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