
4738 if (buf[0] == '1') {
4739 int rc = kmem_cache_shrink(s);
4740
4741 if (rc)
4742 return rc;
4743 } else
4744 return -EINVAL;
4745 return length;
4746}
4747SLAB_ATTR(shrink);
4748
4749#ifdef CONFIG_NUMA
4750static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4751{
4752 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4753}
4754
4755static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4756 const char *buf, size_t length)
4757{
4758 unsigned long ratio;
4759 int err;
4760
4761 err = strict_strtoul(buf, 10, &ratio);
4762 if (err)
4763 return err;
4764
4765 if (ratio <= 100)
4766 s->remote_node_defrag_ratio = ratio * 10;
4767
4768 return length;
4769}
4770SLAB_ATTR(remote_node_defrag_ratio);
4771#endif
4772
4773#ifdef CONFIG_SLUB_STATS
4774static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4775{
4776 unsigned long sum = 0;
4777 int cpu;
4778 int len;
4779 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4780
4781 if (!data)
4782 return -ENOMEM;
4783
4784 for_each_online_cpu(cpu) {
4785 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4786
4787 data[cpu] = x;
4788 sum += x;
4789 }
4790
4791 len = sprintf(buf, "%lu", sum);
4792
4793#ifdef CONFIG_SMP
4794 for_each_online_cpu(cpu) {
4795 if (data[cpu] && len < PAGE_SIZE - 20)
4796 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4797 }
4798#endif
4799 kfree(data);
4800 return len + sprintf(buf + len, "\n");
4801}
4802
4803static void clear_stat(struct kmem_cache *s, enum stat_item si)
4804{
4805 int cpu;
4806
4807 for_each_online_cpu(cpu)
4808 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4809}
4810
4811#define STAT_ATTR(si, text) \
4812static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4813{ \
4814 return show_stat(s, buf, si); \
4815} \
4816static ssize_t text##_store(struct kmem_cache *s, \
4817 const char *buf, size_t length) \
4818{ \
4819 if (buf[0] != '0') \
4820 return -EINVAL; \
4821 clear_stat(s, si); \
4822 return length; \
4823} \
4824SLAB_ATTR(text); \
4825
4826STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4827STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4828STAT_ATTR(FREE_FASTPATH, free_fastpath);
4829STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4830STAT_ATTR(FREE_FROZEN, free_frozen);
4831STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4832STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4833STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4834STAT_ATTR(ALLOC_SLAB, alloc_slab);
4835STAT_ATTR(ALLOC_REFILL, alloc_refill);
4836STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
4837STAT_ATTR(FREE_SLAB, free_slab);
4838STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4839STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4840STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4841STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4842STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4843STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4844STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
4845STAT_ATTR(ORDER_FALLBACK, order_fallback);
4846STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4847STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4848#endif
4849
4850static struct attribute *slab_attrs[] = {
4851 &slab_size_attr.attr,
4852 &object_size_attr.attr,
4853 &objs_per_slab_attr.attr,
4854 &order_attr.attr,
4855 &min_partial_attr.attr,
4856 &objects_attr.attr,
4857 &objects_partial_attr.attr,
4858 &partial_attr.attr,
4859 &cpu_slabs_attr.attr,
4860 &ctor_attr.attr,
4861 &aliases_attr.attr,
4862 &align_attr.attr,
4863 &hwcache_align_attr.attr,
4864 &reclaim_account_attr.attr,
4865 &destroy_by_rcu_attr.attr,
4866 &shrink_attr.attr,
4867 &reserved_attr.attr,
4868#ifdef CONFIG_SLUB_DEBUG
4869 &total_objects_attr.attr,
4870 &slabs_attr.attr,
4871 &sanity_checks_attr.attr,
4872 &trace_attr.attr,
4873 &red_zone_attr.attr,
4874 &poison_attr.attr,
4875 &store_user_attr.attr,
4876 &validate_attr.attr,
4877 &alloc_calls_attr.attr,
4878 &free_calls_attr.attr,
4879#endif
4880#ifdef CONFIG_ZONE_DMA
4881 &cache_dma_attr.attr,
4882#endif
4883#ifdef CONFIG_NUMA
4884 &remote_node_defrag_ratio_attr.attr,
4885#endif
4886#ifdef CONFIG_SLUB_STATS
4887 &alloc_fastpath_attr.attr,
4888 &alloc_slowpath_attr.attr,
4889 &free_fastpath_attr.attr,
4890 &free_slowpath_attr.attr,
4891 &free_frozen_attr.attr,
4892 &free_add_partial_attr.attr,
4893 &free_remove_partial_attr.attr,
4894 &alloc_from_partial_attr.attr,
4895 &alloc_slab_attr.attr,
4896 &alloc_refill_attr.attr,
4897 &alloc_node_mismatch_attr.attr,
4898 &free_slab_attr.attr,
4899 &cpuslab_flush_attr.attr,
4900 &deactivate_full_attr.attr,
4901 &deactivate_empty_attr.attr,
4902 &deactivate_to_head_attr.attr,
4903 &deactivate_to_tail_attr.attr,
4904 &deactivate_remote_frees_attr.attr,
4905 &deactivate_bypass_attr.attr,
4906 &order_fallback_attr.attr,
4907 &cmpxchg_double_fail_attr.attr,
4908 &cmpxchg_double_cpu_fail_attr.attr,
4909#endif
4910#ifdef CONFIG_FAILSLAB
4911 &failslab_attr.attr,
4912#endif
4913
4914 NULL
4915};
4916
4917static struct attribute_group slab_attr_group = {
4918 .attrs = slab_attrs,
4919};
4920
4921static ssize_t slab_attr_show(struct kobject *kobj,
4922 struct attribute *attr,
4923 char *buf)
4924{
4925 struct slab_attribute *attribute;
4926 struct kmem_cache *s;
4927 int err;
4928
4929 attribute = to_slab_attr(attr);
4930 s = to_slab(kobj);
4931
4932 if (!attribute->show)
4933 return -EIO;
4934
4935 err = attribute->show(s, buf);
4936
4937 return err;
4938}
4939
4940static ssize_t slab_attr_store(struct kobject *kobj,
4941 struct attribute *attr,
4942 const char *buf, size_t len)
4943{
4944 struct slab_attribute *attribute;
4945 struct kmem_cache *s;
4946 int err;
4947
4948 attribute = to_slab_attr(attr);
4949 s = to_slab(kobj);
4950
4951 if (!attribute->store)
4952 return -EIO;
4953
4954 err = attribute->store(s, buf, len);
4955
4956 return err;
4957}
4958
4959static void kmem_cache_release(struct kobject *kobj)
4960{
4961 struct kmem_cache *s = to_slab(kobj);
4962
4963 kfree(s->name);
4964 kfree(s);
4965}
4966
4967static const struct sysfs_ops slab_sysfs_ops = {
4968 .show = slab_attr_show,
4969 .store = slab_attr_store,
4970};
4971
4972static struct kobj_type slab_ktype = {
4973 .sysfs_ops = &slab_sysfs_ops,
4974 .release = kmem_cache_release
4975};
4976
4977static int uevent_filter(struct kset *kset, struct kobject *kobj)
4978{
4979 struct kobj_type *ktype = get_ktype(kobj);
4980
4981 if (ktype == &slab_ktype)
4982 return 1;
4983 return 0;
4984}
4985
4986static const struct kset_uevent_ops slab_uevent_ops = {
4987 .filter = uevent_filter,
4988};
4989
4990static struct kset *slab_kset;
4991
4992#define ID_STR_LENGTH 64
4993
4994/* Create a unique string id for a slab cache:
4995 *
4996 * Format :[flags-]size
4997 */
4998static char *create_unique_id(struct kmem_cache *s)
4999{
5000 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5001 char *p = name;
5002
5003 BUG_ON(!name);
5004
5005 *p++ = ':';
5006 /*
5007 * First flags affecting slabcache operations. We will only
5008 * get here for aliasable slabs so we do not need to support
5009 * too many flags. The flags here must cover all flags that
5010 * are matched during merging to guarantee that the id is
5011 * unique.
5012 */
5013 if (s->flags & SLAB_CACHE_DMA)
5014 *p++ = 'd';
5015 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5016 *p++ = 'a';
5017 if (s->flags & SLAB_DEBUG_FREE)
5018 *p++ = 'F';
5019 if (!(s->flags & SLAB_NOTRACK))
5020 *p++ = 't';
5021 if (p != name + 1)
5022 *p++ = '-';
5023 p += sprintf(p, "%07d", s->size);
5024 BUG_ON(p > name + ID_STR_LENGTH - 1);
5025 return name;
5026}
5027
5028static int sysfs_slab_add(struct kmem_cache *s)
5029{
5030 int err;
5031 const char *name;
5032 int unmergeable;
5033
5034 if (slab_state < SYSFS)
5035 /* Defer until later */
5036 return 0;
5037
5038 unmergeable = slab_unmergeable(s);
5039 if (unmergeable) {
5040 /*
5041 * Slabcache can never be merged so we can use the name proper.
5042 * This is typically the case for debug situations. In that
5043 * case we can catch duplicate names easily.
5044 */
5045 sysfs_remove_link(&slab_kset->kobj, s->name);
5046 name = s->name;
5047 } else {
5048 /*
5049 * Create a unique name for the slab as a target
5050 * for the symlinks.
5051 */
5052 name = create_unique_id(s);
5053 }
5054
5055 s->kobj.kset = slab_kset;
5056 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5057 if (err) {
5058 kobject_put(&s->kobj);
5059 return err;
5060 }
5061
5062 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5063 if (err) {
5064 kobject_del(&s->kobj);
5065 kobject_put(&s->kobj);
5066 return err;
5067 }
5068 kobject_uevent(&s->kobj, KOBJ_ADD);
5069 if (!unmergeable) {
5070 /* Setup first alias */
5071 sysfs_slab_alias(s, s->name);
5072 kfree(name);
5073 }
5074 return 0;
5075}
5076
5077static void sysfs_slab_remove(struct kmem_cache *s)
5078{
5079 if (slab_state < SYSFS)
5080 /*
5081 * Sysfs has not been setup yet so no need to remove the
5082 * cache from sysfs.
5083 */
5084 return;
5085
5086 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5087 kobject_del(&s->kobj);
5088 kobject_put(&s->kobj);
5089}
5090
5091/*
5092 * Need to buffer aliases during bootup until sysfs becomes
5093 * available lest we lose that information.
5094 */
5095struct saved_alias {
5096 struct kmem_cache *s;
5097 const char *name;
5098 struct saved_alias *next;
5099};
5100
5101static struct saved_alias *alias_list;
5102
5103static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5104{
5105 struct saved_alias *al;
5106
5107 if (slab_state == SYSFS) {
5108 /*
5109 * If we have a leftover link then remove it.
5110 */
5111 sysfs_remove_link(&slab_kset->kobj, name);
5112 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5113 }
5114
5115 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5116 if (!al)
5117 return -ENOMEM;
5118
5119 al->s = s;
5120 al->name = name;
5121 al->next = alias_list;
5122 alias_list = al;
5123 return 0;
5124}
5125
5126static int __init slab_sysfs_init(void)
5127{
5128 struct kmem_cache *s;
5129 int err;
5130
5131 down_write(&slub_lock);
5132
5133 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5134 if (!slab_kset) {
5135 up_write(&slub_lock);
5136 printk(KERN_ERR "Cannot register slab subsystem.\n");
5137 return -ENOSYS;
5138 }
5139
5140 slab_state = SYSFS;
5141
5142 list_for_each_entry(s, &slab_caches, list) {
5143 err = sysfs_slab_add(s);
5144 if (err)
5145 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5146 " to sysfs\n", s->name);
5147 }
5148
5149 while (alias_list) {
5150 struct saved_alias *al = alias_list;
5151
5152 alias_list = alias_list->next;
5153 err = sysfs_slab_alias(al->s, al->name);
5154 if (err)
5155 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5156 " %s to sysfs\n", s->name);
5157 kfree(al);
5158 }
5159
5160 up_write(&slub_lock);
5161 resiliency_test();
5162 return 0;
5163}
5164
5165__initcall(slab_sysfs_init);
5166#endif /* CONFIG_SYSFS */
5167
5168/*
5169 * The /proc/slabinfo ABI
5170 */
5171#ifdef CONFIG_SLABINFO
5172static void print_slabinfo_header(struct seq_file *m)
5173{
5174 seq_puts(m, "slabinfo - version: 2.1\n");
5175 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
5176 "<objperslab> <pagesperslab>");
5177 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
5178 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
5179 seq_putc(m, '\n');
5180}
5181
5182static void *s_start(struct seq_file *m, loff_t *pos)
5183{
5184 loff_t n = *pos;
5185
5186 down_read(&slub_lock);
5187 if (!n)
5188 print_slabinfo_header(m);
5189
5190 return seq_list_start(&slab_caches, *pos);
5191}
5192
5193static void *s_next(struct seq_file *m, void *p, loff_t *pos)
5194{
5195 return seq_list_next(p, &slab_caches, pos);
5196}
5197
5198static void s_stop(struct seq_file *m, void *p)
5199{
5200 up_read(&slub_lock);
5201}
5202
5203static int s_show(struct seq_file *m, void *p)
5204{
5205 unsigned long nr_partials = 0;
5206 unsigned long nr_slabs = 0;
5207 unsigned long nr_inuse = 0;
5208 unsigned long nr_objs = 0;
5209 unsigned long nr_free = 0;
5210 struct kmem_cache *s;
5211 int node;
5212
5213 s = list_entry(p, struct kmem_cache, list);
5214
5215 for_each_online_node(node) {
5216 struct kmem_cache_node *n = get_node(s, node);
5217
5218 if (!n)
5219 continue;
5220
5221 nr_partials += n->nr_partial;
5222 nr_slabs += atomic_long_read(&n->nr_slabs);
5223 nr_objs += atomic_long_read(&n->total_objects);
5224 nr_free += count_partial(n, count_free);
5225 }
5226
5227 nr_inuse = nr_objs - nr_free;
5228
5229 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
5230 nr_objs, s->size, oo_objects(s->oo),
5231 (1 << oo_order(s->oo)));
5232 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
5233 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
5234 0UL);
5235 seq_putc(m, '\n');
5236 return 0;
5237}
5238
5239static const struct seq_operations slabinfo_op = {
5240 .start = s_start,
5241 .next = s_next,
5242 .stop = s_stop,
5243 .show = s_show,
5244};
5245
5246static int slabinfo_open(struct inode *inode, struct file *file)
5247{
5248 return seq_open(file, &slabinfo_op);
5249}
5250
5251static const struct file_operations proc_slabinfo_operations = {
5252 .open = slabinfo_open,
5253 .read = seq_read,
5254 .llseek = seq_lseek,
5255 .release = seq_release,
5256};
5257
5258static int __init slab_proc_init(void)
5259{
5260 proc_create("slabinfo", S_IRUGO, NULL, &proc_slabinfo_operations);
5261 return 0;
5262}
5263module_init(slab_proc_init);
5264#endif /* CONFIG_SLABINFO */
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * SLUB: A slab allocator that limits cache line use instead of queuing
4 * objects in per cpu and per node lists.
5 *
6 * The allocator synchronizes using per slab locks or atomic operatios
7 * and only uses a centralized lock to manage a pool of partial slabs.
8 *
9 * (C) 2007 SGI, Christoph Lameter
10 * (C) 2011 Linux Foundation, Christoph Lameter
11 */
12
13#include <linux/mm.h>
14#include <linux/swap.h> /* struct reclaim_state */
15#include <linux/module.h>
16#include <linux/bit_spinlock.h>
17#include <linux/interrupt.h>
18#include <linux/bitops.h>
19#include <linux/slab.h>
20#include "slab.h"
21#include <linux/proc_fs.h>
22#include <linux/seq_file.h>
23#include <linux/kasan.h>
24#include <linux/cpu.h>
25#include <linux/cpuset.h>
26#include <linux/mempolicy.h>
27#include <linux/ctype.h>
28#include <linux/debugobjects.h>
29#include <linux/kallsyms.h>
30#include <linux/memory.h>
31#include <linux/math64.h>
32#include <linux/fault-inject.h>
33#include <linux/stacktrace.h>
34#include <linux/prefetch.h>
35#include <linux/memcontrol.h>
36#include <linux/random.h>
37
38#include <trace/events/kmem.h>
39
40#include "internal.h"
41
42/*
43 * Lock order:
44 * 1. slab_mutex (Global Mutex)
45 * 2. node->list_lock
46 * 3. slab_lock(page) (Only on some arches and for debugging)
47 *
48 * slab_mutex
49 *
50 * The role of the slab_mutex is to protect the list of all the slabs
51 * and to synchronize major metadata changes to slab cache structures.
52 *
53 * The slab_lock is only used for debugging and on arches that do not
54 * have the ability to do a cmpxchg_double. It only protects:
55 * A. page->freelist -> List of object free in a page
56 * B. page->inuse -> Number of objects in use
57 * C. page->objects -> Number of objects in page
58 * D. page->frozen -> frozen state
59 *
60 * If a slab is frozen then it is exempt from list management. It is not
61 * on any list except per cpu partial list. The processor that froze the
62 * slab is the one who can perform list operations on the page. Other
63 * processors may put objects onto the freelist but the processor that
64 * froze the slab is the only one that can retrieve the objects from the
65 * page's freelist.
66 *
67 * The list_lock protects the partial and full list on each node and
68 * the partial slab counter. If taken then no new slabs may be added or
69 * removed from the lists nor make the number of partial slabs be modified.
70 * (Note that the total number of slabs is an atomic value that may be
71 * modified without taking the list lock).
72 *
73 * The list_lock is a centralized lock and thus we avoid taking it as
74 * much as possible. As long as SLUB does not have to handle partial
75 * slabs, operations can continue without any centralized lock. F.e.
76 * allocating a long series of objects that fill up slabs does not require
77 * the list lock.
78 * Interrupts are disabled during allocation and deallocation in order to
79 * make the slab allocator safe to use in the context of an irq. In addition
80 * interrupts are disabled to ensure that the processor does not change
81 * while handling per_cpu slabs, due to kernel preemption.
82 *
83 * SLUB assigns one slab for allocation to each processor.
84 * Allocations only occur from these slabs called cpu slabs.
85 *
86 * Slabs with free elements are kept on a partial list and during regular
87 * operations no list for full slabs is used. If an object in a full slab is
88 * freed then the slab will show up again on the partial lists.
89 * We track full slabs for debugging purposes though because otherwise we
90 * cannot scan all objects.
91 *
92 * Slabs are freed when they become empty. Teardown and setup is
93 * minimal so we rely on the page allocators per cpu caches for
94 * fast frees and allocs.
95 *
96 * page->frozen The slab is frozen and exempt from list processing.
97 * This means that the slab is dedicated to a purpose
98 * such as satisfying allocations for a specific
99 * processor. Objects may be freed in the slab while
100 * it is frozen but slab_free will then skip the usual
101 * list operations. It is up to the processor holding
102 * the slab to integrate the slab into the slab lists
103 * when the slab is no longer needed.
104 *
105 * One use of this flag is to mark slabs that are
106 * used for allocations. Then such a slab becomes a cpu
107 * slab. The cpu slab may be equipped with an additional
108 * freelist that allows lockless access to
109 * free objects in addition to the regular freelist
110 * that requires the slab lock.
111 *
112 * SLAB_DEBUG_FLAGS Slab requires special handling due to debug
113 * options set. This moves slab handling out of
114 * the fast path and disables lockless freelists.
115 */
116
117#ifdef CONFIG_SLUB_DEBUG
118#ifdef CONFIG_SLUB_DEBUG_ON
119DEFINE_STATIC_KEY_TRUE(slub_debug_enabled);
120#else
121DEFINE_STATIC_KEY_FALSE(slub_debug_enabled);
122#endif
123#endif
124
125static inline bool kmem_cache_debug(struct kmem_cache *s)
126{
127 return kmem_cache_debug_flags(s, SLAB_DEBUG_FLAGS);
128}
129
130void *fixup_red_left(struct kmem_cache *s, void *p)
131{
132 if (kmem_cache_debug_flags(s, SLAB_RED_ZONE))
133 p += s->red_left_pad;
134
135 return p;
136}
137
138static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
139{
140#ifdef CONFIG_SLUB_CPU_PARTIAL
141 return !kmem_cache_debug(s);
142#else
143 return false;
144#endif
145}
146
147/*
148 * Issues still to be resolved:
149 *
150 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
151 *
152 * - Variable sizing of the per node arrays
153 */
154
155/* Enable to test recovery from slab corruption on boot */
156#undef SLUB_RESILIENCY_TEST
157
158/* Enable to log cmpxchg failures */
159#undef SLUB_DEBUG_CMPXCHG
160
161/*
162 * Mininum number of partial slabs. These will be left on the partial
163 * lists even if they are empty. kmem_cache_shrink may reclaim them.
164 */
165#define MIN_PARTIAL 5
166
167/*
168 * Maximum number of desirable partial slabs.
169 * The existence of more partial slabs makes kmem_cache_shrink
170 * sort the partial list by the number of objects in use.
171 */
172#define MAX_PARTIAL 10
173
174#define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
175 SLAB_POISON | SLAB_STORE_USER)
176
177/*
178 * These debug flags cannot use CMPXCHG because there might be consistency
179 * issues when checking or reading debug information
180 */
181#define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
182 SLAB_TRACE)
183
184
185/*
186 * Debugging flags that require metadata to be stored in the slab. These get
187 * disabled when slub_debug=O is used and a cache's min order increases with
188 * metadata.
189 */
190#define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
191
192#define OO_SHIFT 16
193#define OO_MASK ((1 << OO_SHIFT) - 1)
194#define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
195
196/* Internal SLUB flags */
197/* Poison object */
198#define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
199/* Use cmpxchg_double */
200#define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
201
202/*
203 * Tracking user of a slab.
204 */
205#define TRACK_ADDRS_COUNT 16
206struct track {
207 unsigned long addr; /* Called from address */
208#ifdef CONFIG_STACKTRACE
209 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
210#endif
211 int cpu; /* Was running on cpu */
212 int pid; /* Pid context */
213 unsigned long when; /* When did the operation occur */
214};
215
216enum track_item { TRACK_ALLOC, TRACK_FREE };
217
218#ifdef CONFIG_SYSFS
219static int sysfs_slab_add(struct kmem_cache *);
220static int sysfs_slab_alias(struct kmem_cache *, const char *);
221#else
222static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
223static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
224 { return 0; }
225#endif
226
227static inline void stat(const struct kmem_cache *s, enum stat_item si)
228{
229#ifdef CONFIG_SLUB_STATS
230 /*
231 * The rmw is racy on a preemptible kernel but this is acceptable, so
232 * avoid this_cpu_add()'s irq-disable overhead.
233 */
234 raw_cpu_inc(s->cpu_slab->stat[si]);
235#endif
236}
237
238/********************************************************************
239 * Core slab cache functions
240 *******************************************************************/
241
242/*
243 * Returns freelist pointer (ptr). With hardening, this is obfuscated
244 * with an XOR of the address where the pointer is held and a per-cache
245 * random number.
246 */
247static inline void *freelist_ptr(const struct kmem_cache *s, void *ptr,
248 unsigned long ptr_addr)
249{
250#ifdef CONFIG_SLAB_FREELIST_HARDENED
251 /*
252 * When CONFIG_KASAN_SW_TAGS is enabled, ptr_addr might be tagged.
253 * Normally, this doesn't cause any issues, as both set_freepointer()
254 * and get_freepointer() are called with a pointer with the same tag.
255 * However, there are some issues with CONFIG_SLUB_DEBUG code. For
256 * example, when __free_slub() iterates over objects in a cache, it
257 * passes untagged pointers to check_object(). check_object() in turns
258 * calls get_freepointer() with an untagged pointer, which causes the
259 * freepointer to be restored incorrectly.
260 */
261 return (void *)((unsigned long)ptr ^ s->random ^
262 swab((unsigned long)kasan_reset_tag((void *)ptr_addr)));
263#else
264 return ptr;
265#endif
266}
267
268/* Returns the freelist pointer recorded at location ptr_addr. */
269static inline void *freelist_dereference(const struct kmem_cache *s,
270 void *ptr_addr)
271{
272 return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr),
273 (unsigned long)ptr_addr);
274}
275
276static inline void *get_freepointer(struct kmem_cache *s, void *object)
277{
278 return freelist_dereference(s, object + s->offset);
279}
280
281static void prefetch_freepointer(const struct kmem_cache *s, void *object)
282{
283 prefetch(object + s->offset);
284}
285
286static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
287{
288 unsigned long freepointer_addr;
289 void *p;
290
291 if (!debug_pagealloc_enabled_static())
292 return get_freepointer(s, object);
293
294 freepointer_addr = (unsigned long)object + s->offset;
295 copy_from_kernel_nofault(&p, (void **)freepointer_addr, sizeof(p));
296 return freelist_ptr(s, p, freepointer_addr);
297}
298
299static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
300{
301 unsigned long freeptr_addr = (unsigned long)object + s->offset;
302
303#ifdef CONFIG_SLAB_FREELIST_HARDENED
304 BUG_ON(object == fp); /* naive detection of double free or corruption */
305#endif
306
307 *(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr);
308}
309
310/* Loop over all objects in a slab */
311#define for_each_object(__p, __s, __addr, __objects) \
312 for (__p = fixup_red_left(__s, __addr); \
313 __p < (__addr) + (__objects) * (__s)->size; \
314 __p += (__s)->size)
315
316static inline unsigned int order_objects(unsigned int order, unsigned int size)
317{
318 return ((unsigned int)PAGE_SIZE << order) / size;
319}
320
321static inline struct kmem_cache_order_objects oo_make(unsigned int order,
322 unsigned int size)
323{
324 struct kmem_cache_order_objects x = {
325 (order << OO_SHIFT) + order_objects(order, size)
326 };
327
328 return x;
329}
330
331static inline unsigned int oo_order(struct kmem_cache_order_objects x)
332{
333 return x.x >> OO_SHIFT;
334}
335
336static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
337{
338 return x.x & OO_MASK;
339}
340
341/*
342 * Per slab locking using the pagelock
343 */
344static __always_inline void slab_lock(struct page *page)
345{
346 VM_BUG_ON_PAGE(PageTail(page), page);
347 bit_spin_lock(PG_locked, &page->flags);
348}
349
350static __always_inline void slab_unlock(struct page *page)
351{
352 VM_BUG_ON_PAGE(PageTail(page), page);
353 __bit_spin_unlock(PG_locked, &page->flags);
354}
355
356/* Interrupts must be disabled (for the fallback code to work right) */
357static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
358 void *freelist_old, unsigned long counters_old,
359 void *freelist_new, unsigned long counters_new,
360 const char *n)
361{
362 VM_BUG_ON(!irqs_disabled());
363#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
364 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
365 if (s->flags & __CMPXCHG_DOUBLE) {
366 if (cmpxchg_double(&page->freelist, &page->counters,
367 freelist_old, counters_old,
368 freelist_new, counters_new))
369 return true;
370 } else
371#endif
372 {
373 slab_lock(page);
374 if (page->freelist == freelist_old &&
375 page->counters == counters_old) {
376 page->freelist = freelist_new;
377 page->counters = counters_new;
378 slab_unlock(page);
379 return true;
380 }
381 slab_unlock(page);
382 }
383
384 cpu_relax();
385 stat(s, CMPXCHG_DOUBLE_FAIL);
386
387#ifdef SLUB_DEBUG_CMPXCHG
388 pr_info("%s %s: cmpxchg double redo ", n, s->name);
389#endif
390
391 return false;
392}
393
394static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
395 void *freelist_old, unsigned long counters_old,
396 void *freelist_new, unsigned long counters_new,
397 const char *n)
398{
399#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
400 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
401 if (s->flags & __CMPXCHG_DOUBLE) {
402 if (cmpxchg_double(&page->freelist, &page->counters,
403 freelist_old, counters_old,
404 freelist_new, counters_new))
405 return true;
406 } else
407#endif
408 {
409 unsigned long flags;
410
411 local_irq_save(flags);
412 slab_lock(page);
413 if (page->freelist == freelist_old &&
414 page->counters == counters_old) {
415 page->freelist = freelist_new;
416 page->counters = counters_new;
417 slab_unlock(page);
418 local_irq_restore(flags);
419 return true;
420 }
421 slab_unlock(page);
422 local_irq_restore(flags);
423 }
424
425 cpu_relax();
426 stat(s, CMPXCHG_DOUBLE_FAIL);
427
428#ifdef SLUB_DEBUG_CMPXCHG
429 pr_info("%s %s: cmpxchg double redo ", n, s->name);
430#endif
431
432 return false;
433}
434
435#ifdef CONFIG_SLUB_DEBUG
436static unsigned long object_map[BITS_TO_LONGS(MAX_OBJS_PER_PAGE)];
437static DEFINE_SPINLOCK(object_map_lock);
438
439/*
440 * Determine a map of object in use on a page.
441 *
442 * Node listlock must be held to guarantee that the page does
443 * not vanish from under us.
444 */
445static unsigned long *get_map(struct kmem_cache *s, struct page *page)
446 __acquires(&object_map_lock)
447{
448 void *p;
449 void *addr = page_address(page);
450
451 VM_BUG_ON(!irqs_disabled());
452
453 spin_lock(&object_map_lock);
454
455 bitmap_zero(object_map, page->objects);
456
457 for (p = page->freelist; p; p = get_freepointer(s, p))
458 set_bit(__obj_to_index(s, addr, p), object_map);
459
460 return object_map;
461}
462
463static void put_map(unsigned long *map) __releases(&object_map_lock)
464{
465 VM_BUG_ON(map != object_map);
466 spin_unlock(&object_map_lock);
467}
468
469static inline unsigned int size_from_object(struct kmem_cache *s)
470{
471 if (s->flags & SLAB_RED_ZONE)
472 return s->size - s->red_left_pad;
473
474 return s->size;
475}
476
477static inline void *restore_red_left(struct kmem_cache *s, void *p)
478{
479 if (s->flags & SLAB_RED_ZONE)
480 p -= s->red_left_pad;
481
482 return p;
483}
484
485/*
486 * Debug settings:
487 */
488#if defined(CONFIG_SLUB_DEBUG_ON)
489static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
490#else
491static slab_flags_t slub_debug;
492#endif
493
494static char *slub_debug_string;
495static int disable_higher_order_debug;
496
497/*
498 * slub is about to manipulate internal object metadata. This memory lies
499 * outside the range of the allocated object, so accessing it would normally
500 * be reported by kasan as a bounds error. metadata_access_enable() is used
501 * to tell kasan that these accesses are OK.
502 */
503static inline void metadata_access_enable(void)
504{
505 kasan_disable_current();
506}
507
508static inline void metadata_access_disable(void)
509{
510 kasan_enable_current();
511}
512
513/*
514 * Object debugging
515 */
516
517/* Verify that a pointer has an address that is valid within a slab page */
518static inline int check_valid_pointer(struct kmem_cache *s,
519 struct page *page, void *object)
520{
521 void *base;
522
523 if (!object)
524 return 1;
525
526 base = page_address(page);
527 object = kasan_reset_tag(object);
528 object = restore_red_left(s, object);
529 if (object < base || object >= base + page->objects * s->size ||
530 (object - base) % s->size) {
531 return 0;
532 }
533
534 return 1;
535}
536
537static void print_section(char *level, char *text, u8 *addr,
538 unsigned int length)
539{
540 metadata_access_enable();
541 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
542 length, 1);
543 metadata_access_disable();
544}
545
546/*
547 * See comment in calculate_sizes().
548 */
549static inline bool freeptr_outside_object(struct kmem_cache *s)
550{
551 return s->offset >= s->inuse;
552}
553
554/*
555 * Return offset of the end of info block which is inuse + free pointer if
556 * not overlapping with object.
557 */
558static inline unsigned int get_info_end(struct kmem_cache *s)
559{
560 if (freeptr_outside_object(s))
561 return s->inuse + sizeof(void *);
562 else
563 return s->inuse;
564}
565
566static struct track *get_track(struct kmem_cache *s, void *object,
567 enum track_item alloc)
568{
569 struct track *p;
570
571 p = object + get_info_end(s);
572
573 return p + alloc;
574}
575
576static void set_track(struct kmem_cache *s, void *object,
577 enum track_item alloc, unsigned long addr)
578{
579 struct track *p = get_track(s, object, alloc);
580
581 if (addr) {
582#ifdef CONFIG_STACKTRACE
583 unsigned int nr_entries;
584
585 metadata_access_enable();
586 nr_entries = stack_trace_save(p->addrs, TRACK_ADDRS_COUNT, 3);
587 metadata_access_disable();
588
589 if (nr_entries < TRACK_ADDRS_COUNT)
590 p->addrs[nr_entries] = 0;
591#endif
592 p->addr = addr;
593 p->cpu = smp_processor_id();
594 p->pid = current->pid;
595 p->when = jiffies;
596 } else {
597 memset(p, 0, sizeof(struct track));
598 }
599}
600
601static void init_tracking(struct kmem_cache *s, void *object)
602{
603 if (!(s->flags & SLAB_STORE_USER))
604 return;
605
606 set_track(s, object, TRACK_FREE, 0UL);
607 set_track(s, object, TRACK_ALLOC, 0UL);
608}
609
610static void print_track(const char *s, struct track *t, unsigned long pr_time)
611{
612 if (!t->addr)
613 return;
614
615 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
616 s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
617#ifdef CONFIG_STACKTRACE
618 {
619 int i;
620 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
621 if (t->addrs[i])
622 pr_err("\t%pS\n", (void *)t->addrs[i]);
623 else
624 break;
625 }
626#endif
627}
628
629void print_tracking(struct kmem_cache *s, void *object)
630{
631 unsigned long pr_time = jiffies;
632 if (!(s->flags & SLAB_STORE_USER))
633 return;
634
635 print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
636 print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
637}
638
639static void print_page_info(struct page *page)
640{
641 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
642 page, page->objects, page->inuse, page->freelist, page->flags);
643
644}
645
646static void slab_bug(struct kmem_cache *s, char *fmt, ...)
647{
648 struct va_format vaf;
649 va_list args;
650
651 va_start(args, fmt);
652 vaf.fmt = fmt;
653 vaf.va = &args;
654 pr_err("=============================================================================\n");
655 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
656 pr_err("-----------------------------------------------------------------------------\n\n");
657
658 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
659 va_end(args);
660}
661
662static void slab_fix(struct kmem_cache *s, char *fmt, ...)
663{
664 struct va_format vaf;
665 va_list args;
666
667 va_start(args, fmt);
668 vaf.fmt = fmt;
669 vaf.va = &args;
670 pr_err("FIX %s: %pV\n", s->name, &vaf);
671 va_end(args);
672}
673
674static bool freelist_corrupted(struct kmem_cache *s, struct page *page,
675 void **freelist, void *nextfree)
676{
677 if ((s->flags & SLAB_CONSISTENCY_CHECKS) &&
678 !check_valid_pointer(s, page, nextfree) && freelist) {
679 object_err(s, page, *freelist, "Freechain corrupt");
680 *freelist = NULL;
681 slab_fix(s, "Isolate corrupted freechain");
682 return true;
683 }
684
685 return false;
686}
687
688static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
689{
690 unsigned int off; /* Offset of last byte */
691 u8 *addr = page_address(page);
692
693 print_tracking(s, p);
694
695 print_page_info(page);
696
697 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
698 p, p - addr, get_freepointer(s, p));
699
700 if (s->flags & SLAB_RED_ZONE)
701 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
702 s->red_left_pad);
703 else if (p > addr + 16)
704 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
705
706 print_section(KERN_ERR, "Object ", p,
707 min_t(unsigned int, s->object_size, PAGE_SIZE));
708 if (s->flags & SLAB_RED_ZONE)
709 print_section(KERN_ERR, "Redzone ", p + s->object_size,
710 s->inuse - s->object_size);
711
712 off = get_info_end(s);
713
714 if (s->flags & SLAB_STORE_USER)
715 off += 2 * sizeof(struct track);
716
717 off += kasan_metadata_size(s);
718
719 if (off != size_from_object(s))
720 /* Beginning of the filler is the free pointer */
721 print_section(KERN_ERR, "Padding ", p + off,
722 size_from_object(s) - off);
723
724 dump_stack();
725}
726
727void object_err(struct kmem_cache *s, struct page *page,
728 u8 *object, char *reason)
729{
730 slab_bug(s, "%s", reason);
731 print_trailer(s, page, object);
732}
733
734static __printf(3, 4) void slab_err(struct kmem_cache *s, struct page *page,
735 const char *fmt, ...)
736{
737 va_list args;
738 char buf[100];
739
740 va_start(args, fmt);
741 vsnprintf(buf, sizeof(buf), fmt, args);
742 va_end(args);
743 slab_bug(s, "%s", buf);
744 print_page_info(page);
745 dump_stack();
746}
747
748static void init_object(struct kmem_cache *s, void *object, u8 val)
749{
750 u8 *p = object;
751
752 if (s->flags & SLAB_RED_ZONE)
753 memset(p - s->red_left_pad, val, s->red_left_pad);
754
755 if (s->flags & __OBJECT_POISON) {
756 memset(p, POISON_FREE, s->object_size - 1);
757 p[s->object_size - 1] = POISON_END;
758 }
759
760 if (s->flags & SLAB_RED_ZONE)
761 memset(p + s->object_size, val, s->inuse - s->object_size);
762}
763
764static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
765 void *from, void *to)
766{
767 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
768 memset(from, data, to - from);
769}
770
771static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
772 u8 *object, char *what,
773 u8 *start, unsigned int value, unsigned int bytes)
774{
775 u8 *fault;
776 u8 *end;
777 u8 *addr = page_address(page);
778
779 metadata_access_enable();
780 fault = memchr_inv(start, value, bytes);
781 metadata_access_disable();
782 if (!fault)
783 return 1;
784
785 end = start + bytes;
786 while (end > fault && end[-1] == value)
787 end--;
788
789 slab_bug(s, "%s overwritten", what);
790 pr_err("INFO: 0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
791 fault, end - 1, fault - addr,
792 fault[0], value);
793 print_trailer(s, page, object);
794
795 restore_bytes(s, what, value, fault, end);
796 return 0;
797}
798
799/*
800 * Object layout:
801 *
802 * object address
803 * Bytes of the object to be managed.
804 * If the freepointer may overlay the object then the free
805 * pointer is at the middle of the object.
806 *
807 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
808 * 0xa5 (POISON_END)
809 *
810 * object + s->object_size
811 * Padding to reach word boundary. This is also used for Redzoning.
812 * Padding is extended by another word if Redzoning is enabled and
813 * object_size == inuse.
814 *
815 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
816 * 0xcc (RED_ACTIVE) for objects in use.
817 *
818 * object + s->inuse
819 * Meta data starts here.
820 *
821 * A. Free pointer (if we cannot overwrite object on free)
822 * B. Tracking data for SLAB_STORE_USER
823 * C. Padding to reach required alignment boundary or at mininum
824 * one word if debugging is on to be able to detect writes
825 * before the word boundary.
826 *
827 * Padding is done using 0x5a (POISON_INUSE)
828 *
829 * object + s->size
830 * Nothing is used beyond s->size.
831 *
832 * If slabcaches are merged then the object_size and inuse boundaries are mostly
833 * ignored. And therefore no slab options that rely on these boundaries
834 * may be used with merged slabcaches.
835 */
836
837static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
838{
839 unsigned long off = get_info_end(s); /* The end of info */
840
841 if (s->flags & SLAB_STORE_USER)
842 /* We also have user information there */
843 off += 2 * sizeof(struct track);
844
845 off += kasan_metadata_size(s);
846
847 if (size_from_object(s) == off)
848 return 1;
849
850 return check_bytes_and_report(s, page, p, "Object padding",
851 p + off, POISON_INUSE, size_from_object(s) - off);
852}
853
854/* Check the pad bytes at the end of a slab page */
855static int slab_pad_check(struct kmem_cache *s, struct page *page)
856{
857 u8 *start;
858 u8 *fault;
859 u8 *end;
860 u8 *pad;
861 int length;
862 int remainder;
863
864 if (!(s->flags & SLAB_POISON))
865 return 1;
866
867 start = page_address(page);
868 length = page_size(page);
869 end = start + length;
870 remainder = length % s->size;
871 if (!remainder)
872 return 1;
873
874 pad = end - remainder;
875 metadata_access_enable();
876 fault = memchr_inv(pad, POISON_INUSE, remainder);
877 metadata_access_disable();
878 if (!fault)
879 return 1;
880 while (end > fault && end[-1] == POISON_INUSE)
881 end--;
882
883 slab_err(s, page, "Padding overwritten. 0x%p-0x%p @offset=%tu",
884 fault, end - 1, fault - start);
885 print_section(KERN_ERR, "Padding ", pad, remainder);
886
887 restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
888 return 0;
889}
890
891static int check_object(struct kmem_cache *s, struct page *page,
892 void *object, u8 val)
893{
894 u8 *p = object;
895 u8 *endobject = object + s->object_size;
896
897 if (s->flags & SLAB_RED_ZONE) {
898 if (!check_bytes_and_report(s, page, object, "Redzone",
899 object - s->red_left_pad, val, s->red_left_pad))
900 return 0;
901
902 if (!check_bytes_and_report(s, page, object, "Redzone",
903 endobject, val, s->inuse - s->object_size))
904 return 0;
905 } else {
906 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
907 check_bytes_and_report(s, page, p, "Alignment padding",
908 endobject, POISON_INUSE,
909 s->inuse - s->object_size);
910 }
911 }
912
913 if (s->flags & SLAB_POISON) {
914 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
915 (!check_bytes_and_report(s, page, p, "Poison", p,
916 POISON_FREE, s->object_size - 1) ||
917 !check_bytes_and_report(s, page, p, "Poison",
918 p + s->object_size - 1, POISON_END, 1)))
919 return 0;
920 /*
921 * check_pad_bytes cleans up on its own.
922 */
923 check_pad_bytes(s, page, p);
924 }
925
926 if (!freeptr_outside_object(s) && val == SLUB_RED_ACTIVE)
927 /*
928 * Object and freepointer overlap. Cannot check
929 * freepointer while object is allocated.
930 */
931 return 1;
932
933 /* Check free pointer validity */
934 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
935 object_err(s, page, p, "Freepointer corrupt");
936 /*
937 * No choice but to zap it and thus lose the remainder
938 * of the free objects in this slab. May cause
939 * another error because the object count is now wrong.
940 */
941 set_freepointer(s, p, NULL);
942 return 0;
943 }
944 return 1;
945}
946
947static int check_slab(struct kmem_cache *s, struct page *page)
948{
949 int maxobj;
950
951 VM_BUG_ON(!irqs_disabled());
952
953 if (!PageSlab(page)) {
954 slab_err(s, page, "Not a valid slab page");
955 return 0;
956 }
957
958 maxobj = order_objects(compound_order(page), s->size);
959 if (page->objects > maxobj) {
960 slab_err(s, page, "objects %u > max %u",
961 page->objects, maxobj);
962 return 0;
963 }
964 if (page->inuse > page->objects) {
965 slab_err(s, page, "inuse %u > max %u",
966 page->inuse, page->objects);
967 return 0;
968 }
969 /* Slab_pad_check fixes things up after itself */
970 slab_pad_check(s, page);
971 return 1;
972}
973
974/*
975 * Determine if a certain object on a page is on the freelist. Must hold the
976 * slab lock to guarantee that the chains are in a consistent state.
977 */
978static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
979{
980 int nr = 0;
981 void *fp;
982 void *object = NULL;
983 int max_objects;
984
985 fp = page->freelist;
986 while (fp && nr <= page->objects) {
987 if (fp == search)
988 return 1;
989 if (!check_valid_pointer(s, page, fp)) {
990 if (object) {
991 object_err(s, page, object,
992 "Freechain corrupt");
993 set_freepointer(s, object, NULL);
994 } else {
995 slab_err(s, page, "Freepointer corrupt");
996 page->freelist = NULL;
997 page->inuse = page->objects;
998 slab_fix(s, "Freelist cleared");
999 return 0;
1000 }
1001 break;
1002 }
1003 object = fp;
1004 fp = get_freepointer(s, object);
1005 nr++;
1006 }
1007
1008 max_objects = order_objects(compound_order(page), s->size);
1009 if (max_objects > MAX_OBJS_PER_PAGE)
1010 max_objects = MAX_OBJS_PER_PAGE;
1011
1012 if (page->objects != max_objects) {
1013 slab_err(s, page, "Wrong number of objects. Found %d but should be %d",
1014 page->objects, max_objects);
1015 page->objects = max_objects;
1016 slab_fix(s, "Number of objects adjusted.");
1017 }
1018 if (page->inuse != page->objects - nr) {
1019 slab_err(s, page, "Wrong object count. Counter is %d but counted were %d",
1020 page->inuse, page->objects - nr);
1021 page->inuse = page->objects - nr;
1022 slab_fix(s, "Object count adjusted.");
1023 }
1024 return search == NULL;
1025}
1026
1027static void trace(struct kmem_cache *s, struct page *page, void *object,
1028 int alloc)
1029{
1030 if (s->flags & SLAB_TRACE) {
1031 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1032 s->name,
1033 alloc ? "alloc" : "free",
1034 object, page->inuse,
1035 page->freelist);
1036
1037 if (!alloc)
1038 print_section(KERN_INFO, "Object ", (void *)object,
1039 s->object_size);
1040
1041 dump_stack();
1042 }
1043}
1044
1045/*
1046 * Tracking of fully allocated slabs for debugging purposes.
1047 */
1048static void add_full(struct kmem_cache *s,
1049 struct kmem_cache_node *n, struct page *page)
1050{
1051 if (!(s->flags & SLAB_STORE_USER))
1052 return;
1053
1054 lockdep_assert_held(&n->list_lock);
1055 list_add(&page->slab_list, &n->full);
1056}
1057
1058static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
1059{
1060 if (!(s->flags & SLAB_STORE_USER))
1061 return;
1062
1063 lockdep_assert_held(&n->list_lock);
1064 list_del(&page->slab_list);
1065}
1066
1067/* Tracking of the number of slabs for debugging purposes */
1068static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1069{
1070 struct kmem_cache_node *n = get_node(s, node);
1071
1072 return atomic_long_read(&n->nr_slabs);
1073}
1074
1075static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1076{
1077 return atomic_long_read(&n->nr_slabs);
1078}
1079
1080static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1081{
1082 struct kmem_cache_node *n = get_node(s, node);
1083
1084 /*
1085 * May be called early in order to allocate a slab for the
1086 * kmem_cache_node structure. Solve the chicken-egg
1087 * dilemma by deferring the increment of the count during
1088 * bootstrap (see early_kmem_cache_node_alloc).
1089 */
1090 if (likely(n)) {
1091 atomic_long_inc(&n->nr_slabs);
1092 atomic_long_add(objects, &n->total_objects);
1093 }
1094}
1095static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1096{
1097 struct kmem_cache_node *n = get_node(s, node);
1098
1099 atomic_long_dec(&n->nr_slabs);
1100 atomic_long_sub(objects, &n->total_objects);
1101}
1102
1103/* Object debug checks for alloc/free paths */
1104static void setup_object_debug(struct kmem_cache *s, struct page *page,
1105 void *object)
1106{
1107 if (!kmem_cache_debug_flags(s, SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))
1108 return;
1109
1110 init_object(s, object, SLUB_RED_INACTIVE);
1111 init_tracking(s, object);
1112}
1113
1114static
1115void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr)
1116{
1117 if (!kmem_cache_debug_flags(s, SLAB_POISON))
1118 return;
1119
1120 metadata_access_enable();
1121 memset(addr, POISON_INUSE, page_size(page));
1122 metadata_access_disable();
1123}
1124
1125static inline int alloc_consistency_checks(struct kmem_cache *s,
1126 struct page *page, void *object)
1127{
1128 if (!check_slab(s, page))
1129 return 0;
1130
1131 if (!check_valid_pointer(s, page, object)) {
1132 object_err(s, page, object, "Freelist Pointer check fails");
1133 return 0;
1134 }
1135
1136 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1137 return 0;
1138
1139 return 1;
1140}
1141
1142static noinline int alloc_debug_processing(struct kmem_cache *s,
1143 struct page *page,
1144 void *object, unsigned long addr)
1145{
1146 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1147 if (!alloc_consistency_checks(s, page, object))
1148 goto bad;
1149 }
1150
1151 /* Success perform special debug activities for allocs */
1152 if (s->flags & SLAB_STORE_USER)
1153 set_track(s, object, TRACK_ALLOC, addr);
1154 trace(s, page, object, 1);
1155 init_object(s, object, SLUB_RED_ACTIVE);
1156 return 1;
1157
1158bad:
1159 if (PageSlab(page)) {
1160 /*
1161 * If this is a slab page then lets do the best we can
1162 * to avoid issues in the future. Marking all objects
1163 * as used avoids touching the remaining objects.
1164 */
1165 slab_fix(s, "Marking all objects used");
1166 page->inuse = page->objects;
1167 page->freelist = NULL;
1168 }
1169 return 0;
1170}
1171
1172static inline int free_consistency_checks(struct kmem_cache *s,
1173 struct page *page, void *object, unsigned long addr)
1174{
1175 if (!check_valid_pointer(s, page, object)) {
1176 slab_err(s, page, "Invalid object pointer 0x%p", object);
1177 return 0;
1178 }
1179
1180 if (on_freelist(s, page, object)) {
1181 object_err(s, page, object, "Object already free");
1182 return 0;
1183 }
1184
1185 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1186 return 0;
1187
1188 if (unlikely(s != page->slab_cache)) {
1189 if (!PageSlab(page)) {
1190 slab_err(s, page, "Attempt to free object(0x%p) outside of slab",
1191 object);
1192 } else if (!page->slab_cache) {
1193 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1194 object);
1195 dump_stack();
1196 } else
1197 object_err(s, page, object,
1198 "page slab pointer corrupt.");
1199 return 0;
1200 }
1201 return 1;
1202}
1203
1204/* Supports checking bulk free of a constructed freelist */
1205static noinline int free_debug_processing(
1206 struct kmem_cache *s, struct page *page,
1207 void *head, void *tail, int bulk_cnt,
1208 unsigned long addr)
1209{
1210 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1211 void *object = head;
1212 int cnt = 0;
1213 unsigned long flags;
1214 int ret = 0;
1215
1216 spin_lock_irqsave(&n->list_lock, flags);
1217 slab_lock(page);
1218
1219 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1220 if (!check_slab(s, page))
1221 goto out;
1222 }
1223
1224next_object:
1225 cnt++;
1226
1227 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1228 if (!free_consistency_checks(s, page, object, addr))
1229 goto out;
1230 }
1231
1232 if (s->flags & SLAB_STORE_USER)
1233 set_track(s, object, TRACK_FREE, addr);
1234 trace(s, page, object, 0);
1235 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1236 init_object(s, object, SLUB_RED_INACTIVE);
1237
1238 /* Reached end of constructed freelist yet? */
1239 if (object != tail) {
1240 object = get_freepointer(s, object);
1241 goto next_object;
1242 }
1243 ret = 1;
1244
1245out:
1246 if (cnt != bulk_cnt)
1247 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1248 bulk_cnt, cnt);
1249
1250 slab_unlock(page);
1251 spin_unlock_irqrestore(&n->list_lock, flags);
1252 if (!ret)
1253 slab_fix(s, "Object at 0x%p not freed", object);
1254 return ret;
1255}
1256
1257/*
1258 * Parse a block of slub_debug options. Blocks are delimited by ';'
1259 *
1260 * @str: start of block
1261 * @flags: returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified
1262 * @slabs: return start of list of slabs, or NULL when there's no list
1263 * @init: assume this is initial parsing and not per-kmem-create parsing
1264 *
1265 * returns the start of next block if there's any, or NULL
1266 */
1267static char *
1268parse_slub_debug_flags(char *str, slab_flags_t *flags, char **slabs, bool init)
1269{
1270 bool higher_order_disable = false;
1271
1272 /* Skip any completely empty blocks */
1273 while (*str && *str == ';')
1274 str++;
1275
1276 if (*str == ',') {
1277 /*
1278 * No options but restriction on slabs. This means full
1279 * debugging for slabs matching a pattern.
1280 */
1281 *flags = DEBUG_DEFAULT_FLAGS;
1282 goto check_slabs;
1283 }
1284 *flags = 0;
1285
1286 /* Determine which debug features should be switched on */
1287 for (; *str && *str != ',' && *str != ';'; str++) {
1288 switch (tolower(*str)) {
1289 case '-':
1290 *flags = 0;
1291 break;
1292 case 'f':
1293 *flags |= SLAB_CONSISTENCY_CHECKS;
1294 break;
1295 case 'z':
1296 *flags |= SLAB_RED_ZONE;
1297 break;
1298 case 'p':
1299 *flags |= SLAB_POISON;
1300 break;
1301 case 'u':
1302 *flags |= SLAB_STORE_USER;
1303 break;
1304 case 't':
1305 *flags |= SLAB_TRACE;
1306 break;
1307 case 'a':
1308 *flags |= SLAB_FAILSLAB;
1309 break;
1310 case 'o':
1311 /*
1312 * Avoid enabling debugging on caches if its minimum
1313 * order would increase as a result.
1314 */
1315 higher_order_disable = true;
1316 break;
1317 default:
1318 if (init)
1319 pr_err("slub_debug option '%c' unknown. skipped\n", *str);
1320 }
1321 }
1322check_slabs:
1323 if (*str == ',')
1324 *slabs = ++str;
1325 else
1326 *slabs = NULL;
1327
1328 /* Skip over the slab list */
1329 while (*str && *str != ';')
1330 str++;
1331
1332 /* Skip any completely empty blocks */
1333 while (*str && *str == ';')
1334 str++;
1335
1336 if (init && higher_order_disable)
1337 disable_higher_order_debug = 1;
1338
1339 if (*str)
1340 return str;
1341 else
1342 return NULL;
1343}
1344
1345static int __init setup_slub_debug(char *str)
1346{
1347 slab_flags_t flags;
1348 char *saved_str;
1349 char *slab_list;
1350 bool global_slub_debug_changed = false;
1351 bool slab_list_specified = false;
1352
1353 slub_debug = DEBUG_DEFAULT_FLAGS;
1354 if (*str++ != '=' || !*str)
1355 /*
1356 * No options specified. Switch on full debugging.
1357 */
1358 goto out;
1359
1360 saved_str = str;
1361 while (str) {
1362 str = parse_slub_debug_flags(str, &flags, &slab_list, true);
1363
1364 if (!slab_list) {
1365 slub_debug = flags;
1366 global_slub_debug_changed = true;
1367 } else {
1368 slab_list_specified = true;
1369 }
1370 }
1371
1372 /*
1373 * For backwards compatibility, a single list of flags with list of
1374 * slabs means debugging is only enabled for those slabs, so the global
1375 * slub_debug should be 0. We can extended that to multiple lists as
1376 * long as there is no option specifying flags without a slab list.
1377 */
1378 if (slab_list_specified) {
1379 if (!global_slub_debug_changed)
1380 slub_debug = 0;
1381 slub_debug_string = saved_str;
1382 }
1383out:
1384 if (slub_debug != 0 || slub_debug_string)
1385 static_branch_enable(&slub_debug_enabled);
1386 if ((static_branch_unlikely(&init_on_alloc) ||
1387 static_branch_unlikely(&init_on_free)) &&
1388 (slub_debug & SLAB_POISON))
1389 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1390 return 1;
1391}
1392
1393__setup("slub_debug", setup_slub_debug);
1394
1395/*
1396 * kmem_cache_flags - apply debugging options to the cache
1397 * @object_size: the size of an object without meta data
1398 * @flags: flags to set
1399 * @name: name of the cache
1400 * @ctor: constructor function
1401 *
1402 * Debug option(s) are applied to @flags. In addition to the debug
1403 * option(s), if a slab name (or multiple) is specified i.e.
1404 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1405 * then only the select slabs will receive the debug option(s).
1406 */
1407slab_flags_t kmem_cache_flags(unsigned int object_size,
1408 slab_flags_t flags, const char *name,
1409 void (*ctor)(void *))
1410{
1411 char *iter;
1412 size_t len;
1413 char *next_block;
1414 slab_flags_t block_flags;
1415
1416 len = strlen(name);
1417 next_block = slub_debug_string;
1418 /* Go through all blocks of debug options, see if any matches our slab's name */
1419 while (next_block) {
1420 next_block = parse_slub_debug_flags(next_block, &block_flags, &iter, false);
1421 if (!iter)
1422 continue;
1423 /* Found a block that has a slab list, search it */
1424 while (*iter) {
1425 char *end, *glob;
1426 size_t cmplen;
1427
1428 end = strchrnul(iter, ',');
1429 if (next_block && next_block < end)
1430 end = next_block - 1;
1431
1432 glob = strnchr(iter, end - iter, '*');
1433 if (glob)
1434 cmplen = glob - iter;
1435 else
1436 cmplen = max_t(size_t, len, (end - iter));
1437
1438 if (!strncmp(name, iter, cmplen)) {
1439 flags |= block_flags;
1440 return flags;
1441 }
1442
1443 if (!*end || *end == ';')
1444 break;
1445 iter = end + 1;
1446 }
1447 }
1448
1449 return flags | slub_debug;
1450}
1451#else /* !CONFIG_SLUB_DEBUG */
1452static inline void setup_object_debug(struct kmem_cache *s,
1453 struct page *page, void *object) {}
1454static inline
1455void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr) {}
1456
1457static inline int alloc_debug_processing(struct kmem_cache *s,
1458 struct page *page, void *object, unsigned long addr) { return 0; }
1459
1460static inline int free_debug_processing(
1461 struct kmem_cache *s, struct page *page,
1462 void *head, void *tail, int bulk_cnt,
1463 unsigned long addr) { return 0; }
1464
1465static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1466 { return 1; }
1467static inline int check_object(struct kmem_cache *s, struct page *page,
1468 void *object, u8 val) { return 1; }
1469static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1470 struct page *page) {}
1471static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1472 struct page *page) {}
1473slab_flags_t kmem_cache_flags(unsigned int object_size,
1474 slab_flags_t flags, const char *name,
1475 void (*ctor)(void *))
1476{
1477 return flags;
1478}
1479#define slub_debug 0
1480
1481#define disable_higher_order_debug 0
1482
1483static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1484 { return 0; }
1485static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1486 { return 0; }
1487static inline void inc_slabs_node(struct kmem_cache *s, int node,
1488 int objects) {}
1489static inline void dec_slabs_node(struct kmem_cache *s, int node,
1490 int objects) {}
1491
1492static bool freelist_corrupted(struct kmem_cache *s, struct page *page,
1493 void **freelist, void *nextfree)
1494{
1495 return false;
1496}
1497#endif /* CONFIG_SLUB_DEBUG */
1498
1499/*
1500 * Hooks for other subsystems that check memory allocations. In a typical
1501 * production configuration these hooks all should produce no code at all.
1502 */
1503static inline void *kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1504{
1505 ptr = kasan_kmalloc_large(ptr, size, flags);
1506 /* As ptr might get tagged, call kmemleak hook after KASAN. */
1507 kmemleak_alloc(ptr, size, 1, flags);
1508 return ptr;
1509}
1510
1511static __always_inline void kfree_hook(void *x)
1512{
1513 kmemleak_free(x);
1514 kasan_kfree_large(x, _RET_IP_);
1515}
1516
1517static __always_inline bool slab_free_hook(struct kmem_cache *s, void *x)
1518{
1519 kmemleak_free_recursive(x, s->flags);
1520
1521 /*
1522 * Trouble is that we may no longer disable interrupts in the fast path
1523 * So in order to make the debug calls that expect irqs to be
1524 * disabled we need to disable interrupts temporarily.
1525 */
1526#ifdef CONFIG_LOCKDEP
1527 {
1528 unsigned long flags;
1529
1530 local_irq_save(flags);
1531 debug_check_no_locks_freed(x, s->object_size);
1532 local_irq_restore(flags);
1533 }
1534#endif
1535 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1536 debug_check_no_obj_freed(x, s->object_size);
1537
1538 /* Use KCSAN to help debug racy use-after-free. */
1539 if (!(s->flags & SLAB_TYPESAFE_BY_RCU))
1540 __kcsan_check_access(x, s->object_size,
1541 KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT);
1542
1543 /* KASAN might put x into memory quarantine, delaying its reuse */
1544 return kasan_slab_free(s, x, _RET_IP_);
1545}
1546
1547static inline bool slab_free_freelist_hook(struct kmem_cache *s,
1548 void **head, void **tail)
1549{
1550
1551 void *object;
1552 void *next = *head;
1553 void *old_tail = *tail ? *tail : *head;
1554 int rsize;
1555
1556 /* Head and tail of the reconstructed freelist */
1557 *head = NULL;
1558 *tail = NULL;
1559
1560 do {
1561 object = next;
1562 next = get_freepointer(s, object);
1563
1564 if (slab_want_init_on_free(s)) {
1565 /*
1566 * Clear the object and the metadata, but don't touch
1567 * the redzone.
1568 */
1569 memset(object, 0, s->object_size);
1570 rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad
1571 : 0;
1572 memset((char *)object + s->inuse, 0,
1573 s->size - s->inuse - rsize);
1574
1575 }
1576 /* If object's reuse doesn't have to be delayed */
1577 if (!slab_free_hook(s, object)) {
1578 /* Move object to the new freelist */
1579 set_freepointer(s, object, *head);
1580 *head = object;
1581 if (!*tail)
1582 *tail = object;
1583 }
1584 } while (object != old_tail);
1585
1586 if (*head == *tail)
1587 *tail = NULL;
1588
1589 return *head != NULL;
1590}
1591
1592static void *setup_object(struct kmem_cache *s, struct page *page,
1593 void *object)
1594{
1595 setup_object_debug(s, page, object);
1596 object = kasan_init_slab_obj(s, object);
1597 if (unlikely(s->ctor)) {
1598 kasan_unpoison_object_data(s, object);
1599 s->ctor(object);
1600 kasan_poison_object_data(s, object);
1601 }
1602 return object;
1603}
1604
1605/*
1606 * Slab allocation and freeing
1607 */
1608static inline struct page *alloc_slab_page(struct kmem_cache *s,
1609 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1610{
1611 struct page *page;
1612 unsigned int order = oo_order(oo);
1613
1614 if (node == NUMA_NO_NODE)
1615 page = alloc_pages(flags, order);
1616 else
1617 page = __alloc_pages_node(node, flags, order);
1618
1619 if (page)
1620 account_slab_page(page, order, s);
1621
1622 return page;
1623}
1624
1625#ifdef CONFIG_SLAB_FREELIST_RANDOM
1626/* Pre-initialize the random sequence cache */
1627static int init_cache_random_seq(struct kmem_cache *s)
1628{
1629 unsigned int count = oo_objects(s->oo);
1630 int err;
1631
1632 /* Bailout if already initialised */
1633 if (s->random_seq)
1634 return 0;
1635
1636 err = cache_random_seq_create(s, count, GFP_KERNEL);
1637 if (err) {
1638 pr_err("SLUB: Unable to initialize free list for %s\n",
1639 s->name);
1640 return err;
1641 }
1642
1643 /* Transform to an offset on the set of pages */
1644 if (s->random_seq) {
1645 unsigned int i;
1646
1647 for (i = 0; i < count; i++)
1648 s->random_seq[i] *= s->size;
1649 }
1650 return 0;
1651}
1652
1653/* Initialize each random sequence freelist per cache */
1654static void __init init_freelist_randomization(void)
1655{
1656 struct kmem_cache *s;
1657
1658 mutex_lock(&slab_mutex);
1659
1660 list_for_each_entry(s, &slab_caches, list)
1661 init_cache_random_seq(s);
1662
1663 mutex_unlock(&slab_mutex);
1664}
1665
1666/* Get the next entry on the pre-computed freelist randomized */
1667static void *next_freelist_entry(struct kmem_cache *s, struct page *page,
1668 unsigned long *pos, void *start,
1669 unsigned long page_limit,
1670 unsigned long freelist_count)
1671{
1672 unsigned int idx;
1673
1674 /*
1675 * If the target page allocation failed, the number of objects on the
1676 * page might be smaller than the usual size defined by the cache.
1677 */
1678 do {
1679 idx = s->random_seq[*pos];
1680 *pos += 1;
1681 if (*pos >= freelist_count)
1682 *pos = 0;
1683 } while (unlikely(idx >= page_limit));
1684
1685 return (char *)start + idx;
1686}
1687
1688/* Shuffle the single linked freelist based on a random pre-computed sequence */
1689static bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1690{
1691 void *start;
1692 void *cur;
1693 void *next;
1694 unsigned long idx, pos, page_limit, freelist_count;
1695
1696 if (page->objects < 2 || !s->random_seq)
1697 return false;
1698
1699 freelist_count = oo_objects(s->oo);
1700 pos = get_random_int() % freelist_count;
1701
1702 page_limit = page->objects * s->size;
1703 start = fixup_red_left(s, page_address(page));
1704
1705 /* First entry is used as the base of the freelist */
1706 cur = next_freelist_entry(s, page, &pos, start, page_limit,
1707 freelist_count);
1708 cur = setup_object(s, page, cur);
1709 page->freelist = cur;
1710
1711 for (idx = 1; idx < page->objects; idx++) {
1712 next = next_freelist_entry(s, page, &pos, start, page_limit,
1713 freelist_count);
1714 next = setup_object(s, page, next);
1715 set_freepointer(s, cur, next);
1716 cur = next;
1717 }
1718 set_freepointer(s, cur, NULL);
1719
1720 return true;
1721}
1722#else
1723static inline int init_cache_random_seq(struct kmem_cache *s)
1724{
1725 return 0;
1726}
1727static inline void init_freelist_randomization(void) { }
1728static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1729{
1730 return false;
1731}
1732#endif /* CONFIG_SLAB_FREELIST_RANDOM */
1733
1734static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1735{
1736 struct page *page;
1737 struct kmem_cache_order_objects oo = s->oo;
1738 gfp_t alloc_gfp;
1739 void *start, *p, *next;
1740 int idx;
1741 bool shuffle;
1742
1743 flags &= gfp_allowed_mask;
1744
1745 if (gfpflags_allow_blocking(flags))
1746 local_irq_enable();
1747
1748 flags |= s->allocflags;
1749
1750 /*
1751 * Let the initial higher-order allocation fail under memory pressure
1752 * so we fall-back to the minimum order allocation.
1753 */
1754 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1755 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1756 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
1757
1758 page = alloc_slab_page(s, alloc_gfp, node, oo);
1759 if (unlikely(!page)) {
1760 oo = s->min;
1761 alloc_gfp = flags;
1762 /*
1763 * Allocation may have failed due to fragmentation.
1764 * Try a lower order alloc if possible
1765 */
1766 page = alloc_slab_page(s, alloc_gfp, node, oo);
1767 if (unlikely(!page))
1768 goto out;
1769 stat(s, ORDER_FALLBACK);
1770 }
1771
1772 page->objects = oo_objects(oo);
1773
1774 page->slab_cache = s;
1775 __SetPageSlab(page);
1776 if (page_is_pfmemalloc(page))
1777 SetPageSlabPfmemalloc(page);
1778
1779 kasan_poison_slab(page);
1780
1781 start = page_address(page);
1782
1783 setup_page_debug(s, page, start);
1784
1785 shuffle = shuffle_freelist(s, page);
1786
1787 if (!shuffle) {
1788 start = fixup_red_left(s, start);
1789 start = setup_object(s, page, start);
1790 page->freelist = start;
1791 for (idx = 0, p = start; idx < page->objects - 1; idx++) {
1792 next = p + s->size;
1793 next = setup_object(s, page, next);
1794 set_freepointer(s, p, next);
1795 p = next;
1796 }
1797 set_freepointer(s, p, NULL);
1798 }
1799
1800 page->inuse = page->objects;
1801 page->frozen = 1;
1802
1803out:
1804 if (gfpflags_allow_blocking(flags))
1805 local_irq_disable();
1806 if (!page)
1807 return NULL;
1808
1809 inc_slabs_node(s, page_to_nid(page), page->objects);
1810
1811 return page;
1812}
1813
1814static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1815{
1816 if (unlikely(flags & GFP_SLAB_BUG_MASK))
1817 flags = kmalloc_fix_flags(flags);
1818
1819 return allocate_slab(s,
1820 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1821}
1822
1823static void __free_slab(struct kmem_cache *s, struct page *page)
1824{
1825 int order = compound_order(page);
1826 int pages = 1 << order;
1827
1828 if (kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS)) {
1829 void *p;
1830
1831 slab_pad_check(s, page);
1832 for_each_object(p, s, page_address(page),
1833 page->objects)
1834 check_object(s, page, p, SLUB_RED_INACTIVE);
1835 }
1836
1837 __ClearPageSlabPfmemalloc(page);
1838 __ClearPageSlab(page);
1839
1840 page->mapping = NULL;
1841 if (current->reclaim_state)
1842 current->reclaim_state->reclaimed_slab += pages;
1843 unaccount_slab_page(page, order, s);
1844 __free_pages(page, order);
1845}
1846
1847static void rcu_free_slab(struct rcu_head *h)
1848{
1849 struct page *page = container_of(h, struct page, rcu_head);
1850
1851 __free_slab(page->slab_cache, page);
1852}
1853
1854static void free_slab(struct kmem_cache *s, struct page *page)
1855{
1856 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) {
1857 call_rcu(&page->rcu_head, rcu_free_slab);
1858 } else
1859 __free_slab(s, page);
1860}
1861
1862static void discard_slab(struct kmem_cache *s, struct page *page)
1863{
1864 dec_slabs_node(s, page_to_nid(page), page->objects);
1865 free_slab(s, page);
1866}
1867
1868/*
1869 * Management of partially allocated slabs.
1870 */
1871static inline void
1872__add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1873{
1874 n->nr_partial++;
1875 if (tail == DEACTIVATE_TO_TAIL)
1876 list_add_tail(&page->slab_list, &n->partial);
1877 else
1878 list_add(&page->slab_list, &n->partial);
1879}
1880
1881static inline void add_partial(struct kmem_cache_node *n,
1882 struct page *page, int tail)
1883{
1884 lockdep_assert_held(&n->list_lock);
1885 __add_partial(n, page, tail);
1886}
1887
1888static inline void remove_partial(struct kmem_cache_node *n,
1889 struct page *page)
1890{
1891 lockdep_assert_held(&n->list_lock);
1892 list_del(&page->slab_list);
1893 n->nr_partial--;
1894}
1895
1896/*
1897 * Remove slab from the partial list, freeze it and
1898 * return the pointer to the freelist.
1899 *
1900 * Returns a list of objects or NULL if it fails.
1901 */
1902static inline void *acquire_slab(struct kmem_cache *s,
1903 struct kmem_cache_node *n, struct page *page,
1904 int mode, int *objects)
1905{
1906 void *freelist;
1907 unsigned long counters;
1908 struct page new;
1909
1910 lockdep_assert_held(&n->list_lock);
1911
1912 /*
1913 * Zap the freelist and set the frozen bit.
1914 * The old freelist is the list of objects for the
1915 * per cpu allocation list.
1916 */
1917 freelist = page->freelist;
1918 counters = page->counters;
1919 new.counters = counters;
1920 *objects = new.objects - new.inuse;
1921 if (mode) {
1922 new.inuse = page->objects;
1923 new.freelist = NULL;
1924 } else {
1925 new.freelist = freelist;
1926 }
1927
1928 VM_BUG_ON(new.frozen);
1929 new.frozen = 1;
1930
1931 if (!__cmpxchg_double_slab(s, page,
1932 freelist, counters,
1933 new.freelist, new.counters,
1934 "acquire_slab"))
1935 return NULL;
1936
1937 remove_partial(n, page);
1938 WARN_ON(!freelist);
1939 return freelist;
1940}
1941
1942static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1943static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1944
1945/*
1946 * Try to allocate a partial slab from a specific node.
1947 */
1948static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1949 struct kmem_cache_cpu *c, gfp_t flags)
1950{
1951 struct page *page, *page2;
1952 void *object = NULL;
1953 unsigned int available = 0;
1954 int objects;
1955
1956 /*
1957 * Racy check. If we mistakenly see no partial slabs then we
1958 * just allocate an empty slab. If we mistakenly try to get a
1959 * partial slab and there is none available then get_partials()
1960 * will return NULL.
1961 */
1962 if (!n || !n->nr_partial)
1963 return NULL;
1964
1965 spin_lock(&n->list_lock);
1966 list_for_each_entry_safe(page, page2, &n->partial, slab_list) {
1967 void *t;
1968
1969 if (!pfmemalloc_match(page, flags))
1970 continue;
1971
1972 t = acquire_slab(s, n, page, object == NULL, &objects);
1973 if (!t)
1974 break;
1975
1976 available += objects;
1977 if (!object) {
1978 c->page = page;
1979 stat(s, ALLOC_FROM_PARTIAL);
1980 object = t;
1981 } else {
1982 put_cpu_partial(s, page, 0);
1983 stat(s, CPU_PARTIAL_NODE);
1984 }
1985 if (!kmem_cache_has_cpu_partial(s)
1986 || available > slub_cpu_partial(s) / 2)
1987 break;
1988
1989 }
1990 spin_unlock(&n->list_lock);
1991 return object;
1992}
1993
1994/*
1995 * Get a page from somewhere. Search in increasing NUMA distances.
1996 */
1997static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1998 struct kmem_cache_cpu *c)
1999{
2000#ifdef CONFIG_NUMA
2001 struct zonelist *zonelist;
2002 struct zoneref *z;
2003 struct zone *zone;
2004 enum zone_type highest_zoneidx = gfp_zone(flags);
2005 void *object;
2006 unsigned int cpuset_mems_cookie;
2007
2008 /*
2009 * The defrag ratio allows a configuration of the tradeoffs between
2010 * inter node defragmentation and node local allocations. A lower
2011 * defrag_ratio increases the tendency to do local allocations
2012 * instead of attempting to obtain partial slabs from other nodes.
2013 *
2014 * If the defrag_ratio is set to 0 then kmalloc() always
2015 * returns node local objects. If the ratio is higher then kmalloc()
2016 * may return off node objects because partial slabs are obtained
2017 * from other nodes and filled up.
2018 *
2019 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
2020 * (which makes defrag_ratio = 1000) then every (well almost)
2021 * allocation will first attempt to defrag slab caches on other nodes.
2022 * This means scanning over all nodes to look for partial slabs which
2023 * may be expensive if we do it every time we are trying to find a slab
2024 * with available objects.
2025 */
2026 if (!s->remote_node_defrag_ratio ||
2027 get_cycles() % 1024 > s->remote_node_defrag_ratio)
2028 return NULL;
2029
2030 do {
2031 cpuset_mems_cookie = read_mems_allowed_begin();
2032 zonelist = node_zonelist(mempolicy_slab_node(), flags);
2033 for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) {
2034 struct kmem_cache_node *n;
2035
2036 n = get_node(s, zone_to_nid(zone));
2037
2038 if (n && cpuset_zone_allowed(zone, flags) &&
2039 n->nr_partial > s->min_partial) {
2040 object = get_partial_node(s, n, c, flags);
2041 if (object) {
2042 /*
2043 * Don't check read_mems_allowed_retry()
2044 * here - if mems_allowed was updated in
2045 * parallel, that was a harmless race
2046 * between allocation and the cpuset
2047 * update
2048 */
2049 return object;
2050 }
2051 }
2052 }
2053 } while (read_mems_allowed_retry(cpuset_mems_cookie));
2054#endif /* CONFIG_NUMA */
2055 return NULL;
2056}
2057
2058/*
2059 * Get a partial page, lock it and return it.
2060 */
2061static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
2062 struct kmem_cache_cpu *c)
2063{
2064 void *object;
2065 int searchnode = node;
2066
2067 if (node == NUMA_NO_NODE)
2068 searchnode = numa_mem_id();
2069
2070 object = get_partial_node(s, get_node(s, searchnode), c, flags);
2071 if (object || node != NUMA_NO_NODE)
2072 return object;
2073
2074 return get_any_partial(s, flags, c);
2075}
2076
2077#ifdef CONFIG_PREEMPTION
2078/*
2079 * Calculate the next globally unique transaction for disambiguation
2080 * during cmpxchg. The transactions start with the cpu number and are then
2081 * incremented by CONFIG_NR_CPUS.
2082 */
2083#define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
2084#else
2085/*
2086 * No preemption supported therefore also no need to check for
2087 * different cpus.
2088 */
2089#define TID_STEP 1
2090#endif
2091
2092static inline unsigned long next_tid(unsigned long tid)
2093{
2094 return tid + TID_STEP;
2095}
2096
2097#ifdef SLUB_DEBUG_CMPXCHG
2098static inline unsigned int tid_to_cpu(unsigned long tid)
2099{
2100 return tid % TID_STEP;
2101}
2102
2103static inline unsigned long tid_to_event(unsigned long tid)
2104{
2105 return tid / TID_STEP;
2106}
2107#endif
2108
2109static inline unsigned int init_tid(int cpu)
2110{
2111 return cpu;
2112}
2113
2114static inline void note_cmpxchg_failure(const char *n,
2115 const struct kmem_cache *s, unsigned long tid)
2116{
2117#ifdef SLUB_DEBUG_CMPXCHG
2118 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
2119
2120 pr_info("%s %s: cmpxchg redo ", n, s->name);
2121
2122#ifdef CONFIG_PREEMPTION
2123 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
2124 pr_warn("due to cpu change %d -> %d\n",
2125 tid_to_cpu(tid), tid_to_cpu(actual_tid));
2126 else
2127#endif
2128 if (tid_to_event(tid) != tid_to_event(actual_tid))
2129 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2130 tid_to_event(tid), tid_to_event(actual_tid));
2131 else
2132 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2133 actual_tid, tid, next_tid(tid));
2134#endif
2135 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2136}
2137
2138static void init_kmem_cache_cpus(struct kmem_cache *s)
2139{
2140 int cpu;
2141
2142 for_each_possible_cpu(cpu)
2143 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
2144}
2145
2146/*
2147 * Remove the cpu slab
2148 */
2149static void deactivate_slab(struct kmem_cache *s, struct page *page,
2150 void *freelist, struct kmem_cache_cpu *c)
2151{
2152 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
2153 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
2154 int lock = 0;
2155 enum slab_modes l = M_NONE, m = M_NONE;
2156 void *nextfree;
2157 int tail = DEACTIVATE_TO_HEAD;
2158 struct page new;
2159 struct page old;
2160
2161 if (page->freelist) {
2162 stat(s, DEACTIVATE_REMOTE_FREES);
2163 tail = DEACTIVATE_TO_TAIL;
2164 }
2165
2166 /*
2167 * Stage one: Free all available per cpu objects back
2168 * to the page freelist while it is still frozen. Leave the
2169 * last one.
2170 *
2171 * There is no need to take the list->lock because the page
2172 * is still frozen.
2173 */
2174 while (freelist && (nextfree = get_freepointer(s, freelist))) {
2175 void *prior;
2176 unsigned long counters;
2177
2178 /*
2179 * If 'nextfree' is invalid, it is possible that the object at
2180 * 'freelist' is already corrupted. So isolate all objects
2181 * starting at 'freelist'.
2182 */
2183 if (freelist_corrupted(s, page, &freelist, nextfree))
2184 break;
2185
2186 do {
2187 prior = page->freelist;
2188 counters = page->counters;
2189 set_freepointer(s, freelist, prior);
2190 new.counters = counters;
2191 new.inuse--;
2192 VM_BUG_ON(!new.frozen);
2193
2194 } while (!__cmpxchg_double_slab(s, page,
2195 prior, counters,
2196 freelist, new.counters,
2197 "drain percpu freelist"));
2198
2199 freelist = nextfree;
2200 }
2201
2202 /*
2203 * Stage two: Ensure that the page is unfrozen while the
2204 * list presence reflects the actual number of objects
2205 * during unfreeze.
2206 *
2207 * We setup the list membership and then perform a cmpxchg
2208 * with the count. If there is a mismatch then the page
2209 * is not unfrozen but the page is on the wrong list.
2210 *
2211 * Then we restart the process which may have to remove
2212 * the page from the list that we just put it on again
2213 * because the number of objects in the slab may have
2214 * changed.
2215 */
2216redo:
2217
2218 old.freelist = page->freelist;
2219 old.counters = page->counters;
2220 VM_BUG_ON(!old.frozen);
2221
2222 /* Determine target state of the slab */
2223 new.counters = old.counters;
2224 if (freelist) {
2225 new.inuse--;
2226 set_freepointer(s, freelist, old.freelist);
2227 new.freelist = freelist;
2228 } else
2229 new.freelist = old.freelist;
2230
2231 new.frozen = 0;
2232
2233 if (!new.inuse && n->nr_partial >= s->min_partial)
2234 m = M_FREE;
2235 else if (new.freelist) {
2236 m = M_PARTIAL;
2237 if (!lock) {
2238 lock = 1;
2239 /*
2240 * Taking the spinlock removes the possibility
2241 * that acquire_slab() will see a slab page that
2242 * is frozen
2243 */
2244 spin_lock(&n->list_lock);
2245 }
2246 } else {
2247 m = M_FULL;
2248 if (kmem_cache_debug(s) && !lock) {
2249 lock = 1;
2250 /*
2251 * This also ensures that the scanning of full
2252 * slabs from diagnostic functions will not see
2253 * any frozen slabs.
2254 */
2255 spin_lock(&n->list_lock);
2256 }
2257 }
2258
2259 if (l != m) {
2260 if (l == M_PARTIAL)
2261 remove_partial(n, page);
2262 else if (l == M_FULL)
2263 remove_full(s, n, page);
2264
2265 if (m == M_PARTIAL)
2266 add_partial(n, page, tail);
2267 else if (m == M_FULL)
2268 add_full(s, n, page);
2269 }
2270
2271 l = m;
2272 if (!__cmpxchg_double_slab(s, page,
2273 old.freelist, old.counters,
2274 new.freelist, new.counters,
2275 "unfreezing slab"))
2276 goto redo;
2277
2278 if (lock)
2279 spin_unlock(&n->list_lock);
2280
2281 if (m == M_PARTIAL)
2282 stat(s, tail);
2283 else if (m == M_FULL)
2284 stat(s, DEACTIVATE_FULL);
2285 else if (m == M_FREE) {
2286 stat(s, DEACTIVATE_EMPTY);
2287 discard_slab(s, page);
2288 stat(s, FREE_SLAB);
2289 }
2290
2291 c->page = NULL;
2292 c->freelist = NULL;
2293}
2294
2295/*
2296 * Unfreeze all the cpu partial slabs.
2297 *
2298 * This function must be called with interrupts disabled
2299 * for the cpu using c (or some other guarantee must be there
2300 * to guarantee no concurrent accesses).
2301 */
2302static void unfreeze_partials(struct kmem_cache *s,
2303 struct kmem_cache_cpu *c)
2304{
2305#ifdef CONFIG_SLUB_CPU_PARTIAL
2306 struct kmem_cache_node *n = NULL, *n2 = NULL;
2307 struct page *page, *discard_page = NULL;
2308
2309 while ((page = slub_percpu_partial(c))) {
2310 struct page new;
2311 struct page old;
2312
2313 slub_set_percpu_partial(c, page);
2314
2315 n2 = get_node(s, page_to_nid(page));
2316 if (n != n2) {
2317 if (n)
2318 spin_unlock(&n->list_lock);
2319
2320 n = n2;
2321 spin_lock(&n->list_lock);
2322 }
2323
2324 do {
2325
2326 old.freelist = page->freelist;
2327 old.counters = page->counters;
2328 VM_BUG_ON(!old.frozen);
2329
2330 new.counters = old.counters;
2331 new.freelist = old.freelist;
2332
2333 new.frozen = 0;
2334
2335 } while (!__cmpxchg_double_slab(s, page,
2336 old.freelist, old.counters,
2337 new.freelist, new.counters,
2338 "unfreezing slab"));
2339
2340 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2341 page->next = discard_page;
2342 discard_page = page;
2343 } else {
2344 add_partial(n, page, DEACTIVATE_TO_TAIL);
2345 stat(s, FREE_ADD_PARTIAL);
2346 }
2347 }
2348
2349 if (n)
2350 spin_unlock(&n->list_lock);
2351
2352 while (discard_page) {
2353 page = discard_page;
2354 discard_page = discard_page->next;
2355
2356 stat(s, DEACTIVATE_EMPTY);
2357 discard_slab(s, page);
2358 stat(s, FREE_SLAB);
2359 }
2360#endif /* CONFIG_SLUB_CPU_PARTIAL */
2361}
2362
2363/*
2364 * Put a page that was just frozen (in __slab_free|get_partial_node) into a
2365 * partial page slot if available.
2366 *
2367 * If we did not find a slot then simply move all the partials to the
2368 * per node partial list.
2369 */
2370static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2371{
2372#ifdef CONFIG_SLUB_CPU_PARTIAL
2373 struct page *oldpage;
2374 int pages;
2375 int pobjects;
2376
2377 preempt_disable();
2378 do {
2379 pages = 0;
2380 pobjects = 0;
2381 oldpage = this_cpu_read(s->cpu_slab->partial);
2382
2383 if (oldpage) {
2384 pobjects = oldpage->pobjects;
2385 pages = oldpage->pages;
2386 if (drain && pobjects > slub_cpu_partial(s)) {
2387 unsigned long flags;
2388 /*
2389 * partial array is full. Move the existing
2390 * set to the per node partial list.
2391 */
2392 local_irq_save(flags);
2393 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2394 local_irq_restore(flags);
2395 oldpage = NULL;
2396 pobjects = 0;
2397 pages = 0;
2398 stat(s, CPU_PARTIAL_DRAIN);
2399 }
2400 }
2401
2402 pages++;
2403 pobjects += page->objects - page->inuse;
2404
2405 page->pages = pages;
2406 page->pobjects = pobjects;
2407 page->next = oldpage;
2408
2409 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2410 != oldpage);
2411 if (unlikely(!slub_cpu_partial(s))) {
2412 unsigned long flags;
2413
2414 local_irq_save(flags);
2415 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2416 local_irq_restore(flags);
2417 }
2418 preempt_enable();
2419#endif /* CONFIG_SLUB_CPU_PARTIAL */
2420}
2421
2422static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2423{
2424 stat(s, CPUSLAB_FLUSH);
2425 deactivate_slab(s, c->page, c->freelist, c);
2426
2427 c->tid = next_tid(c->tid);
2428}
2429
2430/*
2431 * Flush cpu slab.
2432 *
2433 * Called from IPI handler with interrupts disabled.
2434 */
2435static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2436{
2437 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2438
2439 if (c->page)
2440 flush_slab(s, c);
2441
2442 unfreeze_partials(s, c);
2443}
2444
2445static void flush_cpu_slab(void *d)
2446{
2447 struct kmem_cache *s = d;
2448
2449 __flush_cpu_slab(s, smp_processor_id());
2450}
2451
2452static bool has_cpu_slab(int cpu, void *info)
2453{
2454 struct kmem_cache *s = info;
2455 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2456
2457 return c->page || slub_percpu_partial(c);
2458}
2459
2460static void flush_all(struct kmem_cache *s)
2461{
2462 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1);
2463}
2464
2465/*
2466 * Use the cpu notifier to insure that the cpu slabs are flushed when
2467 * necessary.
2468 */
2469static int slub_cpu_dead(unsigned int cpu)
2470{
2471 struct kmem_cache *s;
2472 unsigned long flags;
2473
2474 mutex_lock(&slab_mutex);
2475 list_for_each_entry(s, &slab_caches, list) {
2476 local_irq_save(flags);
2477 __flush_cpu_slab(s, cpu);
2478 local_irq_restore(flags);
2479 }
2480 mutex_unlock(&slab_mutex);
2481 return 0;
2482}
2483
2484/*
2485 * Check if the objects in a per cpu structure fit numa
2486 * locality expectations.
2487 */
2488static inline int node_match(struct page *page, int node)
2489{
2490#ifdef CONFIG_NUMA
2491 if (node != NUMA_NO_NODE && page_to_nid(page) != node)
2492 return 0;
2493#endif
2494 return 1;
2495}
2496
2497#ifdef CONFIG_SLUB_DEBUG
2498static int count_free(struct page *page)
2499{
2500 return page->objects - page->inuse;
2501}
2502
2503static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2504{
2505 return atomic_long_read(&n->total_objects);
2506}
2507#endif /* CONFIG_SLUB_DEBUG */
2508
2509#if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2510static unsigned long count_partial(struct kmem_cache_node *n,
2511 int (*get_count)(struct page *))
2512{
2513 unsigned long flags;
2514 unsigned long x = 0;
2515 struct page *page;
2516
2517 spin_lock_irqsave(&n->list_lock, flags);
2518 list_for_each_entry(page, &n->partial, slab_list)
2519 x += get_count(page);
2520 spin_unlock_irqrestore(&n->list_lock, flags);
2521 return x;
2522}
2523#endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2524
2525static noinline void
2526slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2527{
2528#ifdef CONFIG_SLUB_DEBUG
2529 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2530 DEFAULT_RATELIMIT_BURST);
2531 int node;
2532 struct kmem_cache_node *n;
2533
2534 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2535 return;
2536
2537 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2538 nid, gfpflags, &gfpflags);
2539 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2540 s->name, s->object_size, s->size, oo_order(s->oo),
2541 oo_order(s->min));
2542
2543 if (oo_order(s->min) > get_order(s->object_size))
2544 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2545 s->name);
2546
2547 for_each_kmem_cache_node(s, node, n) {
2548 unsigned long nr_slabs;
2549 unsigned long nr_objs;
2550 unsigned long nr_free;
2551
2552 nr_free = count_partial(n, count_free);
2553 nr_slabs = node_nr_slabs(n);
2554 nr_objs = node_nr_objs(n);
2555
2556 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2557 node, nr_slabs, nr_objs, nr_free);
2558 }
2559#endif
2560}
2561
2562static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2563 int node, struct kmem_cache_cpu **pc)
2564{
2565 void *freelist;
2566 struct kmem_cache_cpu *c = *pc;
2567 struct page *page;
2568
2569 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
2570
2571 freelist = get_partial(s, flags, node, c);
2572
2573 if (freelist)
2574 return freelist;
2575
2576 page = new_slab(s, flags, node);
2577 if (page) {
2578 c = raw_cpu_ptr(s->cpu_slab);
2579 if (c->page)
2580 flush_slab(s, c);
2581
2582 /*
2583 * No other reference to the page yet so we can
2584 * muck around with it freely without cmpxchg
2585 */
2586 freelist = page->freelist;
2587 page->freelist = NULL;
2588
2589 stat(s, ALLOC_SLAB);
2590 c->page = page;
2591 *pc = c;
2592 }
2593
2594 return freelist;
2595}
2596
2597static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2598{
2599 if (unlikely(PageSlabPfmemalloc(page)))
2600 return gfp_pfmemalloc_allowed(gfpflags);
2601
2602 return true;
2603}
2604
2605/*
2606 * Check the page->freelist of a page and either transfer the freelist to the
2607 * per cpu freelist or deactivate the page.
2608 *
2609 * The page is still frozen if the return value is not NULL.
2610 *
2611 * If this function returns NULL then the page has been unfrozen.
2612 *
2613 * This function must be called with interrupt disabled.
2614 */
2615static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2616{
2617 struct page new;
2618 unsigned long counters;
2619 void *freelist;
2620
2621 do {
2622 freelist = page->freelist;
2623 counters = page->counters;
2624
2625 new.counters = counters;
2626 VM_BUG_ON(!new.frozen);
2627
2628 new.inuse = page->objects;
2629 new.frozen = freelist != NULL;
2630
2631 } while (!__cmpxchg_double_slab(s, page,
2632 freelist, counters,
2633 NULL, new.counters,
2634 "get_freelist"));
2635
2636 return freelist;
2637}
2638
2639/*
2640 * Slow path. The lockless freelist is empty or we need to perform
2641 * debugging duties.
2642 *
2643 * Processing is still very fast if new objects have been freed to the
2644 * regular freelist. In that case we simply take over the regular freelist
2645 * as the lockless freelist and zap the regular freelist.
2646 *
2647 * If that is not working then we fall back to the partial lists. We take the
2648 * first element of the freelist as the object to allocate now and move the
2649 * rest of the freelist to the lockless freelist.
2650 *
2651 * And if we were unable to get a new slab from the partial slab lists then
2652 * we need to allocate a new slab. This is the slowest path since it involves
2653 * a call to the page allocator and the setup of a new slab.
2654 *
2655 * Version of __slab_alloc to use when we know that interrupts are
2656 * already disabled (which is the case for bulk allocation).
2657 */
2658static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2659 unsigned long addr, struct kmem_cache_cpu *c)
2660{
2661 void *freelist;
2662 struct page *page;
2663
2664 page = c->page;
2665 if (!page) {
2666 /*
2667 * if the node is not online or has no normal memory, just
2668 * ignore the node constraint
2669 */
2670 if (unlikely(node != NUMA_NO_NODE &&
2671 !node_state(node, N_NORMAL_MEMORY)))
2672 node = NUMA_NO_NODE;
2673 goto new_slab;
2674 }
2675redo:
2676
2677 if (unlikely(!node_match(page, node))) {
2678 /*
2679 * same as above but node_match() being false already
2680 * implies node != NUMA_NO_NODE
2681 */
2682 if (!node_state(node, N_NORMAL_MEMORY)) {
2683 node = NUMA_NO_NODE;
2684 goto redo;
2685 } else {
2686 stat(s, ALLOC_NODE_MISMATCH);
2687 deactivate_slab(s, page, c->freelist, c);
2688 goto new_slab;
2689 }
2690 }
2691
2692 /*
2693 * By rights, we should be searching for a slab page that was
2694 * PFMEMALLOC but right now, we are losing the pfmemalloc
2695 * information when the page leaves the per-cpu allocator
2696 */
2697 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2698 deactivate_slab(s, page, c->freelist, c);
2699 goto new_slab;
2700 }
2701
2702 /* must check again c->freelist in case of cpu migration or IRQ */
2703 freelist = c->freelist;
2704 if (freelist)
2705 goto load_freelist;
2706
2707 freelist = get_freelist(s, page);
2708
2709 if (!freelist) {
2710 c->page = NULL;
2711 stat(s, DEACTIVATE_BYPASS);
2712 goto new_slab;
2713 }
2714
2715 stat(s, ALLOC_REFILL);
2716
2717load_freelist:
2718 /*
2719 * freelist is pointing to the list of objects to be used.
2720 * page is pointing to the page from which the objects are obtained.
2721 * That page must be frozen for per cpu allocations to work.
2722 */
2723 VM_BUG_ON(!c->page->frozen);
2724 c->freelist = get_freepointer(s, freelist);
2725 c->tid = next_tid(c->tid);
2726 return freelist;
2727
2728new_slab:
2729
2730 if (slub_percpu_partial(c)) {
2731 page = c->page = slub_percpu_partial(c);
2732 slub_set_percpu_partial(c, page);
2733 stat(s, CPU_PARTIAL_ALLOC);
2734 goto redo;
2735 }
2736
2737 freelist = new_slab_objects(s, gfpflags, node, &c);
2738
2739 if (unlikely(!freelist)) {
2740 slab_out_of_memory(s, gfpflags, node);
2741 return NULL;
2742 }
2743
2744 page = c->page;
2745 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2746 goto load_freelist;
2747
2748 /* Only entered in the debug case */
2749 if (kmem_cache_debug(s) &&
2750 !alloc_debug_processing(s, page, freelist, addr))
2751 goto new_slab; /* Slab failed checks. Next slab needed */
2752
2753 deactivate_slab(s, page, get_freepointer(s, freelist), c);
2754 return freelist;
2755}
2756
2757/*
2758 * Another one that disabled interrupt and compensates for possible
2759 * cpu changes by refetching the per cpu area pointer.
2760 */
2761static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2762 unsigned long addr, struct kmem_cache_cpu *c)
2763{
2764 void *p;
2765 unsigned long flags;
2766
2767 local_irq_save(flags);
2768#ifdef CONFIG_PREEMPTION
2769 /*
2770 * We may have been preempted and rescheduled on a different
2771 * cpu before disabling interrupts. Need to reload cpu area
2772 * pointer.
2773 */
2774 c = this_cpu_ptr(s->cpu_slab);
2775#endif
2776
2777 p = ___slab_alloc(s, gfpflags, node, addr, c);
2778 local_irq_restore(flags);
2779 return p;
2780}
2781
2782/*
2783 * If the object has been wiped upon free, make sure it's fully initialized by
2784 * zeroing out freelist pointer.
2785 */
2786static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s,
2787 void *obj)
2788{
2789 if (unlikely(slab_want_init_on_free(s)) && obj)
2790 memset((void *)((char *)obj + s->offset), 0, sizeof(void *));
2791}
2792
2793/*
2794 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2795 * have the fastpath folded into their functions. So no function call
2796 * overhead for requests that can be satisfied on the fastpath.
2797 *
2798 * The fastpath works by first checking if the lockless freelist can be used.
2799 * If not then __slab_alloc is called for slow processing.
2800 *
2801 * Otherwise we can simply pick the next object from the lockless free list.
2802 */
2803static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2804 gfp_t gfpflags, int node, unsigned long addr)
2805{
2806 void *object;
2807 struct kmem_cache_cpu *c;
2808 struct page *page;
2809 unsigned long tid;
2810 struct obj_cgroup *objcg = NULL;
2811
2812 s = slab_pre_alloc_hook(s, &objcg, 1, gfpflags);
2813 if (!s)
2814 return NULL;
2815redo:
2816 /*
2817 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2818 * enabled. We may switch back and forth between cpus while
2819 * reading from one cpu area. That does not matter as long
2820 * as we end up on the original cpu again when doing the cmpxchg.
2821 *
2822 * We should guarantee that tid and kmem_cache are retrieved on
2823 * the same cpu. It could be different if CONFIG_PREEMPTION so we need
2824 * to check if it is matched or not.
2825 */
2826 do {
2827 tid = this_cpu_read(s->cpu_slab->tid);
2828 c = raw_cpu_ptr(s->cpu_slab);
2829 } while (IS_ENABLED(CONFIG_PREEMPTION) &&
2830 unlikely(tid != READ_ONCE(c->tid)));
2831
2832 /*
2833 * Irqless object alloc/free algorithm used here depends on sequence
2834 * of fetching cpu_slab's data. tid should be fetched before anything
2835 * on c to guarantee that object and page associated with previous tid
2836 * won't be used with current tid. If we fetch tid first, object and
2837 * page could be one associated with next tid and our alloc/free
2838 * request will be failed. In this case, we will retry. So, no problem.
2839 */
2840 barrier();
2841
2842 /*
2843 * The transaction ids are globally unique per cpu and per operation on
2844 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2845 * occurs on the right processor and that there was no operation on the
2846 * linked list in between.
2847 */
2848
2849 object = c->freelist;
2850 page = c->page;
2851 if (unlikely(!object || !node_match(page, node))) {
2852 object = __slab_alloc(s, gfpflags, node, addr, c);
2853 stat(s, ALLOC_SLOWPATH);
2854 } else {
2855 void *next_object = get_freepointer_safe(s, object);
2856
2857 /*
2858 * The cmpxchg will only match if there was no additional
2859 * operation and if we are on the right processor.
2860 *
2861 * The cmpxchg does the following atomically (without lock
2862 * semantics!)
2863 * 1. Relocate first pointer to the current per cpu area.
2864 * 2. Verify that tid and freelist have not been changed
2865 * 3. If they were not changed replace tid and freelist
2866 *
2867 * Since this is without lock semantics the protection is only
2868 * against code executing on this cpu *not* from access by
2869 * other cpus.
2870 */
2871 if (unlikely(!this_cpu_cmpxchg_double(
2872 s->cpu_slab->freelist, s->cpu_slab->tid,
2873 object, tid,
2874 next_object, next_tid(tid)))) {
2875
2876 note_cmpxchg_failure("slab_alloc", s, tid);
2877 goto redo;
2878 }
2879 prefetch_freepointer(s, next_object);
2880 stat(s, ALLOC_FASTPATH);
2881 }
2882
2883 maybe_wipe_obj_freeptr(s, object);
2884
2885 if (unlikely(slab_want_init_on_alloc(gfpflags, s)) && object)
2886 memset(object, 0, s->object_size);
2887
2888 slab_post_alloc_hook(s, objcg, gfpflags, 1, &object);
2889
2890 return object;
2891}
2892
2893static __always_inline void *slab_alloc(struct kmem_cache *s,
2894 gfp_t gfpflags, unsigned long addr)
2895{
2896 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2897}
2898
2899void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2900{
2901 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2902
2903 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2904 s->size, gfpflags);
2905
2906 return ret;
2907}
2908EXPORT_SYMBOL(kmem_cache_alloc);
2909
2910#ifdef CONFIG_TRACING
2911void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2912{
2913 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2914 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2915 ret = kasan_kmalloc(s, ret, size, gfpflags);
2916 return ret;
2917}
2918EXPORT_SYMBOL(kmem_cache_alloc_trace);
2919#endif
2920
2921#ifdef CONFIG_NUMA
2922void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2923{
2924 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2925
2926 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2927 s->object_size, s->size, gfpflags, node);
2928
2929 return ret;
2930}
2931EXPORT_SYMBOL(kmem_cache_alloc_node);
2932
2933#ifdef CONFIG_TRACING
2934void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2935 gfp_t gfpflags,
2936 int node, size_t size)
2937{
2938 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2939
2940 trace_kmalloc_node(_RET_IP_, ret,
2941 size, s->size, gfpflags, node);
2942
2943 ret = kasan_kmalloc(s, ret, size, gfpflags);
2944 return ret;
2945}
2946EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2947#endif
2948#endif /* CONFIG_NUMA */
2949
2950/*
2951 * Slow path handling. This may still be called frequently since objects
2952 * have a longer lifetime than the cpu slabs in most processing loads.
2953 *
2954 * So we still attempt to reduce cache line usage. Just take the slab
2955 * lock and free the item. If there is no additional partial page
2956 * handling required then we can return immediately.
2957 */
2958static void __slab_free(struct kmem_cache *s, struct page *page,
2959 void *head, void *tail, int cnt,
2960 unsigned long addr)
2961
2962{
2963 void *prior;
2964 int was_frozen;
2965 struct page new;
2966 unsigned long counters;
2967 struct kmem_cache_node *n = NULL;
2968 unsigned long flags;
2969
2970 stat(s, FREE_SLOWPATH);
2971
2972 if (kmem_cache_debug(s) &&
2973 !free_debug_processing(s, page, head, tail, cnt, addr))
2974 return;
2975
2976 do {
2977 if (unlikely(n)) {
2978 spin_unlock_irqrestore(&n->list_lock, flags);
2979 n = NULL;
2980 }
2981 prior = page->freelist;
2982 counters = page->counters;
2983 set_freepointer(s, tail, prior);
2984 new.counters = counters;
2985 was_frozen = new.frozen;
2986 new.inuse -= cnt;
2987 if ((!new.inuse || !prior) && !was_frozen) {
2988
2989 if (kmem_cache_has_cpu_partial(s) && !prior) {
2990
2991 /*
2992 * Slab was on no list before and will be
2993 * partially empty
2994 * We can defer the list move and instead
2995 * freeze it.
2996 */
2997 new.frozen = 1;
2998
2999 } else { /* Needs to be taken off a list */
3000
3001 n = get_node(s, page_to_nid(page));
3002 /*
3003 * Speculatively acquire the list_lock.
3004 * If the cmpxchg does not succeed then we may
3005 * drop the list_lock without any processing.
3006 *
3007 * Otherwise the list_lock will synchronize with
3008 * other processors updating the list of slabs.
3009 */
3010 spin_lock_irqsave(&n->list_lock, flags);
3011
3012 }
3013 }
3014
3015 } while (!cmpxchg_double_slab(s, page,
3016 prior, counters,
3017 head, new.counters,
3018 "__slab_free"));
3019
3020 if (likely(!n)) {
3021
3022 /*
3023 * If we just froze the page then put it onto the
3024 * per cpu partial list.
3025 */
3026 if (new.frozen && !was_frozen) {
3027 put_cpu_partial(s, page, 1);
3028 stat(s, CPU_PARTIAL_FREE);
3029 }
3030 /*
3031 * The list lock was not taken therefore no list
3032 * activity can be necessary.
3033 */
3034 if (was_frozen)
3035 stat(s, FREE_FROZEN);
3036 return;
3037 }
3038
3039 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
3040 goto slab_empty;
3041
3042 /*
3043 * Objects left in the slab. If it was not on the partial list before
3044 * then add it.
3045 */
3046 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
3047 remove_full(s, n, page);
3048 add_partial(n, page, DEACTIVATE_TO_TAIL);
3049 stat(s, FREE_ADD_PARTIAL);
3050 }
3051 spin_unlock_irqrestore(&n->list_lock, flags);
3052 return;
3053
3054slab_empty:
3055 if (prior) {
3056 /*
3057 * Slab on the partial list.
3058 */
3059 remove_partial(n, page);
3060 stat(s, FREE_REMOVE_PARTIAL);
3061 } else {
3062 /* Slab must be on the full list */
3063 remove_full(s, n, page);
3064 }
3065
3066 spin_unlock_irqrestore(&n->list_lock, flags);
3067 stat(s, FREE_SLAB);
3068 discard_slab(s, page);
3069}
3070
3071/*
3072 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
3073 * can perform fastpath freeing without additional function calls.
3074 *
3075 * The fastpath is only possible if we are freeing to the current cpu slab
3076 * of this processor. This typically the case if we have just allocated
3077 * the item before.
3078 *
3079 * If fastpath is not possible then fall back to __slab_free where we deal
3080 * with all sorts of special processing.
3081 *
3082 * Bulk free of a freelist with several objects (all pointing to the
3083 * same page) possible by specifying head and tail ptr, plus objects
3084 * count (cnt). Bulk free indicated by tail pointer being set.
3085 */
3086static __always_inline void do_slab_free(struct kmem_cache *s,
3087 struct page *page, void *head, void *tail,
3088 int cnt, unsigned long addr)
3089{
3090 void *tail_obj = tail ? : head;
3091 struct kmem_cache_cpu *c;
3092 unsigned long tid;
3093
3094 memcg_slab_free_hook(s, page, head);
3095redo:
3096 /*
3097 * Determine the currently cpus per cpu slab.
3098 * The cpu may change afterward. However that does not matter since
3099 * data is retrieved via this pointer. If we are on the same cpu
3100 * during the cmpxchg then the free will succeed.
3101 */
3102 do {
3103 tid = this_cpu_read(s->cpu_slab->tid);
3104 c = raw_cpu_ptr(s->cpu_slab);
3105 } while (IS_ENABLED(CONFIG_PREEMPTION) &&
3106 unlikely(tid != READ_ONCE(c->tid)));
3107
3108 /* Same with comment on barrier() in slab_alloc_node() */
3109 barrier();
3110
3111 if (likely(page == c->page)) {
3112 void **freelist = READ_ONCE(c->freelist);
3113
3114 set_freepointer(s, tail_obj, freelist);
3115
3116 if (unlikely(!this_cpu_cmpxchg_double(
3117 s->cpu_slab->freelist, s->cpu_slab->tid,
3118 freelist, tid,
3119 head, next_tid(tid)))) {
3120
3121 note_cmpxchg_failure("slab_free", s, tid);
3122 goto redo;
3123 }
3124 stat(s, FREE_FASTPATH);
3125 } else
3126 __slab_free(s, page, head, tail_obj, cnt, addr);
3127
3128}
3129
3130static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
3131 void *head, void *tail, int cnt,
3132 unsigned long addr)
3133{
3134 /*
3135 * With KASAN enabled slab_free_freelist_hook modifies the freelist
3136 * to remove objects, whose reuse must be delayed.
3137 */
3138 if (slab_free_freelist_hook(s, &head, &tail))
3139 do_slab_free(s, page, head, tail, cnt, addr);
3140}
3141
3142#ifdef CONFIG_KASAN_GENERIC
3143void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
3144{
3145 do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr);
3146}
3147#endif
3148
3149void kmem_cache_free(struct kmem_cache *s, void *x)
3150{
3151 s = cache_from_obj(s, x);
3152 if (!s)
3153 return;
3154 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
3155 trace_kmem_cache_free(_RET_IP_, x);
3156}
3157EXPORT_SYMBOL(kmem_cache_free);
3158
3159struct detached_freelist {
3160 struct page *page;
3161 void *tail;
3162 void *freelist;
3163 int cnt;
3164 struct kmem_cache *s;
3165};
3166
3167/*
3168 * This function progressively scans the array with free objects (with
3169 * a limited look ahead) and extract objects belonging to the same
3170 * page. It builds a detached freelist directly within the given
3171 * page/objects. This can happen without any need for
3172 * synchronization, because the objects are owned by running process.
3173 * The freelist is build up as a single linked list in the objects.
3174 * The idea is, that this detached freelist can then be bulk
3175 * transferred to the real freelist(s), but only requiring a single
3176 * synchronization primitive. Look ahead in the array is limited due
3177 * to performance reasons.
3178 */
3179static inline
3180int build_detached_freelist(struct kmem_cache *s, size_t size,
3181 void **p, struct detached_freelist *df)
3182{
3183 size_t first_skipped_index = 0;
3184 int lookahead = 3;
3185 void *object;
3186 struct page *page;
3187
3188 /* Always re-init detached_freelist */
3189 df->page = NULL;
3190
3191 do {
3192 object = p[--size];
3193 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3194 } while (!object && size);
3195
3196 if (!object)
3197 return 0;
3198
3199 page = virt_to_head_page(object);
3200 if (!s) {
3201 /* Handle kalloc'ed objects */
3202 if (unlikely(!PageSlab(page))) {
3203 BUG_ON(!PageCompound(page));
3204 kfree_hook(object);
3205 __free_pages(page, compound_order(page));
3206 p[size] = NULL; /* mark object processed */
3207 return size;
3208 }
3209 /* Derive kmem_cache from object */
3210 df->s = page->slab_cache;
3211 } else {
3212 df->s = cache_from_obj(s, object); /* Support for memcg */
3213 }
3214
3215 /* Start new detached freelist */
3216 df->page = page;
3217 set_freepointer(df->s, object, NULL);
3218 df->tail = object;
3219 df->freelist = object;
3220 p[size] = NULL; /* mark object processed */
3221 df->cnt = 1;
3222
3223 while (size) {
3224 object = p[--size];
3225 if (!object)
3226 continue; /* Skip processed objects */
3227
3228 /* df->page is always set at this point */
3229 if (df->page == virt_to_head_page(object)) {
3230 /* Opportunity build freelist */
3231 set_freepointer(df->s, object, df->freelist);
3232 df->freelist = object;
3233 df->cnt++;
3234 p[size] = NULL; /* mark object processed */
3235
3236 continue;
3237 }
3238
3239 /* Limit look ahead search */
3240 if (!--lookahead)
3241 break;
3242
3243 if (!first_skipped_index)
3244 first_skipped_index = size + 1;
3245 }
3246
3247 return first_skipped_index;
3248}
3249
3250/* Note that interrupts must be enabled when calling this function. */
3251void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3252{
3253 if (WARN_ON(!size))
3254 return;
3255
3256 do {
3257 struct detached_freelist df;
3258
3259 size = build_detached_freelist(s, size, p, &df);
3260 if (!df.page)
3261 continue;
3262
3263 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
3264 } while (likely(size));
3265}
3266EXPORT_SYMBOL(kmem_cache_free_bulk);
3267
3268/* Note that interrupts must be enabled when calling this function. */
3269int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3270 void **p)
3271{
3272 struct kmem_cache_cpu *c;
3273 int i;
3274 struct obj_cgroup *objcg = NULL;
3275
3276 /* memcg and kmem_cache debug support */
3277 s = slab_pre_alloc_hook(s, &objcg, size, flags);
3278 if (unlikely(!s))
3279 return false;
3280 /*
3281 * Drain objects in the per cpu slab, while disabling local
3282 * IRQs, which protects against PREEMPT and interrupts
3283 * handlers invoking normal fastpath.
3284 */
3285 local_irq_disable();
3286 c = this_cpu_ptr(s->cpu_slab);
3287
3288 for (i = 0; i < size; i++) {
3289 void *object = c->freelist;
3290
3291 if (unlikely(!object)) {
3292 /*
3293 * We may have removed an object from c->freelist using
3294 * the fastpath in the previous iteration; in that case,
3295 * c->tid has not been bumped yet.
3296 * Since ___slab_alloc() may reenable interrupts while
3297 * allocating memory, we should bump c->tid now.
3298 */
3299 c->tid = next_tid(c->tid);
3300
3301 /*
3302 * Invoking slow path likely have side-effect
3303 * of re-populating per CPU c->freelist
3304 */
3305 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3306 _RET_IP_, c);
3307 if (unlikely(!p[i]))
3308 goto error;
3309
3310 c = this_cpu_ptr(s->cpu_slab);
3311 maybe_wipe_obj_freeptr(s, p[i]);
3312
3313 continue; /* goto for-loop */
3314 }
3315 c->freelist = get_freepointer(s, object);
3316 p[i] = object;
3317 maybe_wipe_obj_freeptr(s, p[i]);
3318 }
3319 c->tid = next_tid(c->tid);
3320 local_irq_enable();
3321
3322 /* Clear memory outside IRQ disabled fastpath loop */
3323 if (unlikely(slab_want_init_on_alloc(flags, s))) {
3324 int j;
3325
3326 for (j = 0; j < i; j++)
3327 memset(p[j], 0, s->object_size);
3328 }
3329
3330 /* memcg and kmem_cache debug support */
3331 slab_post_alloc_hook(s, objcg, flags, size, p);
3332 return i;
3333error:
3334 local_irq_enable();
3335 slab_post_alloc_hook(s, objcg, flags, i, p);
3336 __kmem_cache_free_bulk(s, i, p);
3337 return 0;
3338}
3339EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3340
3341
3342/*
3343 * Object placement in a slab is made very easy because we always start at
3344 * offset 0. If we tune the size of the object to the alignment then we can
3345 * get the required alignment by putting one properly sized object after
3346 * another.
3347 *
3348 * Notice that the allocation order determines the sizes of the per cpu
3349 * caches. Each processor has always one slab available for allocations.
3350 * Increasing the allocation order reduces the number of times that slabs
3351 * must be moved on and off the partial lists and is therefore a factor in
3352 * locking overhead.
3353 */
3354
3355/*
3356 * Mininum / Maximum order of slab pages. This influences locking overhead
3357 * and slab fragmentation. A higher order reduces the number of partial slabs
3358 * and increases the number of allocations possible without having to
3359 * take the list_lock.
3360 */
3361static unsigned int slub_min_order;
3362static unsigned int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3363static unsigned int slub_min_objects;
3364
3365/*
3366 * Calculate the order of allocation given an slab object size.
3367 *
3368 * The order of allocation has significant impact on performance and other
3369 * system components. Generally order 0 allocations should be preferred since
3370 * order 0 does not cause fragmentation in the page allocator. Larger objects
3371 * be problematic to put into order 0 slabs because there may be too much
3372 * unused space left. We go to a higher order if more than 1/16th of the slab
3373 * would be wasted.
3374 *
3375 * In order to reach satisfactory performance we must ensure that a minimum
3376 * number of objects is in one slab. Otherwise we may generate too much
3377 * activity on the partial lists which requires taking the list_lock. This is
3378 * less a concern for large slabs though which are rarely used.
3379 *
3380 * slub_max_order specifies the order where we begin to stop considering the
3381 * number of objects in a slab as critical. If we reach slub_max_order then
3382 * we try to keep the page order as low as possible. So we accept more waste
3383 * of space in favor of a small page order.
3384 *
3385 * Higher order allocations also allow the placement of more objects in a
3386 * slab and thereby reduce object handling overhead. If the user has
3387 * requested a higher mininum order then we start with that one instead of
3388 * the smallest order which will fit the object.
3389 */
3390static inline unsigned int slab_order(unsigned int size,
3391 unsigned int min_objects, unsigned int max_order,
3392 unsigned int fract_leftover)
3393{
3394 unsigned int min_order = slub_min_order;
3395 unsigned int order;
3396
3397 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
3398 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3399
3400 for (order = max(min_order, (unsigned int)get_order(min_objects * size));
3401 order <= max_order; order++) {
3402
3403 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
3404 unsigned int rem;
3405
3406 rem = slab_size % size;
3407
3408 if (rem <= slab_size / fract_leftover)
3409 break;
3410 }
3411
3412 return order;
3413}
3414
3415static inline int calculate_order(unsigned int size)
3416{
3417 unsigned int order;
3418 unsigned int min_objects;
3419 unsigned int max_objects;
3420
3421 /*
3422 * Attempt to find best configuration for a slab. This
3423 * works by first attempting to generate a layout with
3424 * the best configuration and backing off gradually.
3425 *
3426 * First we increase the acceptable waste in a slab. Then
3427 * we reduce the minimum objects required in a slab.
3428 */
3429 min_objects = slub_min_objects;
3430 if (!min_objects)
3431 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3432 max_objects = order_objects(slub_max_order, size);
3433 min_objects = min(min_objects, max_objects);
3434
3435 while (min_objects > 1) {
3436 unsigned int fraction;
3437
3438 fraction = 16;
3439 while (fraction >= 4) {
3440 order = slab_order(size, min_objects,
3441 slub_max_order, fraction);
3442 if (order <= slub_max_order)
3443 return order;
3444 fraction /= 2;
3445 }
3446 min_objects--;
3447 }
3448
3449 /*
3450 * We were unable to place multiple objects in a slab. Now
3451 * lets see if we can place a single object there.
3452 */
3453 order = slab_order(size, 1, slub_max_order, 1);
3454 if (order <= slub_max_order)
3455 return order;
3456
3457 /*
3458 * Doh this slab cannot be placed using slub_max_order.
3459 */
3460 order = slab_order(size, 1, MAX_ORDER, 1);
3461 if (order < MAX_ORDER)
3462 return order;
3463 return -ENOSYS;
3464}
3465
3466static void
3467init_kmem_cache_node(struct kmem_cache_node *n)
3468{
3469 n->nr_partial = 0;
3470 spin_lock_init(&n->list_lock);
3471 INIT_LIST_HEAD(&n->partial);
3472#ifdef CONFIG_SLUB_DEBUG
3473 atomic_long_set(&n->nr_slabs, 0);
3474 atomic_long_set(&n->total_objects, 0);
3475 INIT_LIST_HEAD(&n->full);
3476#endif
3477}
3478
3479static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3480{
3481 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3482 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3483
3484 /*
3485 * Must align to double word boundary for the double cmpxchg
3486 * instructions to work; see __pcpu_double_call_return_bool().
3487 */
3488 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3489 2 * sizeof(void *));
3490
3491 if (!s->cpu_slab)
3492 return 0;
3493
3494 init_kmem_cache_cpus(s);
3495
3496 return 1;
3497}
3498
3499static struct kmem_cache *kmem_cache_node;
3500
3501/*
3502 * No kmalloc_node yet so do it by hand. We know that this is the first
3503 * slab on the node for this slabcache. There are no concurrent accesses
3504 * possible.
3505 *
3506 * Note that this function only works on the kmem_cache_node
3507 * when allocating for the kmem_cache_node. This is used for bootstrapping
3508 * memory on a fresh node that has no slab structures yet.
3509 */
3510static void early_kmem_cache_node_alloc(int node)
3511{
3512 struct page *page;
3513 struct kmem_cache_node *n;
3514
3515 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3516
3517 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3518
3519 BUG_ON(!page);
3520 if (page_to_nid(page) != node) {
3521 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3522 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3523 }
3524
3525 n = page->freelist;
3526 BUG_ON(!n);
3527#ifdef CONFIG_SLUB_DEBUG
3528 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3529 init_tracking(kmem_cache_node, n);
3530#endif
3531 n = kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node),
3532 GFP_KERNEL);
3533 page->freelist = get_freepointer(kmem_cache_node, n);
3534 page->inuse = 1;
3535 page->frozen = 0;
3536 kmem_cache_node->node[node] = n;
3537 init_kmem_cache_node(n);
3538 inc_slabs_node(kmem_cache_node, node, page->objects);
3539
3540 /*
3541 * No locks need to be taken here as it has just been
3542 * initialized and there is no concurrent access.
3543 */
3544 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3545}
3546
3547static void free_kmem_cache_nodes(struct kmem_cache *s)
3548{
3549 int node;
3550 struct kmem_cache_node *n;
3551
3552 for_each_kmem_cache_node(s, node, n) {
3553 s->node[node] = NULL;
3554 kmem_cache_free(kmem_cache_node, n);
3555 }
3556}
3557
3558void __kmem_cache_release(struct kmem_cache *s)
3559{
3560 cache_random_seq_destroy(s);
3561 free_percpu(s->cpu_slab);
3562 free_kmem_cache_nodes(s);
3563}
3564
3565static int init_kmem_cache_nodes(struct kmem_cache *s)
3566{
3567 int node;
3568
3569 for_each_node_state(node, N_NORMAL_MEMORY) {
3570 struct kmem_cache_node *n;
3571
3572 if (slab_state == DOWN) {
3573 early_kmem_cache_node_alloc(node);
3574 continue;
3575 }
3576 n = kmem_cache_alloc_node(kmem_cache_node,
3577 GFP_KERNEL, node);
3578
3579 if (!n) {
3580 free_kmem_cache_nodes(s);
3581 return 0;
3582 }
3583
3584 init_kmem_cache_node(n);
3585 s->node[node] = n;
3586 }
3587 return 1;
3588}
3589
3590static void set_min_partial(struct kmem_cache *s, unsigned long min)
3591{
3592 if (min < MIN_PARTIAL)
3593 min = MIN_PARTIAL;
3594 else if (min > MAX_PARTIAL)
3595 min = MAX_PARTIAL;
3596 s->min_partial = min;
3597}
3598
3599static void set_cpu_partial(struct kmem_cache *s)
3600{
3601#ifdef CONFIG_SLUB_CPU_PARTIAL
3602 /*
3603 * cpu_partial determined the maximum number of objects kept in the
3604 * per cpu partial lists of a processor.
3605 *
3606 * Per cpu partial lists mainly contain slabs that just have one
3607 * object freed. If they are used for allocation then they can be
3608 * filled up again with minimal effort. The slab will never hit the
3609 * per node partial lists and therefore no locking will be required.
3610 *
3611 * This setting also determines
3612 *
3613 * A) The number of objects from per cpu partial slabs dumped to the
3614 * per node list when we reach the limit.
3615 * B) The number of objects in cpu partial slabs to extract from the
3616 * per node list when we run out of per cpu objects. We only fetch
3617 * 50% to keep some capacity around for frees.
3618 */
3619 if (!kmem_cache_has_cpu_partial(s))
3620 slub_set_cpu_partial(s, 0);
3621 else if (s->size >= PAGE_SIZE)
3622 slub_set_cpu_partial(s, 2);
3623 else if (s->size >= 1024)
3624 slub_set_cpu_partial(s, 6);
3625 else if (s->size >= 256)
3626 slub_set_cpu_partial(s, 13);
3627 else
3628 slub_set_cpu_partial(s, 30);
3629#endif
3630}
3631
3632/*
3633 * calculate_sizes() determines the order and the distribution of data within
3634 * a slab object.
3635 */
3636static int calculate_sizes(struct kmem_cache *s, int forced_order)
3637{
3638 slab_flags_t flags = s->flags;
3639 unsigned int size = s->object_size;
3640 unsigned int freepointer_area;
3641 unsigned int order;
3642
3643 /*
3644 * Round up object size to the next word boundary. We can only
3645 * place the free pointer at word boundaries and this determines
3646 * the possible location of the free pointer.
3647 */
3648 size = ALIGN(size, sizeof(void *));
3649 /*
3650 * This is the area of the object where a freepointer can be
3651 * safely written. If redzoning adds more to the inuse size, we
3652 * can't use that portion for writing the freepointer, so
3653 * s->offset must be limited within this for the general case.
3654 */
3655 freepointer_area = size;
3656
3657#ifdef CONFIG_SLUB_DEBUG
3658 /*
3659 * Determine if we can poison the object itself. If the user of
3660 * the slab may touch the object after free or before allocation
3661 * then we should never poison the object itself.
3662 */
3663 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
3664 !s->ctor)
3665 s->flags |= __OBJECT_POISON;
3666 else
3667 s->flags &= ~__OBJECT_POISON;
3668
3669
3670 /*
3671 * If we are Redzoning then check if there is some space between the
3672 * end of the object and the free pointer. If not then add an
3673 * additional word to have some bytes to store Redzone information.
3674 */
3675 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3676 size += sizeof(void *);
3677#endif
3678
3679 /*
3680 * With that we have determined the number of bytes in actual use
3681 * by the object. This is the potential offset to the free pointer.
3682 */
3683 s->inuse = size;
3684
3685 if (((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
3686 s->ctor)) {
3687 /*
3688 * Relocate free pointer after the object if it is not
3689 * permitted to overwrite the first word of the object on
3690 * kmem_cache_free.
3691 *
3692 * This is the case if we do RCU, have a constructor or
3693 * destructor or are poisoning the objects.
3694 *
3695 * The assumption that s->offset >= s->inuse means free
3696 * pointer is outside of the object is used in the
3697 * freeptr_outside_object() function. If that is no
3698 * longer true, the function needs to be modified.
3699 */
3700 s->offset = size;
3701 size += sizeof(void *);
3702 } else if (freepointer_area > sizeof(void *)) {
3703 /*
3704 * Store freelist pointer near middle of object to keep
3705 * it away from the edges of the object to avoid small
3706 * sized over/underflows from neighboring allocations.
3707 */
3708 s->offset = ALIGN(freepointer_area / 2, sizeof(void *));
3709 }
3710
3711#ifdef CONFIG_SLUB_DEBUG
3712 if (flags & SLAB_STORE_USER)
3713 /*
3714 * Need to store information about allocs and frees after
3715 * the object.
3716 */
3717 size += 2 * sizeof(struct track);
3718#endif
3719
3720 kasan_cache_create(s, &size, &s->flags);
3721#ifdef CONFIG_SLUB_DEBUG
3722 if (flags & SLAB_RED_ZONE) {
3723 /*
3724 * Add some empty padding so that we can catch
3725 * overwrites from earlier objects rather than let
3726 * tracking information or the free pointer be
3727 * corrupted if a user writes before the start
3728 * of the object.
3729 */
3730 size += sizeof(void *);
3731
3732 s->red_left_pad = sizeof(void *);
3733 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3734 size += s->red_left_pad;
3735 }
3736#endif
3737
3738 /*
3739 * SLUB stores one object immediately after another beginning from
3740 * offset 0. In order to align the objects we have to simply size
3741 * each object to conform to the alignment.
3742 */
3743 size = ALIGN(size, s->align);
3744 s->size = size;
3745 s->reciprocal_size = reciprocal_value(size);
3746 if (forced_order >= 0)
3747 order = forced_order;
3748 else
3749 order = calculate_order(size);
3750
3751 if ((int)order < 0)
3752 return 0;
3753
3754 s->allocflags = 0;
3755 if (order)
3756 s->allocflags |= __GFP_COMP;
3757
3758 if (s->flags & SLAB_CACHE_DMA)
3759 s->allocflags |= GFP_DMA;
3760
3761 if (s->flags & SLAB_CACHE_DMA32)
3762 s->allocflags |= GFP_DMA32;
3763
3764 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3765 s->allocflags |= __GFP_RECLAIMABLE;
3766
3767 /*
3768 * Determine the number of objects per slab
3769 */
3770 s->oo = oo_make(order, size);
3771 s->min = oo_make(get_order(size), size);
3772 if (oo_objects(s->oo) > oo_objects(s->max))
3773 s->max = s->oo;
3774
3775 return !!oo_objects(s->oo);
3776}
3777
3778static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
3779{
3780 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3781#ifdef CONFIG_SLAB_FREELIST_HARDENED
3782 s->random = get_random_long();
3783#endif
3784
3785 if (!calculate_sizes(s, -1))
3786 goto error;
3787 if (disable_higher_order_debug) {
3788 /*
3789 * Disable debugging flags that store metadata if the min slab
3790 * order increased.
3791 */
3792 if (get_order(s->size) > get_order(s->object_size)) {
3793 s->flags &= ~DEBUG_METADATA_FLAGS;
3794 s->offset = 0;
3795 if (!calculate_sizes(s, -1))
3796 goto error;
3797 }
3798 }
3799
3800#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3801 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3802 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
3803 /* Enable fast mode */
3804 s->flags |= __CMPXCHG_DOUBLE;
3805#endif
3806
3807 /*
3808 * The larger the object size is, the more pages we want on the partial
3809 * list to avoid pounding the page allocator excessively.
3810 */
3811 set_min_partial(s, ilog2(s->size) / 2);
3812
3813 set_cpu_partial(s);
3814
3815#ifdef CONFIG_NUMA
3816 s->remote_node_defrag_ratio = 1000;
3817#endif
3818
3819 /* Initialize the pre-computed randomized freelist if slab is up */
3820 if (slab_state >= UP) {
3821 if (init_cache_random_seq(s))
3822 goto error;
3823 }
3824
3825 if (!init_kmem_cache_nodes(s))
3826 goto error;
3827
3828 if (alloc_kmem_cache_cpus(s))
3829 return 0;
3830
3831 free_kmem_cache_nodes(s);
3832error:
3833 return -EINVAL;
3834}
3835
3836static void list_slab_objects(struct kmem_cache *s, struct page *page,
3837 const char *text)
3838{
3839#ifdef CONFIG_SLUB_DEBUG
3840 void *addr = page_address(page);
3841 unsigned long *map;
3842 void *p;
3843
3844 slab_err(s, page, text, s->name);
3845 slab_lock(page);
3846
3847 map = get_map(s, page);
3848 for_each_object(p, s, addr, page->objects) {
3849
3850 if (!test_bit(__obj_to_index(s, addr, p), map)) {
3851 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3852 print_tracking(s, p);
3853 }
3854 }
3855 put_map(map);
3856 slab_unlock(page);
3857#endif
3858}
3859
3860/*
3861 * Attempt to free all partial slabs on a node.
3862 * This is called from __kmem_cache_shutdown(). We must take list_lock
3863 * because sysfs file might still access partial list after the shutdowning.
3864 */
3865static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3866{
3867 LIST_HEAD(discard);
3868 struct page *page, *h;
3869
3870 BUG_ON(irqs_disabled());
3871 spin_lock_irq(&n->list_lock);
3872 list_for_each_entry_safe(page, h, &n->partial, slab_list) {
3873 if (!page->inuse) {
3874 remove_partial(n, page);
3875 list_add(&page->slab_list, &discard);
3876 } else {
3877 list_slab_objects(s, page,
3878 "Objects remaining in %s on __kmem_cache_shutdown()");
3879 }
3880 }
3881 spin_unlock_irq(&n->list_lock);
3882
3883 list_for_each_entry_safe(page, h, &discard, slab_list)
3884 discard_slab(s, page);
3885}
3886
3887bool __kmem_cache_empty(struct kmem_cache *s)
3888{
3889 int node;
3890 struct kmem_cache_node *n;
3891
3892 for_each_kmem_cache_node(s, node, n)
3893 if (n->nr_partial || slabs_node(s, node))
3894 return false;
3895 return true;
3896}
3897
3898/*
3899 * Release all resources used by a slab cache.
3900 */
3901int __kmem_cache_shutdown(struct kmem_cache *s)
3902{
3903 int node;
3904 struct kmem_cache_node *n;
3905
3906 flush_all(s);
3907 /* Attempt to free all objects */
3908 for_each_kmem_cache_node(s, node, n) {
3909 free_partial(s, n);
3910 if (n->nr_partial || slabs_node(s, node))
3911 return 1;
3912 }
3913 return 0;
3914}
3915
3916/********************************************************************
3917 * Kmalloc subsystem
3918 *******************************************************************/
3919
3920static int __init setup_slub_min_order(char *str)
3921{
3922 get_option(&str, (int *)&slub_min_order);
3923
3924 return 1;
3925}
3926
3927__setup("slub_min_order=", setup_slub_min_order);
3928
3929static int __init setup_slub_max_order(char *str)
3930{
3931 get_option(&str, (int *)&slub_max_order);
3932 slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1);
3933
3934 return 1;
3935}
3936
3937__setup("slub_max_order=", setup_slub_max_order);
3938
3939static int __init setup_slub_min_objects(char *str)
3940{
3941 get_option(&str, (int *)&slub_min_objects);
3942
3943 return 1;
3944}
3945
3946__setup("slub_min_objects=", setup_slub_min_objects);
3947
3948void *__kmalloc(size_t size, gfp_t flags)
3949{
3950 struct kmem_cache *s;
3951 void *ret;
3952
3953 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3954 return kmalloc_large(size, flags);
3955
3956 s = kmalloc_slab(size, flags);
3957
3958 if (unlikely(ZERO_OR_NULL_PTR(s)))
3959 return s;
3960
3961 ret = slab_alloc(s, flags, _RET_IP_);
3962
3963 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3964
3965 ret = kasan_kmalloc(s, ret, size, flags);
3966
3967 return ret;
3968}
3969EXPORT_SYMBOL(__kmalloc);
3970
3971#ifdef CONFIG_NUMA
3972static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3973{
3974 struct page *page;
3975 void *ptr = NULL;
3976 unsigned int order = get_order(size);
3977
3978 flags |= __GFP_COMP;
3979 page = alloc_pages_node(node, flags, order);
3980 if (page) {
3981 ptr = page_address(page);
3982 mod_node_page_state(page_pgdat(page), NR_SLAB_UNRECLAIMABLE_B,
3983 PAGE_SIZE << order);
3984 }
3985
3986 return kmalloc_large_node_hook(ptr, size, flags);
3987}
3988
3989void *__kmalloc_node(size_t size, gfp_t flags, int node)
3990{
3991 struct kmem_cache *s;
3992 void *ret;
3993
3994 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3995 ret = kmalloc_large_node(size, flags, node);
3996
3997 trace_kmalloc_node(_RET_IP_, ret,
3998 size, PAGE_SIZE << get_order(size),
3999 flags, node);
4000
4001 return ret;
4002 }
4003
4004 s = kmalloc_slab(size, flags);
4005
4006 if (unlikely(ZERO_OR_NULL_PTR(s)))
4007 return s;
4008
4009 ret = slab_alloc_node(s, flags, node, _RET_IP_);
4010
4011 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
4012
4013 ret = kasan_kmalloc(s, ret, size, flags);
4014
4015 return ret;
4016}
4017EXPORT_SYMBOL(__kmalloc_node);
4018#endif /* CONFIG_NUMA */
4019
4020#ifdef CONFIG_HARDENED_USERCOPY
4021/*
4022 * Rejects incorrectly sized objects and objects that are to be copied
4023 * to/from userspace but do not fall entirely within the containing slab
4024 * cache's usercopy region.
4025 *
4026 * Returns NULL if check passes, otherwise const char * to name of cache
4027 * to indicate an error.
4028 */
4029void __check_heap_object(const void *ptr, unsigned long n, struct page *page,
4030 bool to_user)
4031{
4032 struct kmem_cache *s;
4033 unsigned int offset;
4034 size_t object_size;
4035
4036 ptr = kasan_reset_tag(ptr);
4037
4038 /* Find object and usable object size. */
4039 s = page->slab_cache;
4040
4041 /* Reject impossible pointers. */
4042 if (ptr < page_address(page))
4043 usercopy_abort("SLUB object not in SLUB page?!", NULL,
4044 to_user, 0, n);
4045
4046 /* Find offset within object. */
4047 offset = (ptr - page_address(page)) % s->size;
4048
4049 /* Adjust for redzone and reject if within the redzone. */
4050 if (kmem_cache_debug_flags(s, SLAB_RED_ZONE)) {
4051 if (offset < s->red_left_pad)
4052 usercopy_abort("SLUB object in left red zone",
4053 s->name, to_user, offset, n);
4054 offset -= s->red_left_pad;
4055 }
4056
4057 /* Allow address range falling entirely within usercopy region. */
4058 if (offset >= s->useroffset &&
4059 offset - s->useroffset <= s->usersize &&
4060 n <= s->useroffset - offset + s->usersize)
4061 return;
4062
4063 /*
4064 * If the copy is still within the allocated object, produce
4065 * a warning instead of rejecting the copy. This is intended
4066 * to be a temporary method to find any missing usercopy
4067 * whitelists.
4068 */
4069 object_size = slab_ksize(s);
4070 if (usercopy_fallback &&
4071 offset <= object_size && n <= object_size - offset) {
4072 usercopy_warn("SLUB object", s->name, to_user, offset, n);
4073 return;
4074 }
4075
4076 usercopy_abort("SLUB object", s->name, to_user, offset, n);
4077}
4078#endif /* CONFIG_HARDENED_USERCOPY */
4079
4080size_t __ksize(const void *object)
4081{
4082 struct page *page;
4083
4084 if (unlikely(object == ZERO_SIZE_PTR))
4085 return 0;
4086
4087 page = virt_to_head_page(object);
4088
4089 if (unlikely(!PageSlab(page))) {
4090 WARN_ON(!PageCompound(page));
4091 return page_size(page);
4092 }
4093
4094 return slab_ksize(page->slab_cache);
4095}
4096EXPORT_SYMBOL(__ksize);
4097
4098void kfree(const void *x)
4099{
4100 struct page *page;
4101 void *object = (void *)x;
4102
4103 trace_kfree(_RET_IP_, x);
4104
4105 if (unlikely(ZERO_OR_NULL_PTR(x)))
4106 return;
4107
4108 page = virt_to_head_page(x);
4109 if (unlikely(!PageSlab(page))) {
4110 unsigned int order = compound_order(page);
4111
4112 BUG_ON(!PageCompound(page));
4113 kfree_hook(object);
4114 mod_node_page_state(page_pgdat(page), NR_SLAB_UNRECLAIMABLE_B,
4115 -(PAGE_SIZE << order));
4116 __free_pages(page, order);
4117 return;
4118 }
4119 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
4120}
4121EXPORT_SYMBOL(kfree);
4122
4123#define SHRINK_PROMOTE_MAX 32
4124
4125/*
4126 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
4127 * up most to the head of the partial lists. New allocations will then
4128 * fill those up and thus they can be removed from the partial lists.
4129 *
4130 * The slabs with the least items are placed last. This results in them
4131 * being allocated from last increasing the chance that the last objects
4132 * are freed in them.
4133 */
4134int __kmem_cache_shrink(struct kmem_cache *s)
4135{
4136 int node;
4137 int i;
4138 struct kmem_cache_node *n;
4139 struct page *page;
4140 struct page *t;
4141 struct list_head discard;
4142 struct list_head promote[SHRINK_PROMOTE_MAX];
4143 unsigned long flags;
4144 int ret = 0;
4145
4146 flush_all(s);
4147 for_each_kmem_cache_node(s, node, n) {
4148 INIT_LIST_HEAD(&discard);
4149 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
4150 INIT_LIST_HEAD(promote + i);
4151
4152 spin_lock_irqsave(&n->list_lock, flags);
4153
4154 /*
4155 * Build lists of slabs to discard or promote.
4156 *
4157 * Note that concurrent frees may occur while we hold the
4158 * list_lock. page->inuse here is the upper limit.
4159 */
4160 list_for_each_entry_safe(page, t, &n->partial, slab_list) {
4161 int free = page->objects - page->inuse;
4162
4163 /* Do not reread page->inuse */
4164 barrier();
4165
4166 /* We do not keep full slabs on the list */
4167 BUG_ON(free <= 0);
4168
4169 if (free == page->objects) {
4170 list_move(&page->slab_list, &discard);
4171 n->nr_partial--;
4172 } else if (free <= SHRINK_PROMOTE_MAX)
4173 list_move(&page->slab_list, promote + free - 1);
4174 }
4175
4176 /*
4177 * Promote the slabs filled up most to the head of the
4178 * partial list.
4179 */
4180 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
4181 list_splice(promote + i, &n->partial);
4182
4183 spin_unlock_irqrestore(&n->list_lock, flags);
4184
4185 /* Release empty slabs */
4186 list_for_each_entry_safe(page, t, &discard, slab_list)
4187 discard_slab(s, page);
4188
4189 if (slabs_node(s, node))
4190 ret = 1;
4191 }
4192
4193 return ret;
4194}
4195
4196static int slab_mem_going_offline_callback(void *arg)
4197{
4198 struct kmem_cache *s;
4199
4200 mutex_lock(&slab_mutex);
4201 list_for_each_entry(s, &slab_caches, list)
4202 __kmem_cache_shrink(s);
4203 mutex_unlock(&slab_mutex);
4204
4205 return 0;
4206}
4207
4208static void slab_mem_offline_callback(void *arg)
4209{
4210 struct kmem_cache_node *n;
4211 struct kmem_cache *s;
4212 struct memory_notify *marg = arg;
4213 int offline_node;
4214
4215 offline_node = marg->status_change_nid_normal;
4216
4217 /*
4218 * If the node still has available memory. we need kmem_cache_node
4219 * for it yet.
4220 */
4221 if (offline_node < 0)
4222 return;
4223
4224 mutex_lock(&slab_mutex);
4225 list_for_each_entry(s, &slab_caches, list) {
4226 n = get_node(s, offline_node);
4227 if (n) {
4228 /*
4229 * if n->nr_slabs > 0, slabs still exist on the node
4230 * that is going down. We were unable to free them,
4231 * and offline_pages() function shouldn't call this
4232 * callback. So, we must fail.
4233 */
4234 BUG_ON(slabs_node(s, offline_node));
4235
4236 s->node[offline_node] = NULL;
4237 kmem_cache_free(kmem_cache_node, n);
4238 }
4239 }
4240 mutex_unlock(&slab_mutex);
4241}
4242
4243static int slab_mem_going_online_callback(void *arg)
4244{
4245 struct kmem_cache_node *n;
4246 struct kmem_cache *s;
4247 struct memory_notify *marg = arg;
4248 int nid = marg->status_change_nid_normal;
4249 int ret = 0;
4250
4251 /*
4252 * If the node's memory is already available, then kmem_cache_node is
4253 * already created. Nothing to do.
4254 */
4255 if (nid < 0)
4256 return 0;
4257
4258 /*
4259 * We are bringing a node online. No memory is available yet. We must
4260 * allocate a kmem_cache_node structure in order to bring the node
4261 * online.
4262 */
4263 mutex_lock(&slab_mutex);
4264 list_for_each_entry(s, &slab_caches, list) {
4265 /*
4266 * XXX: kmem_cache_alloc_node will fallback to other nodes
4267 * since memory is not yet available from the node that
4268 * is brought up.
4269 */
4270 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4271 if (!n) {
4272 ret = -ENOMEM;
4273 goto out;
4274 }
4275 init_kmem_cache_node(n);
4276 s->node[nid] = n;
4277 }
4278out:
4279 mutex_unlock(&slab_mutex);
4280 return ret;
4281}
4282
4283static int slab_memory_callback(struct notifier_block *self,
4284 unsigned long action, void *arg)
4285{
4286 int ret = 0;
4287
4288 switch (action) {
4289 case MEM_GOING_ONLINE:
4290 ret = slab_mem_going_online_callback(arg);
4291 break;
4292 case MEM_GOING_OFFLINE:
4293 ret = slab_mem_going_offline_callback(arg);
4294 break;
4295 case MEM_OFFLINE:
4296 case MEM_CANCEL_ONLINE:
4297 slab_mem_offline_callback(arg);
4298 break;
4299 case MEM_ONLINE:
4300 case MEM_CANCEL_OFFLINE:
4301 break;
4302 }
4303 if (ret)
4304 ret = notifier_from_errno(ret);
4305 else
4306 ret = NOTIFY_OK;
4307 return ret;
4308}
4309
4310static struct notifier_block slab_memory_callback_nb = {
4311 .notifier_call = slab_memory_callback,
4312 .priority = SLAB_CALLBACK_PRI,
4313};
4314
4315/********************************************************************
4316 * Basic setup of slabs
4317 *******************************************************************/
4318
4319/*
4320 * Used for early kmem_cache structures that were allocated using
4321 * the page allocator. Allocate them properly then fix up the pointers
4322 * that may be pointing to the wrong kmem_cache structure.
4323 */
4324
4325static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4326{
4327 int node;
4328 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4329 struct kmem_cache_node *n;
4330
4331 memcpy(s, static_cache, kmem_cache->object_size);
4332
4333 /*
4334 * This runs very early, and only the boot processor is supposed to be
4335 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4336 * IPIs around.
4337 */
4338 __flush_cpu_slab(s, smp_processor_id());
4339 for_each_kmem_cache_node(s, node, n) {
4340 struct page *p;
4341
4342 list_for_each_entry(p, &n->partial, slab_list)
4343 p->slab_cache = s;
4344
4345#ifdef CONFIG_SLUB_DEBUG
4346 list_for_each_entry(p, &n->full, slab_list)
4347 p->slab_cache = s;
4348#endif
4349 }
4350 list_add(&s->list, &slab_caches);
4351 return s;
4352}
4353
4354void __init kmem_cache_init(void)
4355{
4356 static __initdata struct kmem_cache boot_kmem_cache,
4357 boot_kmem_cache_node;
4358
4359 if (debug_guardpage_minorder())
4360 slub_max_order = 0;
4361
4362 kmem_cache_node = &boot_kmem_cache_node;
4363 kmem_cache = &boot_kmem_cache;
4364
4365 create_boot_cache(kmem_cache_node, "kmem_cache_node",
4366 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
4367
4368 register_hotmemory_notifier(&slab_memory_callback_nb);
4369
4370 /* Able to allocate the per node structures */
4371 slab_state = PARTIAL;
4372
4373 create_boot_cache(kmem_cache, "kmem_cache",
4374 offsetof(struct kmem_cache, node) +
4375 nr_node_ids * sizeof(struct kmem_cache_node *),
4376 SLAB_HWCACHE_ALIGN, 0, 0);
4377
4378 kmem_cache = bootstrap(&boot_kmem_cache);
4379 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
4380
4381 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4382 setup_kmalloc_cache_index_table();
4383 create_kmalloc_caches(0);
4384
4385 /* Setup random freelists for each cache */
4386 init_freelist_randomization();
4387
4388 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
4389 slub_cpu_dead);
4390
4391 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
4392 cache_line_size(),
4393 slub_min_order, slub_max_order, slub_min_objects,
4394 nr_cpu_ids, nr_node_ids);
4395}
4396
4397void __init kmem_cache_init_late(void)
4398{
4399}
4400
4401struct kmem_cache *
4402__kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
4403 slab_flags_t flags, void (*ctor)(void *))
4404{
4405 struct kmem_cache *s;
4406
4407 s = find_mergeable(size, align, flags, name, ctor);
4408 if (s) {
4409 s->refcount++;
4410
4411 /*
4412 * Adjust the object sizes so that we clear
4413 * the complete object on kzalloc.
4414 */
4415 s->object_size = max(s->object_size, size);
4416 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
4417
4418 if (sysfs_slab_alias(s, name)) {
4419 s->refcount--;
4420 s = NULL;
4421 }
4422 }
4423
4424 return s;
4425}
4426
4427int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
4428{
4429 int err;
4430
4431 err = kmem_cache_open(s, flags);
4432 if (err)
4433 return err;
4434
4435 /* Mutex is not taken during early boot */
4436 if (slab_state <= UP)
4437 return 0;
4438
4439 err = sysfs_slab_add(s);
4440 if (err)
4441 __kmem_cache_release(s);
4442
4443 return err;
4444}
4445
4446void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4447{
4448 struct kmem_cache *s;
4449 void *ret;
4450
4451 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4452 return kmalloc_large(size, gfpflags);
4453
4454 s = kmalloc_slab(size, gfpflags);
4455
4456 if (unlikely(ZERO_OR_NULL_PTR(s)))
4457 return s;
4458
4459 ret = slab_alloc(s, gfpflags, caller);
4460
4461 /* Honor the call site pointer we received. */
4462 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4463
4464 return ret;
4465}
4466EXPORT_SYMBOL(__kmalloc_track_caller);
4467
4468#ifdef CONFIG_NUMA
4469void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4470 int node, unsigned long caller)
4471{
4472 struct kmem_cache *s;
4473 void *ret;
4474
4475 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4476 ret = kmalloc_large_node(size, gfpflags, node);
4477
4478 trace_kmalloc_node(caller, ret,
4479 size, PAGE_SIZE << get_order(size),
4480 gfpflags, node);
4481
4482 return ret;
4483 }
4484
4485 s = kmalloc_slab(size, gfpflags);
4486
4487 if (unlikely(ZERO_OR_NULL_PTR(s)))
4488 return s;
4489
4490 ret = slab_alloc_node(s, gfpflags, node, caller);
4491
4492 /* Honor the call site pointer we received. */
4493 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4494
4495 return ret;
4496}
4497EXPORT_SYMBOL(__kmalloc_node_track_caller);
4498#endif
4499
4500#ifdef CONFIG_SYSFS
4501static int count_inuse(struct page *page)
4502{
4503 return page->inuse;
4504}
4505
4506static int count_total(struct page *page)
4507{
4508 return page->objects;
4509}
4510#endif
4511
4512#ifdef CONFIG_SLUB_DEBUG
4513static void validate_slab(struct kmem_cache *s, struct page *page)
4514{
4515 void *p;
4516 void *addr = page_address(page);
4517 unsigned long *map;
4518
4519 slab_lock(page);
4520
4521 if (!check_slab(s, page) || !on_freelist(s, page, NULL))
4522 goto unlock;
4523
4524 /* Now we know that a valid freelist exists */
4525 map = get_map(s, page);
4526 for_each_object(p, s, addr, page->objects) {
4527 u8 val = test_bit(__obj_to_index(s, addr, p), map) ?
4528 SLUB_RED_INACTIVE : SLUB_RED_ACTIVE;
4529
4530 if (!check_object(s, page, p, val))
4531 break;
4532 }
4533 put_map(map);
4534unlock:
4535 slab_unlock(page);
4536}
4537
4538static int validate_slab_node(struct kmem_cache *s,
4539 struct kmem_cache_node *n)
4540{
4541 unsigned long count = 0;
4542 struct page *page;
4543 unsigned long flags;
4544
4545 spin_lock_irqsave(&n->list_lock, flags);
4546
4547 list_for_each_entry(page, &n->partial, slab_list) {
4548 validate_slab(s, page);
4549 count++;
4550 }
4551 if (count != n->nr_partial)
4552 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4553 s->name, count, n->nr_partial);
4554
4555 if (!(s->flags & SLAB_STORE_USER))
4556 goto out;
4557
4558 list_for_each_entry(page, &n->full, slab_list) {
4559 validate_slab(s, page);
4560 count++;
4561 }
4562 if (count != atomic_long_read(&n->nr_slabs))
4563 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4564 s->name, count, atomic_long_read(&n->nr_slabs));
4565
4566out:
4567 spin_unlock_irqrestore(&n->list_lock, flags);
4568 return count;
4569}
4570
4571static long validate_slab_cache(struct kmem_cache *s)
4572{
4573 int node;
4574 unsigned long count = 0;
4575 struct kmem_cache_node *n;
4576
4577 flush_all(s);
4578 for_each_kmem_cache_node(s, node, n)
4579 count += validate_slab_node(s, n);
4580
4581 return count;
4582}
4583/*
4584 * Generate lists of code addresses where slabcache objects are allocated
4585 * and freed.
4586 */
4587
4588struct location {
4589 unsigned long count;
4590 unsigned long addr;
4591 long long sum_time;
4592 long min_time;
4593 long max_time;
4594 long min_pid;
4595 long max_pid;
4596 DECLARE_BITMAP(cpus, NR_CPUS);
4597 nodemask_t nodes;
4598};
4599
4600struct loc_track {
4601 unsigned long max;
4602 unsigned long count;
4603 struct location *loc;
4604};
4605
4606static void free_loc_track(struct loc_track *t)
4607{
4608 if (t->max)
4609 free_pages((unsigned long)t->loc,
4610 get_order(sizeof(struct location) * t->max));
4611}
4612
4613static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4614{
4615 struct location *l;
4616 int order;
4617
4618 order = get_order(sizeof(struct location) * max);
4619
4620 l = (void *)__get_free_pages(flags, order);
4621 if (!l)
4622 return 0;
4623
4624 if (t->count) {
4625 memcpy(l, t->loc, sizeof(struct location) * t->count);
4626 free_loc_track(t);
4627 }
4628 t->max = max;
4629 t->loc = l;
4630 return 1;
4631}
4632
4633static int add_location(struct loc_track *t, struct kmem_cache *s,
4634 const struct track *track)
4635{
4636 long start, end, pos;
4637 struct location *l;
4638 unsigned long caddr;
4639 unsigned long age = jiffies - track->when;
4640
4641 start = -1;
4642 end = t->count;
4643
4644 for ( ; ; ) {
4645 pos = start + (end - start + 1) / 2;
4646
4647 /*
4648 * There is nothing at "end". If we end up there
4649 * we need to add something to before end.
4650 */
4651 if (pos == end)
4652 break;
4653
4654 caddr = t->loc[pos].addr;
4655 if (track->addr == caddr) {
4656
4657 l = &t->loc[pos];
4658 l->count++;
4659 if (track->when) {
4660 l->sum_time += age;
4661 if (age < l->min_time)
4662 l->min_time = age;
4663 if (age > l->max_time)
4664 l->max_time = age;
4665
4666 if (track->pid < l->min_pid)
4667 l->min_pid = track->pid;
4668 if (track->pid > l->max_pid)
4669 l->max_pid = track->pid;
4670
4671 cpumask_set_cpu(track->cpu,
4672 to_cpumask(l->cpus));
4673 }
4674 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4675 return 1;
4676 }
4677
4678 if (track->addr < caddr)
4679 end = pos;
4680 else
4681 start = pos;
4682 }
4683
4684 /*
4685 * Not found. Insert new tracking element.
4686 */
4687 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4688 return 0;
4689
4690 l = t->loc + pos;
4691 if (pos < t->count)
4692 memmove(l + 1, l,
4693 (t->count - pos) * sizeof(struct location));
4694 t->count++;
4695 l->count = 1;
4696 l->addr = track->addr;
4697 l->sum_time = age;
4698 l->min_time = age;
4699 l->max_time = age;
4700 l->min_pid = track->pid;
4701 l->max_pid = track->pid;
4702 cpumask_clear(to_cpumask(l->cpus));
4703 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4704 nodes_clear(l->nodes);
4705 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4706 return 1;
4707}
4708
4709static void process_slab(struct loc_track *t, struct kmem_cache *s,
4710 struct page *page, enum track_item alloc)
4711{
4712 void *addr = page_address(page);
4713 void *p;
4714 unsigned long *map;
4715
4716 map = get_map(s, page);
4717 for_each_object(p, s, addr, page->objects)
4718 if (!test_bit(__obj_to_index(s, addr, p), map))
4719 add_location(t, s, get_track(s, p, alloc));
4720 put_map(map);
4721}
4722
4723static int list_locations(struct kmem_cache *s, char *buf,
4724 enum track_item alloc)
4725{
4726 int len = 0;
4727 unsigned long i;
4728 struct loc_track t = { 0, 0, NULL };
4729 int node;
4730 struct kmem_cache_node *n;
4731
4732 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4733 GFP_KERNEL)) {
4734 return sprintf(buf, "Out of memory\n");
4735 }
4736 /* Push back cpu slabs */
4737 flush_all(s);
4738
4739 for_each_kmem_cache_node(s, node, n) {
4740 unsigned long flags;
4741 struct page *page;
4742
4743 if (!atomic_long_read(&n->nr_slabs))
4744 continue;
4745
4746 spin_lock_irqsave(&n->list_lock, flags);
4747 list_for_each_entry(page, &n->partial, slab_list)
4748 process_slab(&t, s, page, alloc);
4749 list_for_each_entry(page, &n->full, slab_list)
4750 process_slab(&t, s, page, alloc);
4751 spin_unlock_irqrestore(&n->list_lock, flags);
4752 }
4753
4754 for (i = 0; i < t.count; i++) {
4755 struct location *l = &t.loc[i];
4756
4757 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4758 break;
4759 len += sprintf(buf + len, "%7ld ", l->count);
4760
4761 if (l->addr)
4762 len += sprintf(buf + len, "%pS", (void *)l->addr);
4763 else
4764 len += sprintf(buf + len, "<not-available>");
4765
4766 if (l->sum_time != l->min_time) {
4767 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4768 l->min_time,
4769 (long)div_u64(l->sum_time, l->count),
4770 l->max_time);
4771 } else
4772 len += sprintf(buf + len, " age=%ld",
4773 l->min_time);
4774
4775 if (l->min_pid != l->max_pid)
4776 len += sprintf(buf + len, " pid=%ld-%ld",
4777 l->min_pid, l->max_pid);
4778 else
4779 len += sprintf(buf + len, " pid=%ld",
4780 l->min_pid);
4781
4782 if (num_online_cpus() > 1 &&
4783 !cpumask_empty(to_cpumask(l->cpus)) &&
4784 len < PAGE_SIZE - 60)
4785 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4786 " cpus=%*pbl",
4787 cpumask_pr_args(to_cpumask(l->cpus)));
4788
4789 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4790 len < PAGE_SIZE - 60)
4791 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4792 " nodes=%*pbl",
4793 nodemask_pr_args(&l->nodes));
4794
4795 len += sprintf(buf + len, "\n");
4796 }
4797
4798 free_loc_track(&t);
4799 if (!t.count)
4800 len += sprintf(buf, "No data\n");
4801 return len;
4802}
4803#endif /* CONFIG_SLUB_DEBUG */
4804
4805#ifdef SLUB_RESILIENCY_TEST
4806static void __init resiliency_test(void)
4807{
4808 u8 *p;
4809 int type = KMALLOC_NORMAL;
4810
4811 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4812
4813 pr_err("SLUB resiliency testing\n");
4814 pr_err("-----------------------\n");
4815 pr_err("A. Corruption after allocation\n");
4816
4817 p = kzalloc(16, GFP_KERNEL);
4818 p[16] = 0x12;
4819 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4820 p + 16);
4821
4822 validate_slab_cache(kmalloc_caches[type][4]);
4823
4824 /* Hmmm... The next two are dangerous */
4825 p = kzalloc(32, GFP_KERNEL);
4826 p[32 + sizeof(void *)] = 0x34;
4827 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4828 p);
4829 pr_err("If allocated object is overwritten then not detectable\n\n");
4830
4831 validate_slab_cache(kmalloc_caches[type][5]);
4832 p = kzalloc(64, GFP_KERNEL);
4833 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4834 *p = 0x56;
4835 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4836 p);
4837 pr_err("If allocated object is overwritten then not detectable\n\n");
4838 validate_slab_cache(kmalloc_caches[type][6]);
4839
4840 pr_err("\nB. Corruption after free\n");
4841 p = kzalloc(128, GFP_KERNEL);
4842 kfree(p);
4843 *p = 0x78;
4844 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4845 validate_slab_cache(kmalloc_caches[type][7]);
4846
4847 p = kzalloc(256, GFP_KERNEL);
4848 kfree(p);
4849 p[50] = 0x9a;
4850 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4851 validate_slab_cache(kmalloc_caches[type][8]);
4852
4853 p = kzalloc(512, GFP_KERNEL);
4854 kfree(p);
4855 p[512] = 0xab;
4856 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4857 validate_slab_cache(kmalloc_caches[type][9]);
4858}
4859#else
4860#ifdef CONFIG_SYSFS
4861static void resiliency_test(void) {};
4862#endif
4863#endif /* SLUB_RESILIENCY_TEST */
4864
4865#ifdef CONFIG_SYSFS
4866enum slab_stat_type {
4867 SL_ALL, /* All slabs */
4868 SL_PARTIAL, /* Only partially allocated slabs */
4869 SL_CPU, /* Only slabs used for cpu caches */
4870 SL_OBJECTS, /* Determine allocated objects not slabs */
4871 SL_TOTAL /* Determine object capacity not slabs */
4872};
4873
4874#define SO_ALL (1 << SL_ALL)
4875#define SO_PARTIAL (1 << SL_PARTIAL)
4876#define SO_CPU (1 << SL_CPU)
4877#define SO_OBJECTS (1 << SL_OBJECTS)
4878#define SO_TOTAL (1 << SL_TOTAL)
4879
4880#ifdef CONFIG_MEMCG
4881static bool memcg_sysfs_enabled = IS_ENABLED(CONFIG_SLUB_MEMCG_SYSFS_ON);
4882
4883static int __init setup_slub_memcg_sysfs(char *str)
4884{
4885 int v;
4886
4887 if (get_option(&str, &v) > 0)
4888 memcg_sysfs_enabled = v;
4889
4890 return 1;
4891}
4892
4893__setup("slub_memcg_sysfs=", setup_slub_memcg_sysfs);
4894#endif
4895
4896static ssize_t show_slab_objects(struct kmem_cache *s,
4897 char *buf, unsigned long flags)
4898{
4899 unsigned long total = 0;
4900 int node;
4901 int x;
4902 unsigned long *nodes;
4903
4904 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
4905 if (!nodes)
4906 return -ENOMEM;
4907
4908 if (flags & SO_CPU) {
4909 int cpu;
4910
4911 for_each_possible_cpu(cpu) {
4912 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4913 cpu);
4914 int node;
4915 struct page *page;
4916
4917 page = READ_ONCE(c->page);
4918 if (!page)
4919 continue;
4920
4921 node = page_to_nid(page);
4922 if (flags & SO_TOTAL)
4923 x = page->objects;
4924 else if (flags & SO_OBJECTS)
4925 x = page->inuse;
4926 else
4927 x = 1;
4928
4929 total += x;
4930 nodes[node] += x;
4931
4932 page = slub_percpu_partial_read_once(c);
4933 if (page) {
4934 node = page_to_nid(page);
4935 if (flags & SO_TOTAL)
4936 WARN_ON_ONCE(1);
4937 else if (flags & SO_OBJECTS)
4938 WARN_ON_ONCE(1);
4939 else
4940 x = page->pages;
4941 total += x;
4942 nodes[node] += x;
4943 }
4944 }
4945 }
4946
4947 /*
4948 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
4949 * already held which will conflict with an existing lock order:
4950 *
4951 * mem_hotplug_lock->slab_mutex->kernfs_mutex
4952 *
4953 * We don't really need mem_hotplug_lock (to hold off
4954 * slab_mem_going_offline_callback) here because slab's memory hot
4955 * unplug code doesn't destroy the kmem_cache->node[] data.
4956 */
4957
4958#ifdef CONFIG_SLUB_DEBUG
4959 if (flags & SO_ALL) {
4960 struct kmem_cache_node *n;
4961
4962 for_each_kmem_cache_node(s, node, n) {
4963
4964 if (flags & SO_TOTAL)
4965 x = atomic_long_read(&n->total_objects);
4966 else if (flags & SO_OBJECTS)
4967 x = atomic_long_read(&n->total_objects) -
4968 count_partial(n, count_free);
4969 else
4970 x = atomic_long_read(&n->nr_slabs);
4971 total += x;
4972 nodes[node] += x;
4973 }
4974
4975 } else
4976#endif
4977 if (flags & SO_PARTIAL) {
4978 struct kmem_cache_node *n;
4979
4980 for_each_kmem_cache_node(s, node, n) {
4981 if (flags & SO_TOTAL)
4982 x = count_partial(n, count_total);
4983 else if (flags & SO_OBJECTS)
4984 x = count_partial(n, count_inuse);
4985 else
4986 x = n->nr_partial;
4987 total += x;
4988 nodes[node] += x;
4989 }
4990 }
4991 x = sprintf(buf, "%lu", total);
4992#ifdef CONFIG_NUMA
4993 for (node = 0; node < nr_node_ids; node++)
4994 if (nodes[node])
4995 x += sprintf(buf + x, " N%d=%lu",
4996 node, nodes[node]);
4997#endif
4998 kfree(nodes);
4999 return x + sprintf(buf + x, "\n");
5000}
5001
5002#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
5003#define to_slab(n) container_of(n, struct kmem_cache, kobj)
5004
5005struct slab_attribute {
5006 struct attribute attr;
5007 ssize_t (*show)(struct kmem_cache *s, char *buf);
5008 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
5009};
5010
5011#define SLAB_ATTR_RO(_name) \
5012 static struct slab_attribute _name##_attr = \
5013 __ATTR(_name, 0400, _name##_show, NULL)
5014
5015#define SLAB_ATTR(_name) \
5016 static struct slab_attribute _name##_attr = \
5017 __ATTR(_name, 0600, _name##_show, _name##_store)
5018
5019static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
5020{
5021 return sprintf(buf, "%u\n", s->size);
5022}
5023SLAB_ATTR_RO(slab_size);
5024
5025static ssize_t align_show(struct kmem_cache *s, char *buf)
5026{
5027 return sprintf(buf, "%u\n", s->align);
5028}
5029SLAB_ATTR_RO(align);
5030
5031static ssize_t object_size_show(struct kmem_cache *s, char *buf)
5032{
5033 return sprintf(buf, "%u\n", s->object_size);
5034}
5035SLAB_ATTR_RO(object_size);
5036
5037static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
5038{
5039 return sprintf(buf, "%u\n", oo_objects(s->oo));
5040}
5041SLAB_ATTR_RO(objs_per_slab);
5042
5043static ssize_t order_show(struct kmem_cache *s, char *buf)
5044{
5045 return sprintf(buf, "%u\n", oo_order(s->oo));
5046}
5047SLAB_ATTR_RO(order);
5048
5049static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
5050{
5051 return sprintf(buf, "%lu\n", s->min_partial);
5052}
5053
5054static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
5055 size_t length)
5056{
5057 unsigned long min;
5058 int err;
5059
5060 err = kstrtoul(buf, 10, &min);
5061 if (err)
5062 return err;
5063
5064 set_min_partial(s, min);
5065 return length;
5066}
5067SLAB_ATTR(min_partial);
5068
5069static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
5070{
5071 return sprintf(buf, "%u\n", slub_cpu_partial(s));
5072}
5073
5074static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
5075 size_t length)
5076{
5077 unsigned int objects;
5078 int err;
5079
5080 err = kstrtouint(buf, 10, &objects);
5081 if (err)
5082 return err;
5083 if (objects && !kmem_cache_has_cpu_partial(s))
5084 return -EINVAL;
5085
5086 slub_set_cpu_partial(s, objects);
5087 flush_all(s);
5088 return length;
5089}
5090SLAB_ATTR(cpu_partial);
5091
5092static ssize_t ctor_show(struct kmem_cache *s, char *buf)
5093{
5094 if (!s->ctor)
5095 return 0;
5096 return sprintf(buf, "%pS\n", s->ctor);
5097}
5098SLAB_ATTR_RO(ctor);
5099
5100static ssize_t aliases_show(struct kmem_cache *s, char *buf)
5101{
5102 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
5103}
5104SLAB_ATTR_RO(aliases);
5105
5106static ssize_t partial_show(struct kmem_cache *s, char *buf)
5107{
5108 return show_slab_objects(s, buf, SO_PARTIAL);
5109}
5110SLAB_ATTR_RO(partial);
5111
5112static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
5113{
5114 return show_slab_objects(s, buf, SO_CPU);
5115}
5116SLAB_ATTR_RO(cpu_slabs);
5117
5118static ssize_t objects_show(struct kmem_cache *s, char *buf)
5119{
5120 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5121}
5122SLAB_ATTR_RO(objects);
5123
5124static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5125{
5126 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5127}
5128SLAB_ATTR_RO(objects_partial);
5129
5130static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5131{
5132 int objects = 0;
5133 int pages = 0;
5134 int cpu;
5135 int len;
5136
5137 for_each_online_cpu(cpu) {
5138 struct page *page;
5139
5140 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5141
5142 if (page) {
5143 pages += page->pages;
5144 objects += page->pobjects;
5145 }
5146 }
5147
5148 len = sprintf(buf, "%d(%d)", objects, pages);
5149
5150#ifdef CONFIG_SMP
5151 for_each_online_cpu(cpu) {
5152 struct page *page;
5153
5154 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5155
5156 if (page && len < PAGE_SIZE - 20)
5157 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
5158 page->pobjects, page->pages);
5159 }
5160#endif
5161 return len + sprintf(buf + len, "\n");
5162}
5163SLAB_ATTR_RO(slabs_cpu_partial);
5164
5165static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5166{
5167 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5168}
5169SLAB_ATTR_RO(reclaim_account);
5170
5171static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5172{
5173 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5174}
5175SLAB_ATTR_RO(hwcache_align);
5176
5177#ifdef CONFIG_ZONE_DMA
5178static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5179{
5180 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5181}
5182SLAB_ATTR_RO(cache_dma);
5183#endif
5184
5185static ssize_t usersize_show(struct kmem_cache *s, char *buf)
5186{
5187 return sprintf(buf, "%u\n", s->usersize);
5188}
5189SLAB_ATTR_RO(usersize);
5190
5191static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5192{
5193 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
5194}
5195SLAB_ATTR_RO(destroy_by_rcu);
5196
5197#ifdef CONFIG_SLUB_DEBUG
5198static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5199{
5200 return show_slab_objects(s, buf, SO_ALL);
5201}
5202SLAB_ATTR_RO(slabs);
5203
5204static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5205{
5206 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5207}
5208SLAB_ATTR_RO(total_objects);
5209
5210static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5211{
5212 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5213}
5214SLAB_ATTR_RO(sanity_checks);
5215
5216static ssize_t trace_show(struct kmem_cache *s, char *buf)
5217{
5218 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5219}
5220SLAB_ATTR_RO(trace);
5221
5222static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5223{
5224 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5225}
5226
5227SLAB_ATTR_RO(red_zone);
5228
5229static ssize_t poison_show(struct kmem_cache *s, char *buf)
5230{
5231 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
5232}
5233
5234SLAB_ATTR_RO(poison);
5235
5236static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5237{
5238 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5239}
5240
5241SLAB_ATTR_RO(store_user);
5242
5243static ssize_t validate_show(struct kmem_cache *s, char *buf)
5244{
5245 return 0;
5246}
5247
5248static ssize_t validate_store(struct kmem_cache *s,
5249 const char *buf, size_t length)
5250{
5251 int ret = -EINVAL;
5252
5253 if (buf[0] == '1') {
5254 ret = validate_slab_cache(s);
5255 if (ret >= 0)
5256 ret = length;
5257 }
5258 return ret;
5259}
5260SLAB_ATTR(validate);
5261
5262static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
5263{
5264 if (!(s->flags & SLAB_STORE_USER))
5265 return -ENOSYS;
5266 return list_locations(s, buf, TRACK_ALLOC);
5267}
5268SLAB_ATTR_RO(alloc_calls);
5269
5270static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
5271{
5272 if (!(s->flags & SLAB_STORE_USER))
5273 return -ENOSYS;
5274 return list_locations(s, buf, TRACK_FREE);
5275}
5276SLAB_ATTR_RO(free_calls);
5277#endif /* CONFIG_SLUB_DEBUG */
5278
5279#ifdef CONFIG_FAILSLAB
5280static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5281{
5282 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5283}
5284SLAB_ATTR_RO(failslab);
5285#endif
5286
5287static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5288{
5289 return 0;
5290}
5291
5292static ssize_t shrink_store(struct kmem_cache *s,
5293 const char *buf, size_t length)
5294{
5295 if (buf[0] == '1')
5296 kmem_cache_shrink(s);
5297 else
5298 return -EINVAL;
5299 return length;
5300}
5301SLAB_ATTR(shrink);
5302
5303#ifdef CONFIG_NUMA
5304static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5305{
5306 return sprintf(buf, "%u\n", s->remote_node_defrag_ratio / 10);
5307}
5308
5309static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5310 const char *buf, size_t length)
5311{
5312 unsigned int ratio;
5313 int err;
5314
5315 err = kstrtouint(buf, 10, &ratio);
5316 if (err)
5317 return err;
5318 if (ratio > 100)
5319 return -ERANGE;
5320
5321 s->remote_node_defrag_ratio = ratio * 10;
5322
5323 return length;
5324}
5325SLAB_ATTR(remote_node_defrag_ratio);
5326#endif
5327
5328#ifdef CONFIG_SLUB_STATS
5329static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5330{
5331 unsigned long sum = 0;
5332 int cpu;
5333 int len;
5334 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
5335
5336 if (!data)
5337 return -ENOMEM;
5338
5339 for_each_online_cpu(cpu) {
5340 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5341
5342 data[cpu] = x;
5343 sum += x;
5344 }
5345
5346 len = sprintf(buf, "%lu", sum);
5347
5348#ifdef CONFIG_SMP
5349 for_each_online_cpu(cpu) {
5350 if (data[cpu] && len < PAGE_SIZE - 20)
5351 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5352 }
5353#endif
5354 kfree(data);
5355 return len + sprintf(buf + len, "\n");
5356}
5357
5358static void clear_stat(struct kmem_cache *s, enum stat_item si)
5359{
5360 int cpu;
5361
5362 for_each_online_cpu(cpu)
5363 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5364}
5365
5366#define STAT_ATTR(si, text) \
5367static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5368{ \
5369 return show_stat(s, buf, si); \
5370} \
5371static ssize_t text##_store(struct kmem_cache *s, \
5372 const char *buf, size_t length) \
5373{ \
5374 if (buf[0] != '0') \
5375 return -EINVAL; \
5376 clear_stat(s, si); \
5377 return length; \
5378} \
5379SLAB_ATTR(text); \
5380
5381STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5382STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5383STAT_ATTR(FREE_FASTPATH, free_fastpath);
5384STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5385STAT_ATTR(FREE_FROZEN, free_frozen);
5386STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5387STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5388STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5389STAT_ATTR(ALLOC_SLAB, alloc_slab);
5390STAT_ATTR(ALLOC_REFILL, alloc_refill);
5391STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5392STAT_ATTR(FREE_SLAB, free_slab);
5393STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5394STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5395STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5396STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5397STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5398STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5399STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5400STAT_ATTR(ORDER_FALLBACK, order_fallback);
5401STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5402STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5403STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5404STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5405STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5406STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5407#endif /* CONFIG_SLUB_STATS */
5408
5409static struct attribute *slab_attrs[] = {
5410 &slab_size_attr.attr,
5411 &object_size_attr.attr,
5412 &objs_per_slab_attr.attr,
5413 &order_attr.attr,
5414 &min_partial_attr.attr,
5415 &cpu_partial_attr.attr,
5416 &objects_attr.attr,
5417 &objects_partial_attr.attr,
5418 &partial_attr.attr,
5419 &cpu_slabs_attr.attr,
5420 &ctor_attr.attr,
5421 &aliases_attr.attr,
5422 &align_attr.attr,
5423 &hwcache_align_attr.attr,
5424 &reclaim_account_attr.attr,
5425 &destroy_by_rcu_attr.attr,
5426 &shrink_attr.attr,
5427 &slabs_cpu_partial_attr.attr,
5428#ifdef CONFIG_SLUB_DEBUG
5429 &total_objects_attr.attr,
5430 &slabs_attr.attr,
5431 &sanity_checks_attr.attr,
5432 &trace_attr.attr,
5433 &red_zone_attr.attr,
5434 &poison_attr.attr,
5435 &store_user_attr.attr,
5436 &validate_attr.attr,
5437 &alloc_calls_attr.attr,
5438 &free_calls_attr.attr,
5439#endif
5440#ifdef CONFIG_ZONE_DMA
5441 &cache_dma_attr.attr,
5442#endif
5443#ifdef CONFIG_NUMA
5444 &remote_node_defrag_ratio_attr.attr,
5445#endif
5446#ifdef CONFIG_SLUB_STATS
5447 &alloc_fastpath_attr.attr,
5448 &alloc_slowpath_attr.attr,
5449 &free_fastpath_attr.attr,
5450 &free_slowpath_attr.attr,
5451 &free_frozen_attr.attr,
5452 &free_add_partial_attr.attr,
5453 &free_remove_partial_attr.attr,
5454 &alloc_from_partial_attr.attr,
5455 &alloc_slab_attr.attr,
5456 &alloc_refill_attr.attr,
5457 &alloc_node_mismatch_attr.attr,
5458 &free_slab_attr.attr,
5459 &cpuslab_flush_attr.attr,
5460 &deactivate_full_attr.attr,
5461 &deactivate_empty_attr.attr,
5462 &deactivate_to_head_attr.attr,
5463 &deactivate_to_tail_attr.attr,
5464 &deactivate_remote_frees_attr.attr,
5465 &deactivate_bypass_attr.attr,
5466 &order_fallback_attr.attr,
5467 &cmpxchg_double_fail_attr.attr,
5468 &cmpxchg_double_cpu_fail_attr.attr,
5469 &cpu_partial_alloc_attr.attr,
5470 &cpu_partial_free_attr.attr,
5471 &cpu_partial_node_attr.attr,
5472 &cpu_partial_drain_attr.attr,
5473#endif
5474#ifdef CONFIG_FAILSLAB
5475 &failslab_attr.attr,
5476#endif
5477 &usersize_attr.attr,
5478
5479 NULL
5480};
5481
5482static const struct attribute_group slab_attr_group = {
5483 .attrs = slab_attrs,
5484};
5485
5486static ssize_t slab_attr_show(struct kobject *kobj,
5487 struct attribute *attr,
5488 char *buf)
5489{
5490 struct slab_attribute *attribute;
5491 struct kmem_cache *s;
5492 int err;
5493
5494 attribute = to_slab_attr(attr);
5495 s = to_slab(kobj);
5496
5497 if (!attribute->show)
5498 return -EIO;
5499
5500 err = attribute->show(s, buf);
5501
5502 return err;
5503}
5504
5505static ssize_t slab_attr_store(struct kobject *kobj,
5506 struct attribute *attr,
5507 const char *buf, size_t len)
5508{
5509 struct slab_attribute *attribute;
5510 struct kmem_cache *s;
5511 int err;
5512
5513 attribute = to_slab_attr(attr);
5514 s = to_slab(kobj);
5515
5516 if (!attribute->store)
5517 return -EIO;
5518
5519 err = attribute->store(s, buf, len);
5520 return err;
5521}
5522
5523static void kmem_cache_release(struct kobject *k)
5524{
5525 slab_kmem_cache_release(to_slab(k));
5526}
5527
5528static const struct sysfs_ops slab_sysfs_ops = {
5529 .show = slab_attr_show,
5530 .store = slab_attr_store,
5531};
5532
5533static struct kobj_type slab_ktype = {
5534 .sysfs_ops = &slab_sysfs_ops,
5535 .release = kmem_cache_release,
5536};
5537
5538static struct kset *slab_kset;
5539
5540static inline struct kset *cache_kset(struct kmem_cache *s)
5541{
5542 return slab_kset;
5543}
5544
5545#define ID_STR_LENGTH 64
5546
5547/* Create a unique string id for a slab cache:
5548 *
5549 * Format :[flags-]size
5550 */
5551static char *create_unique_id(struct kmem_cache *s)
5552{
5553 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5554 char *p = name;
5555
5556 BUG_ON(!name);
5557
5558 *p++ = ':';
5559 /*
5560 * First flags affecting slabcache operations. We will only
5561 * get here for aliasable slabs so we do not need to support
5562 * too many flags. The flags here must cover all flags that
5563 * are matched during merging to guarantee that the id is
5564 * unique.
5565 */
5566 if (s->flags & SLAB_CACHE_DMA)
5567 *p++ = 'd';
5568 if (s->flags & SLAB_CACHE_DMA32)
5569 *p++ = 'D';
5570 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5571 *p++ = 'a';
5572 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5573 *p++ = 'F';
5574 if (s->flags & SLAB_ACCOUNT)
5575 *p++ = 'A';
5576 if (p != name + 1)
5577 *p++ = '-';
5578 p += sprintf(p, "%07u", s->size);
5579
5580 BUG_ON(p > name + ID_STR_LENGTH - 1);
5581 return name;
5582}
5583
5584static int sysfs_slab_add(struct kmem_cache *s)
5585{
5586 int err;
5587 const char *name;
5588 struct kset *kset = cache_kset(s);
5589 int unmergeable = slab_unmergeable(s);
5590
5591 if (!kset) {
5592 kobject_init(&s->kobj, &slab_ktype);
5593 return 0;
5594 }
5595
5596 if (!unmergeable && disable_higher_order_debug &&
5597 (slub_debug & DEBUG_METADATA_FLAGS))
5598 unmergeable = 1;
5599
5600 if (unmergeable) {
5601 /*
5602 * Slabcache can never be merged so we can use the name proper.
5603 * This is typically the case for debug situations. In that
5604 * case we can catch duplicate names easily.
5605 */
5606 sysfs_remove_link(&slab_kset->kobj, s->name);
5607 name = s->name;
5608 } else {
5609 /*
5610 * Create a unique name for the slab as a target
5611 * for the symlinks.
5612 */
5613 name = create_unique_id(s);
5614 }
5615
5616 s->kobj.kset = kset;
5617 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5618 if (err) {
5619 kobject_put(&s->kobj);
5620 goto out;
5621 }
5622
5623 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5624 if (err)
5625 goto out_del_kobj;
5626
5627 if (!unmergeable) {
5628 /* Setup first alias */
5629 sysfs_slab_alias(s, s->name);
5630 }
5631out:
5632 if (!unmergeable)
5633 kfree(name);
5634 return err;
5635out_del_kobj:
5636 kobject_del(&s->kobj);
5637 goto out;
5638}
5639
5640void sysfs_slab_unlink(struct kmem_cache *s)
5641{
5642 if (slab_state >= FULL)
5643 kobject_del(&s->kobj);
5644}
5645
5646void sysfs_slab_release(struct kmem_cache *s)
5647{
5648 if (slab_state >= FULL)
5649 kobject_put(&s->kobj);
5650}
5651
5652/*
5653 * Need to buffer aliases during bootup until sysfs becomes
5654 * available lest we lose that information.
5655 */
5656struct saved_alias {
5657 struct kmem_cache *s;
5658 const char *name;
5659 struct saved_alias *next;
5660};
5661
5662static struct saved_alias *alias_list;
5663
5664static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5665{
5666 struct saved_alias *al;
5667
5668 if (slab_state == FULL) {
5669 /*
5670 * If we have a leftover link then remove it.
5671 */
5672 sysfs_remove_link(&slab_kset->kobj, name);
5673 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5674 }
5675
5676 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5677 if (!al)
5678 return -ENOMEM;
5679
5680 al->s = s;
5681 al->name = name;
5682 al->next = alias_list;
5683 alias_list = al;
5684 return 0;
5685}
5686
5687static int __init slab_sysfs_init(void)
5688{
5689 struct kmem_cache *s;
5690 int err;
5691
5692 mutex_lock(&slab_mutex);
5693
5694 slab_kset = kset_create_and_add("slab", NULL, kernel_kobj);
5695 if (!slab_kset) {
5696 mutex_unlock(&slab_mutex);
5697 pr_err("Cannot register slab subsystem.\n");
5698 return -ENOSYS;
5699 }
5700
5701 slab_state = FULL;
5702
5703 list_for_each_entry(s, &slab_caches, list) {
5704 err = sysfs_slab_add(s);
5705 if (err)
5706 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5707 s->name);
5708 }
5709
5710 while (alias_list) {
5711 struct saved_alias *al = alias_list;
5712
5713 alias_list = alias_list->next;
5714 err = sysfs_slab_alias(al->s, al->name);
5715 if (err)
5716 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5717 al->name);
5718 kfree(al);
5719 }
5720
5721 mutex_unlock(&slab_mutex);
5722 resiliency_test();
5723 return 0;
5724}
5725
5726__initcall(slab_sysfs_init);
5727#endif /* CONFIG_SYSFS */
5728
5729/*
5730 * The /proc/slabinfo ABI
5731 */
5732#ifdef CONFIG_SLUB_DEBUG
5733void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5734{
5735 unsigned long nr_slabs = 0;
5736 unsigned long nr_objs = 0;
5737 unsigned long nr_free = 0;
5738 int node;
5739 struct kmem_cache_node *n;
5740
5741 for_each_kmem_cache_node(s, node, n) {
5742 nr_slabs += node_nr_slabs(n);
5743 nr_objs += node_nr_objs(n);
5744 nr_free += count_partial(n, count_free);
5745 }
5746
5747 sinfo->active_objs = nr_objs - nr_free;
5748 sinfo->num_objs = nr_objs;
5749 sinfo->active_slabs = nr_slabs;
5750 sinfo->num_slabs = nr_slabs;
5751 sinfo->objects_per_slab = oo_objects(s->oo);
5752 sinfo->cache_order = oo_order(s->oo);
5753}
5754
5755void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5756{
5757}
5758
5759ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5760 size_t count, loff_t *ppos)
5761{
5762 return -EIO;
5763}
5764#endif /* CONFIG_SLUB_DEBUG */