summaryrefslogtreecommitdiff
path: root/mm/slub.c (plain)
blob: f1827060b18f807bbb3d867bef8f3db052181835
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 __printf(3, 4) 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#ifdef CONFIG_AMLOGIC_SLAB_TRACE
1416 slab_trace_add_page(page, order, s, flags);
1417#endif
1418 return page;
1419}
1420
1421#ifdef CONFIG_SLAB_FREELIST_RANDOM
1422/* Pre-initialize the random sequence cache */
1423static int init_cache_random_seq(struct kmem_cache *s)
1424{
1425 int err;
1426 unsigned long i, count = oo_objects(s->oo);
1427
1428 /* Bailout if already initialised */
1429 if (s->random_seq)
1430 return 0;
1431
1432 err = cache_random_seq_create(s, count, GFP_KERNEL);
1433 if (err) {
1434 pr_err("SLUB: Unable to initialize free list for %s\n",
1435 s->name);
1436 return err;
1437 }
1438
1439 /* Transform to an offset on the set of pages */
1440 if (s->random_seq) {
1441 for (i = 0; i < count; i++)
1442 s->random_seq[i] *= s->size;
1443 }
1444 return 0;
1445}
1446
1447/* Initialize each random sequence freelist per cache */
1448static void __init init_freelist_randomization(void)
1449{
1450 struct kmem_cache *s;
1451
1452 mutex_lock(&slab_mutex);
1453
1454 list_for_each_entry(s, &slab_caches, list)
1455 init_cache_random_seq(s);
1456
1457 mutex_unlock(&slab_mutex);
1458}
1459
1460/* Get the next entry on the pre-computed freelist randomized */
1461static void *next_freelist_entry(struct kmem_cache *s, struct page *page,
1462 unsigned long *pos, void *start,
1463 unsigned long page_limit,
1464 unsigned long freelist_count)
1465{
1466 unsigned int idx;
1467
1468 /*
1469 * If the target page allocation failed, the number of objects on the
1470 * page might be smaller than the usual size defined by the cache.
1471 */
1472 do {
1473 idx = s->random_seq[*pos];
1474 *pos += 1;
1475 if (*pos >= freelist_count)
1476 *pos = 0;
1477 } while (unlikely(idx >= page_limit));
1478
1479 return (char *)start + idx;
1480}
1481
1482/* Shuffle the single linked freelist based on a random pre-computed sequence */
1483static bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1484{
1485 void *start;
1486 void *cur;
1487 void *next;
1488 unsigned long idx, pos, page_limit, freelist_count;
1489
1490 if (page->objects < 2 || !s->random_seq)
1491 return false;
1492
1493 freelist_count = oo_objects(s->oo);
1494 pos = get_random_int() % freelist_count;
1495
1496 page_limit = page->objects * s->size;
1497 start = fixup_red_left(s, page_address(page));
1498
1499 /* First entry is used as the base of the freelist */
1500 cur = next_freelist_entry(s, page, &pos, start, page_limit,
1501 freelist_count);
1502 page->freelist = cur;
1503
1504 for (idx = 1; idx < page->objects; idx++) {
1505 setup_object(s, page, cur);
1506 next = next_freelist_entry(s, page, &pos, start, page_limit,
1507 freelist_count);
1508 set_freepointer(s, cur, next);
1509 cur = next;
1510 }
1511 setup_object(s, page, cur);
1512 set_freepointer(s, cur, NULL);
1513
1514 return true;
1515}
1516#else
1517static inline int init_cache_random_seq(struct kmem_cache *s)
1518{
1519 return 0;
1520}
1521static inline void init_freelist_randomization(void) { }
1522static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1523{
1524 return false;
1525}
1526#endif /* CONFIG_SLAB_FREELIST_RANDOM */
1527
1528static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1529{
1530 struct page *page;
1531 struct kmem_cache_order_objects oo = s->oo;
1532 gfp_t alloc_gfp;
1533 void *start, *p;
1534 int idx, order;
1535 bool shuffle;
1536
1537 flags &= gfp_allowed_mask;
1538
1539 if (gfpflags_allow_blocking(flags))
1540 local_irq_enable();
1541
1542 flags |= s->allocflags;
1543
1544 /*
1545 * Let the initial higher-order allocation fail under memory pressure
1546 * so we fall-back to the minimum order allocation.
1547 */
1548 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1549 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1550 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
1551
1552 page = alloc_slab_page(s, alloc_gfp, node, oo);
1553 if (unlikely(!page)) {
1554 oo = s->min;
1555 alloc_gfp = flags;
1556 /*
1557 * Allocation may have failed due to fragmentation.
1558 * Try a lower order alloc if possible
1559 */
1560 page = alloc_slab_page(s, alloc_gfp, node, oo);
1561 if (unlikely(!page))
1562 goto out;
1563 stat(s, ORDER_FALLBACK);
1564 }
1565
1566 if (kmemcheck_enabled &&
1567 !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1568 int pages = 1 << oo_order(oo);
1569
1570 kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node);
1571
1572 /*
1573 * Objects from caches that have a constructor don't get
1574 * cleared when they're allocated, so we need to do it here.
1575 */
1576 if (s->ctor)
1577 kmemcheck_mark_uninitialized_pages(page, pages);
1578 else
1579 kmemcheck_mark_unallocated_pages(page, pages);
1580 }
1581
1582 page->objects = oo_objects(oo);
1583
1584 order = compound_order(page);
1585 page->slab_cache = s;
1586 __SetPageSlab(page);
1587 if (page_is_pfmemalloc(page))
1588 SetPageSlabPfmemalloc(page);
1589
1590 start = page_address(page);
1591
1592 if (unlikely(s->flags & SLAB_POISON))
1593 memset(start, POISON_INUSE, PAGE_SIZE << order);
1594
1595 kasan_poison_slab(page);
1596
1597 shuffle = shuffle_freelist(s, page);
1598
1599 if (!shuffle) {
1600 for_each_object_idx(p, idx, s, start, page->objects) {
1601 setup_object(s, page, p);
1602 if (likely(idx < page->objects))
1603 set_freepointer(s, p, p + s->size);
1604 else
1605 set_freepointer(s, p, NULL);
1606 }
1607 page->freelist = fixup_red_left(s, start);
1608 }
1609
1610 page->inuse = page->objects;
1611 page->frozen = 1;
1612
1613out:
1614 if (gfpflags_allow_blocking(flags))
1615 local_irq_disable();
1616 if (!page)
1617 return NULL;
1618
1619 mod_zone_page_state(page_zone(page),
1620 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1621 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1622 1 << oo_order(oo));
1623
1624 inc_slabs_node(s, page_to_nid(page), page->objects);
1625
1626 return page;
1627}
1628
1629static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1630{
1631 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1632 gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK;
1633 flags &= ~GFP_SLAB_BUG_MASK;
1634 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
1635 invalid_mask, &invalid_mask, flags, &flags);
1636 }
1637
1638 return allocate_slab(s,
1639 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1640}
1641
1642static void __free_slab(struct kmem_cache *s, struct page *page)
1643{
1644 int order = compound_order(page);
1645 int pages = 1 << order;
1646
1647 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1648 void *p;
1649
1650 slab_pad_check(s, page);
1651 for_each_object(p, s, page_address(page),
1652 page->objects)
1653 check_object(s, page, p, SLUB_RED_INACTIVE);
1654 }
1655
1656 kmemcheck_free_shadow(page, compound_order(page));
1657
1658 mod_zone_page_state(page_zone(page),
1659 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1660 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1661 -pages);
1662
1663 __ClearPageSlabPfmemalloc(page);
1664 __ClearPageSlab(page);
1665
1666 page_mapcount_reset(page);
1667 if (current->reclaim_state)
1668 current->reclaim_state->reclaimed_slab += pages;
1669 memcg_uncharge_slab(page, order, s);
1670#ifdef CONFIG_AMLOGIC_SLAB_TRACE
1671 slab_trace_remove_page(page, order, s);
1672#endif
1673 __free_pages(page, order);
1674}
1675
1676#define need_reserve_slab_rcu \
1677 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1678
1679static void rcu_free_slab(struct rcu_head *h)
1680{
1681 struct page *page;
1682
1683 if (need_reserve_slab_rcu)
1684 page = virt_to_head_page(h);
1685 else
1686 page = container_of((struct list_head *)h, struct page, lru);
1687
1688 __free_slab(page->slab_cache, page);
1689}
1690
1691static void free_slab(struct kmem_cache *s, struct page *page)
1692{
1693 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1694 struct rcu_head *head;
1695
1696 if (need_reserve_slab_rcu) {
1697 int order = compound_order(page);
1698 int offset = (PAGE_SIZE << order) - s->reserved;
1699
1700 VM_BUG_ON(s->reserved != sizeof(*head));
1701 head = page_address(page) + offset;
1702 } else {
1703 head = &page->rcu_head;
1704 }
1705
1706 call_rcu(head, rcu_free_slab);
1707 } else
1708 __free_slab(s, page);
1709}
1710
1711static void discard_slab(struct kmem_cache *s, struct page *page)
1712{
1713 dec_slabs_node(s, page_to_nid(page), page->objects);
1714 free_slab(s, page);
1715}
1716
1717/*
1718 * Management of partially allocated slabs.
1719 */
1720static inline void
1721__add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1722{
1723 n->nr_partial++;
1724 if (tail == DEACTIVATE_TO_TAIL)
1725 list_add_tail(&page->lru, &n->partial);
1726 else
1727 list_add(&page->lru, &n->partial);
1728}
1729
1730static inline void add_partial(struct kmem_cache_node *n,
1731 struct page *page, int tail)
1732{
1733 lockdep_assert_held(&n->list_lock);
1734 __add_partial(n, page, tail);
1735}
1736
1737static inline void remove_partial(struct kmem_cache_node *n,
1738 struct page *page)
1739{
1740 lockdep_assert_held(&n->list_lock);
1741 list_del(&page->lru);
1742 n->nr_partial--;
1743}
1744
1745/*
1746 * Remove slab from the partial list, freeze it and
1747 * return the pointer to the freelist.
1748 *
1749 * Returns a list of objects or NULL if it fails.
1750 */
1751static inline void *acquire_slab(struct kmem_cache *s,
1752 struct kmem_cache_node *n, struct page *page,
1753 int mode, int *objects)
1754{
1755 void *freelist;
1756 unsigned long counters;
1757 struct page new;
1758
1759 lockdep_assert_held(&n->list_lock);
1760
1761 /*
1762 * Zap the freelist and set the frozen bit.
1763 * The old freelist is the list of objects for the
1764 * per cpu allocation list.
1765 */
1766 freelist = page->freelist;
1767 counters = page->counters;
1768 new.counters = counters;
1769 *objects = new.objects - new.inuse;
1770 if (mode) {
1771 new.inuse = page->objects;
1772 new.freelist = NULL;
1773 } else {
1774 new.freelist = freelist;
1775 }
1776
1777 VM_BUG_ON(new.frozen);
1778 new.frozen = 1;
1779
1780 if (!__cmpxchg_double_slab(s, page,
1781 freelist, counters,
1782 new.freelist, new.counters,
1783 "acquire_slab"))
1784 return NULL;
1785
1786 remove_partial(n, page);
1787 WARN_ON(!freelist);
1788 return freelist;
1789}
1790
1791static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1792static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1793
1794/*
1795 * Try to allocate a partial slab from a specific node.
1796 */
1797static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1798 struct kmem_cache_cpu *c, gfp_t flags)
1799{
1800 struct page *page, *page2;
1801 void *object = NULL;
1802 unsigned int available = 0;
1803 int objects;
1804
1805 /*
1806 * Racy check. If we mistakenly see no partial slabs then we
1807 * just allocate an empty slab. If we mistakenly try to get a
1808 * partial slab and there is none available then get_partials()
1809 * will return NULL.
1810 */
1811 if (!n || !n->nr_partial)
1812 return NULL;
1813
1814 spin_lock(&n->list_lock);
1815 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1816 void *t;
1817
1818 if (!pfmemalloc_match(page, flags))
1819 continue;
1820
1821 t = acquire_slab(s, n, page, object == NULL, &objects);
1822 if (!t)
1823 break;
1824
1825 available += objects;
1826 if (!object) {
1827 c->page = page;
1828 stat(s, ALLOC_FROM_PARTIAL);
1829 object = t;
1830 } else {
1831 put_cpu_partial(s, page, 0);
1832 stat(s, CPU_PARTIAL_NODE);
1833 }
1834 if (!kmem_cache_has_cpu_partial(s)
1835 || available > s->cpu_partial / 2)
1836 break;
1837
1838 }
1839 spin_unlock(&n->list_lock);
1840 return object;
1841}
1842
1843/*
1844 * Get a page from somewhere. Search in increasing NUMA distances.
1845 */
1846static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1847 struct kmem_cache_cpu *c)
1848{
1849#ifdef CONFIG_NUMA
1850 struct zonelist *zonelist;
1851 struct zoneref *z;
1852 struct zone *zone;
1853 enum zone_type high_zoneidx = gfp_zone(flags);
1854 void *object;
1855 unsigned int cpuset_mems_cookie;
1856
1857 /*
1858 * The defrag ratio allows a configuration of the tradeoffs between
1859 * inter node defragmentation and node local allocations. A lower
1860 * defrag_ratio increases the tendency to do local allocations
1861 * instead of attempting to obtain partial slabs from other nodes.
1862 *
1863 * If the defrag_ratio is set to 0 then kmalloc() always
1864 * returns node local objects. If the ratio is higher then kmalloc()
1865 * may return off node objects because partial slabs are obtained
1866 * from other nodes and filled up.
1867 *
1868 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
1869 * (which makes defrag_ratio = 1000) then every (well almost)
1870 * allocation will first attempt to defrag slab caches on other nodes.
1871 * This means scanning over all nodes to look for partial slabs which
1872 * may be expensive if we do it every time we are trying to find a slab
1873 * with available objects.
1874 */
1875 if (!s->remote_node_defrag_ratio ||
1876 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1877 return NULL;
1878
1879 do {
1880 cpuset_mems_cookie = read_mems_allowed_begin();
1881 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1882 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1883 struct kmem_cache_node *n;
1884
1885 n = get_node(s, zone_to_nid(zone));
1886
1887 if (n && cpuset_zone_allowed(zone, flags) &&
1888 n->nr_partial > s->min_partial) {
1889 object = get_partial_node(s, n, c, flags);
1890 if (object) {
1891 /*
1892 * Don't check read_mems_allowed_retry()
1893 * here - if mems_allowed was updated in
1894 * parallel, that was a harmless race
1895 * between allocation and the cpuset
1896 * update
1897 */
1898 return object;
1899 }
1900 }
1901 }
1902 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1903#endif
1904 return NULL;
1905}
1906
1907/*
1908 * Get a partial page, lock it and return it.
1909 */
1910static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1911 struct kmem_cache_cpu *c)
1912{
1913 void *object;
1914 int searchnode = node;
1915
1916 if (node == NUMA_NO_NODE)
1917 searchnode = numa_mem_id();
1918 else if (!node_present_pages(node))
1919 searchnode = node_to_mem_node(node);
1920
1921 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1922 if (object || node != NUMA_NO_NODE)
1923 return object;
1924
1925 return get_any_partial(s, flags, c);
1926}
1927
1928#ifdef CONFIG_PREEMPT
1929/*
1930 * Calculate the next globally unique transaction for disambiguiation
1931 * during cmpxchg. The transactions start with the cpu number and are then
1932 * incremented by CONFIG_NR_CPUS.
1933 */
1934#define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1935#else
1936/*
1937 * No preemption supported therefore also no need to check for
1938 * different cpus.
1939 */
1940#define TID_STEP 1
1941#endif
1942
1943static inline unsigned long next_tid(unsigned long tid)
1944{
1945 return tid + TID_STEP;
1946}
1947
1948static inline unsigned int tid_to_cpu(unsigned long tid)
1949{
1950 return tid % TID_STEP;
1951}
1952
1953static inline unsigned long tid_to_event(unsigned long tid)
1954{
1955 return tid / TID_STEP;
1956}
1957
1958static inline unsigned int init_tid(int cpu)
1959{
1960 return cpu;
1961}
1962
1963static inline void note_cmpxchg_failure(const char *n,
1964 const struct kmem_cache *s, unsigned long tid)
1965{
1966#ifdef SLUB_DEBUG_CMPXCHG
1967 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1968
1969 pr_info("%s %s: cmpxchg redo ", n, s->name);
1970
1971#ifdef CONFIG_PREEMPT
1972 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1973 pr_warn("due to cpu change %d -> %d\n",
1974 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1975 else
1976#endif
1977 if (tid_to_event(tid) != tid_to_event(actual_tid))
1978 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1979 tid_to_event(tid), tid_to_event(actual_tid));
1980 else
1981 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1982 actual_tid, tid, next_tid(tid));
1983#endif
1984 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1985}
1986
1987static void init_kmem_cache_cpus(struct kmem_cache *s)
1988{
1989 int cpu;
1990
1991 for_each_possible_cpu(cpu)
1992 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1993}
1994
1995/*
1996 * Remove the cpu slab
1997 */
1998static void deactivate_slab(struct kmem_cache *s, struct page *page,
1999 void *freelist)
2000{
2001 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
2002 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
2003 int lock = 0;
2004 enum slab_modes l = M_NONE, m = M_NONE;
2005 void *nextfree;
2006 int tail = DEACTIVATE_TO_HEAD;
2007 struct page new;
2008 struct page old;
2009
2010 if (page->freelist) {
2011 stat(s, DEACTIVATE_REMOTE_FREES);
2012 tail = DEACTIVATE_TO_TAIL;
2013 }
2014
2015 /*
2016 * Stage one: Free all available per cpu objects back
2017 * to the page freelist while it is still frozen. Leave the
2018 * last one.
2019 *
2020 * There is no need to take the list->lock because the page
2021 * is still frozen.
2022 */
2023 while (freelist && (nextfree = get_freepointer(s, freelist))) {
2024 void *prior;
2025 unsigned long counters;
2026
2027 do {
2028 prior = page->freelist;
2029 counters = page->counters;
2030 set_freepointer(s, freelist, prior);
2031 new.counters = counters;
2032 new.inuse--;
2033 VM_BUG_ON(!new.frozen);
2034
2035 } while (!__cmpxchg_double_slab(s, page,
2036 prior, counters,
2037 freelist, new.counters,
2038 "drain percpu freelist"));
2039
2040 freelist = nextfree;
2041 }
2042
2043 /*
2044 * Stage two: Ensure that the page is unfrozen while the
2045 * list presence reflects the actual number of objects
2046 * during unfreeze.
2047 *
2048 * We setup the list membership and then perform a cmpxchg
2049 * with the count. If there is a mismatch then the page
2050 * is not unfrozen but the page is on the wrong list.
2051 *
2052 * Then we restart the process which may have to remove
2053 * the page from the list that we just put it on again
2054 * because the number of objects in the slab may have
2055 * changed.
2056 */
2057redo:
2058
2059 old.freelist = page->freelist;
2060 old.counters = page->counters;
2061 VM_BUG_ON(!old.frozen);
2062
2063 /* Determine target state of the slab */
2064 new.counters = old.counters;
2065 if (freelist) {
2066 new.inuse--;
2067 set_freepointer(s, freelist, old.freelist);
2068 new.freelist = freelist;
2069 } else
2070 new.freelist = old.freelist;
2071
2072 new.frozen = 0;
2073
2074 if (!new.inuse && n->nr_partial >= s->min_partial)
2075 m = M_FREE;
2076 else if (new.freelist) {
2077 m = M_PARTIAL;
2078 if (!lock) {
2079 lock = 1;
2080 /*
2081 * Taking the spinlock removes the possiblity
2082 * that acquire_slab() will see a slab page that
2083 * is frozen
2084 */
2085 spin_lock(&n->list_lock);
2086 }
2087 } else {
2088 m = M_FULL;
2089 if (kmem_cache_debug(s) && !lock) {
2090 lock = 1;
2091 /*
2092 * This also ensures that the scanning of full
2093 * slabs from diagnostic functions will not see
2094 * any frozen slabs.
2095 */
2096 spin_lock(&n->list_lock);
2097 }
2098 }
2099
2100 if (l != m) {
2101
2102 if (l == M_PARTIAL)
2103
2104 remove_partial(n, page);
2105
2106 else if (l == M_FULL)
2107
2108 remove_full(s, n, page);
2109
2110 if (m == M_PARTIAL) {
2111
2112 add_partial(n, page, tail);
2113 stat(s, tail);
2114
2115 } else if (m == M_FULL) {
2116
2117 stat(s, DEACTIVATE_FULL);
2118 add_full(s, n, page);
2119
2120 }
2121 }
2122
2123 l = m;
2124 if (!__cmpxchg_double_slab(s, page,
2125 old.freelist, old.counters,
2126 new.freelist, new.counters,
2127 "unfreezing slab"))
2128 goto redo;
2129
2130 if (lock)
2131 spin_unlock(&n->list_lock);
2132
2133 if (m == M_FREE) {
2134 stat(s, DEACTIVATE_EMPTY);
2135 discard_slab(s, page);
2136 stat(s, FREE_SLAB);
2137 }
2138}
2139
2140/*
2141 * Unfreeze all the cpu partial slabs.
2142 *
2143 * This function must be called with interrupts disabled
2144 * for the cpu using c (or some other guarantee must be there
2145 * to guarantee no concurrent accesses).
2146 */
2147static void unfreeze_partials(struct kmem_cache *s,
2148 struct kmem_cache_cpu *c)
2149{
2150#ifdef CONFIG_SLUB_CPU_PARTIAL
2151 struct kmem_cache_node *n = NULL, *n2 = NULL;
2152 struct page *page, *discard_page = NULL;
2153
2154 while ((page = c->partial)) {
2155 struct page new;
2156 struct page old;
2157
2158 c->partial = page->next;
2159
2160 n2 = get_node(s, page_to_nid(page));
2161 if (n != n2) {
2162 if (n)
2163 spin_unlock(&n->list_lock);
2164
2165 n = n2;
2166 spin_lock(&n->list_lock);
2167 }
2168
2169 do {
2170
2171 old.freelist = page->freelist;
2172 old.counters = page->counters;
2173 VM_BUG_ON(!old.frozen);
2174
2175 new.counters = old.counters;
2176 new.freelist = old.freelist;
2177
2178 new.frozen = 0;
2179
2180 } while (!__cmpxchg_double_slab(s, page,
2181 old.freelist, old.counters,
2182 new.freelist, new.counters,
2183 "unfreezing slab"));
2184
2185 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2186 page->next = discard_page;
2187 discard_page = page;
2188 } else {
2189 add_partial(n, page, DEACTIVATE_TO_TAIL);
2190 stat(s, FREE_ADD_PARTIAL);
2191 }
2192 }
2193
2194 if (n)
2195 spin_unlock(&n->list_lock);
2196
2197 while (discard_page) {
2198 page = discard_page;
2199 discard_page = discard_page->next;
2200
2201 stat(s, DEACTIVATE_EMPTY);
2202 discard_slab(s, page);
2203 stat(s, FREE_SLAB);
2204 }
2205#endif
2206}
2207
2208/*
2209 * Put a page that was just frozen (in __slab_free) into a partial page
2210 * slot if available. This is done without interrupts disabled and without
2211 * preemption disabled. The cmpxchg is racy and may put the partial page
2212 * onto a random cpus partial slot.
2213 *
2214 * If we did not find a slot then simply move all the partials to the
2215 * per node partial list.
2216 */
2217static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2218{
2219#ifdef CONFIG_SLUB_CPU_PARTIAL
2220 struct page *oldpage;
2221 int pages;
2222 int pobjects;
2223
2224 preempt_disable();
2225 do {
2226 pages = 0;
2227 pobjects = 0;
2228 oldpage = this_cpu_read(s->cpu_slab->partial);
2229
2230 if (oldpage) {
2231 pobjects = oldpage->pobjects;
2232 pages = oldpage->pages;
2233 if (drain && pobjects > s->cpu_partial) {
2234 unsigned long flags;
2235 /*
2236 * partial array is full. Move the existing
2237 * set to the per node partial list.
2238 */
2239 local_irq_save(flags);
2240 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2241 local_irq_restore(flags);
2242 oldpage = NULL;
2243 pobjects = 0;
2244 pages = 0;
2245 stat(s, CPU_PARTIAL_DRAIN);
2246 }
2247 }
2248
2249 pages++;
2250 pobjects += page->objects - page->inuse;
2251
2252 page->pages = pages;
2253 page->pobjects = pobjects;
2254 page->next = oldpage;
2255
2256 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2257 != oldpage);
2258 if (unlikely(!s->cpu_partial)) {
2259 unsigned long flags;
2260
2261 local_irq_save(flags);
2262 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2263 local_irq_restore(flags);
2264 }
2265 preempt_enable();
2266#endif
2267}
2268
2269static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2270{
2271 stat(s, CPUSLAB_FLUSH);
2272 deactivate_slab(s, c->page, c->freelist);
2273
2274 c->tid = next_tid(c->tid);
2275 c->page = NULL;
2276 c->freelist = NULL;
2277}
2278
2279/*
2280 * Flush cpu slab.
2281 *
2282 * Called from IPI handler with interrupts disabled.
2283 */
2284static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2285{
2286 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2287
2288 if (likely(c)) {
2289 if (c->page)
2290 flush_slab(s, c);
2291
2292 unfreeze_partials(s, c);
2293 }
2294}
2295
2296static void flush_cpu_slab(void *d)
2297{
2298 struct kmem_cache *s = d;
2299
2300 __flush_cpu_slab(s, smp_processor_id());
2301}
2302
2303static bool has_cpu_slab(int cpu, void *info)
2304{
2305 struct kmem_cache *s = info;
2306 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2307
2308 return c->page || c->partial;
2309}
2310
2311static void flush_all(struct kmem_cache *s)
2312{
2313 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2314}
2315
2316/*
2317 * Use the cpu notifier to insure that the cpu slabs are flushed when
2318 * necessary.
2319 */
2320static int slub_cpu_dead(unsigned int cpu)
2321{
2322 struct kmem_cache *s;
2323 unsigned long flags;
2324
2325 mutex_lock(&slab_mutex);
2326 list_for_each_entry(s, &slab_caches, list) {
2327 local_irq_save(flags);
2328 __flush_cpu_slab(s, cpu);
2329 local_irq_restore(flags);
2330 }
2331 mutex_unlock(&slab_mutex);
2332 return 0;
2333}
2334
2335/*
2336 * Check if the objects in a per cpu structure fit numa
2337 * locality expectations.
2338 */
2339static inline int node_match(struct page *page, int node)
2340{
2341#ifdef CONFIG_NUMA
2342 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2343 return 0;
2344#endif
2345 return 1;
2346}
2347
2348#ifdef CONFIG_SLUB_DEBUG
2349static int count_free(struct page *page)
2350{
2351 return page->objects - page->inuse;
2352}
2353
2354static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2355{
2356 return atomic_long_read(&n->total_objects);
2357}
2358#endif /* CONFIG_SLUB_DEBUG */
2359
2360#if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2361static unsigned long count_partial(struct kmem_cache_node *n,
2362 int (*get_count)(struct page *))
2363{
2364 unsigned long flags;
2365 unsigned long x = 0;
2366 struct page *page;
2367
2368 spin_lock_irqsave(&n->list_lock, flags);
2369 list_for_each_entry(page, &n->partial, lru)
2370 x += get_count(page);
2371 spin_unlock_irqrestore(&n->list_lock, flags);
2372 return x;
2373}
2374#endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2375
2376static noinline void
2377slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2378{
2379#ifdef CONFIG_SLUB_DEBUG
2380 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2381 DEFAULT_RATELIMIT_BURST);
2382 int node;
2383 struct kmem_cache_node *n;
2384
2385 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2386 return;
2387
2388 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2389 nid, gfpflags, &gfpflags);
2390 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2391 s->name, s->object_size, s->size, oo_order(s->oo),
2392 oo_order(s->min));
2393
2394 if (oo_order(s->min) > get_order(s->object_size))
2395 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2396 s->name);
2397
2398 for_each_kmem_cache_node(s, node, n) {
2399 unsigned long nr_slabs;
2400 unsigned long nr_objs;
2401 unsigned long nr_free;
2402
2403 nr_free = count_partial(n, count_free);
2404 nr_slabs = node_nr_slabs(n);
2405 nr_objs = node_nr_objs(n);
2406
2407 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2408 node, nr_slabs, nr_objs, nr_free);
2409 }
2410#endif
2411}
2412
2413static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2414 int node, struct kmem_cache_cpu **pc)
2415{
2416 void *freelist;
2417 struct kmem_cache_cpu *c = *pc;
2418 struct page *page;
2419
2420 freelist = get_partial(s, flags, node, c);
2421
2422 if (freelist)
2423 return freelist;
2424
2425 page = new_slab(s, flags, node);
2426 if (page) {
2427 c = raw_cpu_ptr(s->cpu_slab);
2428 if (c->page)
2429 flush_slab(s, c);
2430
2431 /*
2432 * No other reference to the page yet so we can
2433 * muck around with it freely without cmpxchg
2434 */
2435 freelist = page->freelist;
2436 page->freelist = NULL;
2437
2438 stat(s, ALLOC_SLAB);
2439 c->page = page;
2440 *pc = c;
2441 } else
2442 freelist = NULL;
2443
2444 return freelist;
2445}
2446
2447static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2448{
2449 if (unlikely(PageSlabPfmemalloc(page)))
2450 return gfp_pfmemalloc_allowed(gfpflags);
2451
2452 return true;
2453}
2454
2455/*
2456 * Check the page->freelist of a page and either transfer the freelist to the
2457 * per cpu freelist or deactivate the page.
2458 *
2459 * The page is still frozen if the return value is not NULL.
2460 *
2461 * If this function returns NULL then the page has been unfrozen.
2462 *
2463 * This function must be called with interrupt disabled.
2464 */
2465static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2466{
2467 struct page new;
2468 unsigned long counters;
2469 void *freelist;
2470
2471 do {
2472 freelist = page->freelist;
2473 counters = page->counters;
2474
2475 new.counters = counters;
2476 VM_BUG_ON(!new.frozen);
2477
2478 new.inuse = page->objects;
2479 new.frozen = freelist != NULL;
2480
2481 } while (!__cmpxchg_double_slab(s, page,
2482 freelist, counters,
2483 NULL, new.counters,
2484 "get_freelist"));
2485
2486 return freelist;
2487}
2488
2489/*
2490 * Slow path. The lockless freelist is empty or we need to perform
2491 * debugging duties.
2492 *
2493 * Processing is still very fast if new objects have been freed to the
2494 * regular freelist. In that case we simply take over the regular freelist
2495 * as the lockless freelist and zap the regular freelist.
2496 *
2497 * If that is not working then we fall back to the partial lists. We take the
2498 * first element of the freelist as the object to allocate now and move the
2499 * rest of the freelist to the lockless freelist.
2500 *
2501 * And if we were unable to get a new slab from the partial slab lists then
2502 * we need to allocate a new slab. This is the slowest path since it involves
2503 * a call to the page allocator and the setup of a new slab.
2504 *
2505 * Version of __slab_alloc to use when we know that interrupts are
2506 * already disabled (which is the case for bulk allocation).
2507 */
2508static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2509 unsigned long addr, struct kmem_cache_cpu *c)
2510{
2511 void *freelist;
2512 struct page *page;
2513
2514 page = c->page;
2515 if (!page)
2516 goto new_slab;
2517redo:
2518
2519 if (unlikely(!node_match(page, node))) {
2520 int searchnode = node;
2521
2522 if (node != NUMA_NO_NODE && !node_present_pages(node))
2523 searchnode = node_to_mem_node(node);
2524
2525 if (unlikely(!node_match(page, searchnode))) {
2526 stat(s, ALLOC_NODE_MISMATCH);
2527 deactivate_slab(s, page, c->freelist);
2528 c->page = NULL;
2529 c->freelist = NULL;
2530 goto new_slab;
2531 }
2532 }
2533
2534 /*
2535 * By rights, we should be searching for a slab page that was
2536 * PFMEMALLOC but right now, we are losing the pfmemalloc
2537 * information when the page leaves the per-cpu allocator
2538 */
2539 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2540 deactivate_slab(s, page, c->freelist);
2541 c->page = NULL;
2542 c->freelist = NULL;
2543 goto new_slab;
2544 }
2545
2546 /* must check again c->freelist in case of cpu migration or IRQ */
2547 freelist = c->freelist;
2548 if (freelist)
2549 goto load_freelist;
2550
2551 freelist = get_freelist(s, page);
2552
2553 if (!freelist) {
2554 c->page = NULL;
2555 stat(s, DEACTIVATE_BYPASS);
2556 goto new_slab;
2557 }
2558
2559 stat(s, ALLOC_REFILL);
2560
2561load_freelist:
2562 /*
2563 * freelist is pointing to the list of objects to be used.
2564 * page is pointing to the page from which the objects are obtained.
2565 * That page must be frozen for per cpu allocations to work.
2566 */
2567 VM_BUG_ON(!c->page->frozen);
2568 c->freelist = get_freepointer(s, freelist);
2569 c->tid = next_tid(c->tid);
2570#ifdef CONFIG_AMLOGIC_SLAB_TRACE
2571 slab_trace_mark_object(freelist, addr, s);
2572#endif
2573 return freelist;
2574
2575new_slab:
2576
2577 if (c->partial) {
2578 page = c->page = c->partial;
2579 c->partial = page->next;
2580 stat(s, CPU_PARTIAL_ALLOC);
2581 c->freelist = NULL;
2582 goto redo;
2583 }
2584
2585 freelist = new_slab_objects(s, gfpflags, node, &c);
2586
2587 if (unlikely(!freelist)) {
2588 slab_out_of_memory(s, gfpflags, node);
2589 return NULL;
2590 }
2591
2592 page = c->page;
2593 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2594 goto load_freelist;
2595
2596 /* Only entered in the debug case */
2597 if (kmem_cache_debug(s) &&
2598 !alloc_debug_processing(s, page, freelist, addr))
2599 goto new_slab; /* Slab failed checks. Next slab needed */
2600
2601 deactivate_slab(s, page, get_freepointer(s, freelist));
2602 c->page = NULL;
2603 c->freelist = NULL;
2604#ifdef CONFIG_AMLOGIC_SLAB_TRACE
2605 slab_trace_mark_object(freelist, addr, s);
2606#endif
2607 return freelist;
2608}
2609
2610/*
2611 * Another one that disabled interrupt and compensates for possible
2612 * cpu changes by refetching the per cpu area pointer.
2613 */
2614static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2615 unsigned long addr, struct kmem_cache_cpu *c)
2616{
2617 void *p;
2618 unsigned long flags;
2619
2620 local_irq_save(flags);
2621#ifdef CONFIG_PREEMPT
2622 /*
2623 * We may have been preempted and rescheduled on a different
2624 * cpu before disabling interrupts. Need to reload cpu area
2625 * pointer.
2626 */
2627 c = this_cpu_ptr(s->cpu_slab);
2628#endif
2629
2630 p = ___slab_alloc(s, gfpflags, node, addr, c);
2631 local_irq_restore(flags);
2632 return p;
2633}
2634
2635/*
2636 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2637 * have the fastpath folded into their functions. So no function call
2638 * overhead for requests that can be satisfied on the fastpath.
2639 *
2640 * The fastpath works by first checking if the lockless freelist can be used.
2641 * If not then __slab_alloc is called for slow processing.
2642 *
2643 * Otherwise we can simply pick the next object from the lockless free list.
2644 */
2645static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2646 gfp_t gfpflags, int node, unsigned long addr)
2647{
2648 void *object;
2649 struct kmem_cache_cpu *c;
2650 struct page *page;
2651 unsigned long tid;
2652
2653 s = slab_pre_alloc_hook(s, gfpflags);
2654 if (!s)
2655 return NULL;
2656redo:
2657 /*
2658 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2659 * enabled. We may switch back and forth between cpus while
2660 * reading from one cpu area. That does not matter as long
2661 * as we end up on the original cpu again when doing the cmpxchg.
2662 *
2663 * We should guarantee that tid and kmem_cache are retrieved on
2664 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2665 * to check if it is matched or not.
2666 */
2667 do {
2668 tid = this_cpu_read(s->cpu_slab->tid);
2669 c = raw_cpu_ptr(s->cpu_slab);
2670 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2671 unlikely(tid != READ_ONCE(c->tid)));
2672
2673 /*
2674 * Irqless object alloc/free algorithm used here depends on sequence
2675 * of fetching cpu_slab's data. tid should be fetched before anything
2676 * on c to guarantee that object and page associated with previous tid
2677 * won't be used with current tid. If we fetch tid first, object and
2678 * page could be one associated with next tid and our alloc/free
2679 * request will be failed. In this case, we will retry. So, no problem.
2680 */
2681 barrier();
2682
2683 /*
2684 * The transaction ids are globally unique per cpu and per operation on
2685 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2686 * occurs on the right processor and that there was no operation on the
2687 * linked list in between.
2688 */
2689
2690 object = c->freelist;
2691 page = c->page;
2692 if (unlikely(!object || !node_match(page, node))) {
2693 object = __slab_alloc(s, gfpflags, node, addr, c);
2694 stat(s, ALLOC_SLOWPATH);
2695 } else {
2696 void *next_object = get_freepointer_safe(s, object);
2697
2698 /*
2699 * The cmpxchg will only match if there was no additional
2700 * operation and if we are on the right processor.
2701 *
2702 * The cmpxchg does the following atomically (without lock
2703 * semantics!)
2704 * 1. Relocate first pointer to the current per cpu area.
2705 * 2. Verify that tid and freelist have not been changed
2706 * 3. If they were not changed replace tid and freelist
2707 *
2708 * Since this is without lock semantics the protection is only
2709 * against code executing on this cpu *not* from access by
2710 * other cpus.
2711 */
2712 if (unlikely(!this_cpu_cmpxchg_double(
2713 s->cpu_slab->freelist, s->cpu_slab->tid,
2714 object, tid,
2715 next_object, next_tid(tid)))) {
2716
2717 note_cmpxchg_failure("slab_alloc", s, tid);
2718 goto redo;
2719 }
2720 prefetch_freepointer(s, next_object);
2721 stat(s, ALLOC_FASTPATH);
2722 #ifdef CONFIG_AMLOGIC_SLAB_TRACE
2723 slab_trace_mark_object(object, addr, s);
2724 #endif
2725 }
2726
2727 if (unlikely(gfpflags & __GFP_ZERO) && object)
2728 memset(object, 0, s->object_size);
2729
2730 slab_post_alloc_hook(s, gfpflags, 1, &object);
2731
2732 return object;
2733}
2734
2735static __always_inline void *slab_alloc(struct kmem_cache *s,
2736 gfp_t gfpflags, unsigned long addr)
2737{
2738 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2739}
2740
2741void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2742{
2743 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2744
2745 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2746 s->size, gfpflags);
2747
2748 return ret;
2749}
2750EXPORT_SYMBOL(kmem_cache_alloc);
2751
2752#ifdef CONFIG_TRACING
2753void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2754{
2755 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2756 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2757 kasan_kmalloc(s, ret, size, gfpflags);
2758 return ret;
2759}
2760EXPORT_SYMBOL(kmem_cache_alloc_trace);
2761#endif
2762
2763#ifdef CONFIG_NUMA
2764void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2765{
2766 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2767
2768 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2769 s->object_size, s->size, gfpflags, node);
2770
2771 return ret;
2772}
2773EXPORT_SYMBOL(kmem_cache_alloc_node);
2774
2775#ifdef CONFIG_TRACING
2776void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2777 gfp_t gfpflags,
2778 int node, size_t size)
2779{
2780 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2781
2782 trace_kmalloc_node(_RET_IP_, ret,
2783 size, s->size, gfpflags, node);
2784
2785 kasan_kmalloc(s, ret, size, gfpflags);
2786 return ret;
2787}
2788EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2789#endif
2790#endif
2791
2792/*
2793 * Slow path handling. This may still be called frequently since objects
2794 * have a longer lifetime than the cpu slabs in most processing loads.
2795 *
2796 * So we still attempt to reduce cache line usage. Just take the slab
2797 * lock and free the item. If there is no additional partial page
2798 * handling required then we can return immediately.
2799 */
2800static void __slab_free(struct kmem_cache *s, struct page *page,
2801 void *head, void *tail, int cnt,
2802 unsigned long addr)
2803
2804{
2805 void *prior;
2806 int was_frozen;
2807 struct page new;
2808 unsigned long counters;
2809 struct kmem_cache_node *n = NULL;
2810 unsigned long uninitialized_var(flags);
2811
2812 stat(s, FREE_SLOWPATH);
2813
2814 if (kmem_cache_debug(s) &&
2815 !free_debug_processing(s, page, head, tail, cnt, addr))
2816 return;
2817
2818 do {
2819 if (unlikely(n)) {
2820 spin_unlock_irqrestore(&n->list_lock, flags);
2821 n = NULL;
2822 }
2823 prior = page->freelist;
2824 counters = page->counters;
2825 set_freepointer(s, tail, prior);
2826 new.counters = counters;
2827 was_frozen = new.frozen;
2828 new.inuse -= cnt;
2829 if ((!new.inuse || !prior) && !was_frozen) {
2830
2831 if (kmem_cache_has_cpu_partial(s) && !prior) {
2832
2833 /*
2834 * Slab was on no list before and will be
2835 * partially empty
2836 * We can defer the list move and instead
2837 * freeze it.
2838 */
2839 new.frozen = 1;
2840
2841 } else { /* Needs to be taken off a list */
2842
2843 n = get_node(s, page_to_nid(page));
2844 /*
2845 * Speculatively acquire the list_lock.
2846 * If the cmpxchg does not succeed then we may
2847 * drop the list_lock without any processing.
2848 *
2849 * Otherwise the list_lock will synchronize with
2850 * other processors updating the list of slabs.
2851 */
2852 spin_lock_irqsave(&n->list_lock, flags);
2853
2854 }
2855 }
2856
2857 } while (!cmpxchg_double_slab(s, page,
2858 prior, counters,
2859 head, new.counters,
2860 "__slab_free"));
2861
2862 if (likely(!n)) {
2863
2864 /*
2865 * If we just froze the page then put it onto the
2866 * per cpu partial list.
2867 */
2868 if (new.frozen && !was_frozen) {
2869 put_cpu_partial(s, page, 1);
2870 stat(s, CPU_PARTIAL_FREE);
2871 }
2872 /*
2873 * The list lock was not taken therefore no list
2874 * activity can be necessary.
2875 */
2876 if (was_frozen)
2877 stat(s, FREE_FROZEN);
2878 return;
2879 }
2880
2881 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2882 goto slab_empty;
2883
2884 /*
2885 * Objects left in the slab. If it was not on the partial list before
2886 * then add it.
2887 */
2888 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2889 if (kmem_cache_debug(s))
2890 remove_full(s, n, page);
2891 add_partial(n, page, DEACTIVATE_TO_TAIL);
2892 stat(s, FREE_ADD_PARTIAL);
2893 }
2894 spin_unlock_irqrestore(&n->list_lock, flags);
2895 return;
2896
2897slab_empty:
2898 if (prior) {
2899 /*
2900 * Slab on the partial list.
2901 */
2902 remove_partial(n, page);
2903 stat(s, FREE_REMOVE_PARTIAL);
2904 } else {
2905 /* Slab must be on the full list */
2906 remove_full(s, n, page);
2907 }
2908
2909 spin_unlock_irqrestore(&n->list_lock, flags);
2910 stat(s, FREE_SLAB);
2911 discard_slab(s, page);
2912}
2913
2914/*
2915 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2916 * can perform fastpath freeing without additional function calls.
2917 *
2918 * The fastpath is only possible if we are freeing to the current cpu slab
2919 * of this processor. This typically the case if we have just allocated
2920 * the item before.
2921 *
2922 * If fastpath is not possible then fall back to __slab_free where we deal
2923 * with all sorts of special processing.
2924 *
2925 * Bulk free of a freelist with several objects (all pointing to the
2926 * same page) possible by specifying head and tail ptr, plus objects
2927 * count (cnt). Bulk free indicated by tail pointer being set.
2928 */
2929static __always_inline void do_slab_free(struct kmem_cache *s,
2930 struct page *page, void *head, void *tail,
2931 int cnt, unsigned long addr)
2932{
2933 void *tail_obj = tail ? : head;
2934 struct kmem_cache_cpu *c;
2935 unsigned long tid;
2936redo:
2937 /*
2938 * Determine the currently cpus per cpu slab.
2939 * The cpu may change afterward. However that does not matter since
2940 * data is retrieved via this pointer. If we are on the same cpu
2941 * during the cmpxchg then the free will succeed.
2942 */
2943 do {
2944 tid = this_cpu_read(s->cpu_slab->tid);
2945 c = raw_cpu_ptr(s->cpu_slab);
2946 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2947 unlikely(tid != READ_ONCE(c->tid)));
2948
2949 /* Same with comment on barrier() in slab_alloc_node() */
2950 barrier();
2951
2952#ifdef CONFIG_AMLOGIC_SLAB_TRACE
2953 slab_trace_remove_object(head, s);
2954#endif
2955 if (likely(page == c->page)) {
2956 set_freepointer(s, tail_obj, c->freelist);
2957
2958 if (unlikely(!this_cpu_cmpxchg_double(
2959 s->cpu_slab->freelist, s->cpu_slab->tid,
2960 c->freelist, tid,
2961 head, next_tid(tid)))) {
2962
2963 note_cmpxchg_failure("slab_free", s, tid);
2964 goto redo;
2965 }
2966 stat(s, FREE_FASTPATH);
2967 } else
2968 __slab_free(s, page, head, tail_obj, cnt, addr);
2969
2970}
2971
2972static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
2973 void *head, void *tail, int cnt,
2974 unsigned long addr)
2975{
2976 slab_free_freelist_hook(s, head, tail);
2977 /*
2978 * slab_free_freelist_hook() could have put the items into quarantine.
2979 * If so, no need to free them.
2980 */
2981 if (s->flags & SLAB_KASAN && !(s->flags & SLAB_DESTROY_BY_RCU))
2982 return;
2983 do_slab_free(s, page, head, tail, cnt, addr);
2984}
2985
2986#ifdef CONFIG_KASAN
2987void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
2988{
2989 do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr);
2990}
2991#endif
2992
2993void kmem_cache_free(struct kmem_cache *s, void *x)
2994{
2995 s = cache_from_obj(s, x);
2996 if (!s)
2997 return;
2998 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
2999 trace_kmem_cache_free(_RET_IP_, x);
3000}
3001EXPORT_SYMBOL(kmem_cache_free);
3002
3003struct detached_freelist {
3004 struct page *page;
3005 void *tail;
3006 void *freelist;
3007 int cnt;
3008 struct kmem_cache *s;
3009};
3010
3011/*
3012 * This function progressively scans the array with free objects (with
3013 * a limited look ahead) and extract objects belonging to the same
3014 * page. It builds a detached freelist directly within the given
3015 * page/objects. This can happen without any need for
3016 * synchronization, because the objects are owned by running process.
3017 * The freelist is build up as a single linked list in the objects.
3018 * The idea is, that this detached freelist can then be bulk
3019 * transferred to the real freelist(s), but only requiring a single
3020 * synchronization primitive. Look ahead in the array is limited due
3021 * to performance reasons.
3022 */
3023static inline
3024int build_detached_freelist(struct kmem_cache *s, size_t size,
3025 void **p, struct detached_freelist *df)
3026{
3027 size_t first_skipped_index = 0;
3028 int lookahead = 3;
3029 void *object;
3030 struct page *page;
3031
3032 /* Always re-init detached_freelist */
3033 df->page = NULL;
3034
3035 do {
3036 object = p[--size];
3037 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3038 } while (!object && size);
3039
3040 if (!object)
3041 return 0;
3042
3043 page = virt_to_head_page(object);
3044 if (!s) {
3045 /* Handle kalloc'ed objects */
3046 if (unlikely(!PageSlab(page))) {
3047 BUG_ON(!PageCompound(page));
3048 kfree_hook(object);
3049 #ifdef CONFIG_AMLOGIC_SLAB_TRACE
3050 slab_trace_remove_page(page, compound_order(page), s);
3051 #endif
3052 __free_pages(page, compound_order(page));
3053 p[size] = NULL; /* mark object processed */
3054 return size;
3055 }
3056 /* Derive kmem_cache from object */
3057 df->s = page->slab_cache;
3058 } else {
3059 df->s = cache_from_obj(s, object); /* Support for memcg */
3060 }
3061
3062 /* Start new detached freelist */
3063 df->page = page;
3064 set_freepointer(df->s, object, NULL);
3065 df->tail = object;
3066 df->freelist = object;
3067 p[size] = NULL; /* mark object processed */
3068 df->cnt = 1;
3069
3070 while (size) {
3071 object = p[--size];
3072 if (!object)
3073 continue; /* Skip processed objects */
3074
3075 /* df->page is always set at this point */
3076 if (df->page == virt_to_head_page(object)) {
3077 /* Opportunity build freelist */
3078 set_freepointer(df->s, object, df->freelist);
3079 df->freelist = object;
3080 df->cnt++;
3081 p[size] = NULL; /* mark object processed */
3082
3083 continue;
3084 }
3085
3086 /* Limit look ahead search */
3087 if (!--lookahead)
3088 break;
3089
3090 if (!first_skipped_index)
3091 first_skipped_index = size + 1;
3092 }
3093
3094 return first_skipped_index;
3095}
3096
3097/* Note that interrupts must be enabled when calling this function. */
3098void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3099{
3100 if (WARN_ON(!size))
3101 return;
3102
3103 do {
3104 struct detached_freelist df;
3105
3106 size = build_detached_freelist(s, size, p, &df);
3107 if (unlikely(!df.page))
3108 continue;
3109
3110 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
3111 } while (likely(size));
3112}
3113EXPORT_SYMBOL(kmem_cache_free_bulk);
3114
3115/* Note that interrupts must be enabled when calling this function. */
3116int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3117 void **p)
3118{
3119 struct kmem_cache_cpu *c;
3120 int i;
3121
3122 /* memcg and kmem_cache debug support */
3123 s = slab_pre_alloc_hook(s, flags);
3124 if (unlikely(!s))
3125 return false;
3126 /*
3127 * Drain objects in the per cpu slab, while disabling local
3128 * IRQs, which protects against PREEMPT and interrupts
3129 * handlers invoking normal fastpath.
3130 */
3131 local_irq_disable();
3132 c = this_cpu_ptr(s->cpu_slab);
3133
3134 for (i = 0; i < size; i++) {
3135 void *object = c->freelist;
3136
3137 if (unlikely(!object)) {
3138 /*
3139 * Invoking slow path likely have side-effect
3140 * of re-populating per CPU c->freelist
3141 */
3142 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3143 _RET_IP_, c);
3144 if (unlikely(!p[i]))
3145 goto error;
3146
3147 c = this_cpu_ptr(s->cpu_slab);
3148 continue; /* goto for-loop */
3149 }
3150 c->freelist = get_freepointer(s, object);
3151 p[i] = object;
3152 }
3153 c->tid = next_tid(c->tid);
3154 local_irq_enable();
3155
3156 /* Clear memory outside IRQ disabled fastpath loop */
3157 if (unlikely(flags & __GFP_ZERO)) {
3158 int j;
3159
3160 for (j = 0; j < i; j++)
3161 memset(p[j], 0, s->object_size);
3162 }
3163
3164 /* memcg and kmem_cache debug support */
3165 slab_post_alloc_hook(s, flags, size, p);
3166 return i;
3167error:
3168 local_irq_enable();
3169 slab_post_alloc_hook(s, flags, i, p);
3170 __kmem_cache_free_bulk(s, i, p);
3171 return 0;
3172}
3173EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3174
3175
3176/*
3177 * Object placement in a slab is made very easy because we always start at
3178 * offset 0. If we tune the size of the object to the alignment then we can
3179 * get the required alignment by putting one properly sized object after
3180 * another.
3181 *
3182 * Notice that the allocation order determines the sizes of the per cpu
3183 * caches. Each processor has always one slab available for allocations.
3184 * Increasing the allocation order reduces the number of times that slabs
3185 * must be moved on and off the partial lists and is therefore a factor in
3186 * locking overhead.
3187 */
3188
3189/*
3190 * Mininum / Maximum order of slab pages. This influences locking overhead
3191 * and slab fragmentation. A higher order reduces the number of partial slabs
3192 * and increases the number of allocations possible without having to
3193 * take the list_lock.
3194 */
3195static int slub_min_order;
3196static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3197static int slub_min_objects;
3198
3199/*
3200 * Calculate the order of allocation given an slab object size.
3201 *
3202 * The order of allocation has significant impact on performance and other
3203 * system components. Generally order 0 allocations should be preferred since
3204 * order 0 does not cause fragmentation in the page allocator. Larger objects
3205 * be problematic to put into order 0 slabs because there may be too much
3206 * unused space left. We go to a higher order if more than 1/16th of the slab
3207 * would be wasted.
3208 *
3209 * In order to reach satisfactory performance we must ensure that a minimum
3210 * number of objects is in one slab. Otherwise we may generate too much
3211 * activity on the partial lists which requires taking the list_lock. This is
3212 * less a concern for large slabs though which are rarely used.
3213 *
3214 * slub_max_order specifies the order where we begin to stop considering the
3215 * number of objects in a slab as critical. If we reach slub_max_order then
3216 * we try to keep the page order as low as possible. So we accept more waste
3217 * of space in favor of a small page order.
3218 *
3219 * Higher order allocations also allow the placement of more objects in a
3220 * slab and thereby reduce object handling overhead. If the user has
3221 * requested a higher mininum order then we start with that one instead of
3222 * the smallest order which will fit the object.
3223 */
3224static inline int slab_order(int size, int min_objects,
3225 int max_order, int fract_leftover, int reserved)
3226{
3227 int order;
3228 int rem;
3229 int min_order = slub_min_order;
3230
3231 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
3232 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3233
3234 for (order = max(min_order, get_order(min_objects * size + reserved));
3235 order <= max_order; order++) {
3236
3237 unsigned long slab_size = PAGE_SIZE << order;
3238
3239 rem = (slab_size - reserved) % size;
3240
3241 if (rem <= slab_size / fract_leftover)
3242 break;
3243 }
3244
3245 return order;
3246}
3247
3248static inline int calculate_order(int size, int reserved)
3249{
3250 int order;
3251 int min_objects;
3252 int fraction;
3253 int max_objects;
3254
3255 /*
3256 * Attempt to find best configuration for a slab. This
3257 * works by first attempting to generate a layout with
3258 * the best configuration and backing off gradually.
3259 *
3260 * First we increase the acceptable waste in a slab. Then
3261 * we reduce the minimum objects required in a slab.
3262 */
3263 min_objects = slub_min_objects;
3264 if (!min_objects)
3265 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3266 max_objects = order_objects(slub_max_order, size, reserved);
3267 min_objects = min(min_objects, max_objects);
3268
3269 while (min_objects > 1) {
3270 fraction = 16;
3271 while (fraction >= 4) {
3272 order = slab_order(size, min_objects,
3273 slub_max_order, fraction, reserved);
3274 if (order <= slub_max_order)
3275 return order;
3276 fraction /= 2;
3277 }
3278 min_objects--;
3279 }
3280
3281 /*
3282 * We were unable to place multiple objects in a slab. Now
3283 * lets see if we can place a single object there.
3284 */
3285 order = slab_order(size, 1, slub_max_order, 1, reserved);
3286 if (order <= slub_max_order)
3287 return order;
3288
3289 /*
3290 * Doh this slab cannot be placed using slub_max_order.
3291 */
3292 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
3293 if (order < MAX_ORDER)
3294 return order;
3295 return -ENOSYS;
3296}
3297
3298static void
3299init_kmem_cache_node(struct kmem_cache_node *n)
3300{
3301 n->nr_partial = 0;
3302 spin_lock_init(&n->list_lock);
3303 INIT_LIST_HEAD(&n->partial);
3304#ifdef CONFIG_SLUB_DEBUG
3305 atomic_long_set(&n->nr_slabs, 0);
3306 atomic_long_set(&n->total_objects, 0);
3307 INIT_LIST_HEAD(&n->full);
3308#endif
3309}
3310
3311static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3312{
3313 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3314 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3315
3316 /*
3317 * Must align to double word boundary for the double cmpxchg
3318 * instructions to work; see __pcpu_double_call_return_bool().
3319 */
3320 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3321 2 * sizeof(void *));
3322
3323 if (!s->cpu_slab)
3324 return 0;
3325
3326 init_kmem_cache_cpus(s);
3327
3328 return 1;
3329}
3330
3331static struct kmem_cache *kmem_cache_node;
3332
3333/*
3334 * No kmalloc_node yet so do it by hand. We know that this is the first
3335 * slab on the node for this slabcache. There are no concurrent accesses
3336 * possible.
3337 *
3338 * Note that this function only works on the kmem_cache_node
3339 * when allocating for the kmem_cache_node. This is used for bootstrapping
3340 * memory on a fresh node that has no slab structures yet.
3341 */
3342static void early_kmem_cache_node_alloc(int node)
3343{
3344 struct page *page;
3345 struct kmem_cache_node *n;
3346
3347 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3348
3349 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3350
3351 BUG_ON(!page);
3352 if (page_to_nid(page) != node) {
3353 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3354 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3355 }
3356
3357 n = page->freelist;
3358 BUG_ON(!n);
3359 page->freelist = get_freepointer(kmem_cache_node, n);
3360 page->inuse = 1;
3361 page->frozen = 0;
3362 kmem_cache_node->node[node] = n;
3363#ifdef CONFIG_SLUB_DEBUG
3364 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3365 init_tracking(kmem_cache_node, n);
3366#endif
3367 kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node),
3368 GFP_KERNEL);
3369 init_kmem_cache_node(n);
3370 inc_slabs_node(kmem_cache_node, node, page->objects);
3371
3372 /*
3373 * No locks need to be taken here as it has just been
3374 * initialized and there is no concurrent access.
3375 */
3376 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3377}
3378
3379static void free_kmem_cache_nodes(struct kmem_cache *s)
3380{
3381 int node;
3382 struct kmem_cache_node *n;
3383
3384 for_each_kmem_cache_node(s, node, n) {
3385 kmem_cache_free(kmem_cache_node, n);
3386 s->node[node] = NULL;
3387 }
3388}
3389
3390void __kmem_cache_release(struct kmem_cache *s)
3391{
3392 cache_random_seq_destroy(s);
3393 free_percpu(s->cpu_slab);
3394 free_kmem_cache_nodes(s);
3395}
3396
3397static int init_kmem_cache_nodes(struct kmem_cache *s)
3398{
3399 int node;
3400
3401 for_each_node_state(node, N_NORMAL_MEMORY) {
3402 struct kmem_cache_node *n;
3403
3404 if (slab_state == DOWN) {
3405 early_kmem_cache_node_alloc(node);
3406 continue;
3407 }
3408 n = kmem_cache_alloc_node(kmem_cache_node,
3409 GFP_KERNEL, node);
3410
3411 if (!n) {
3412 free_kmem_cache_nodes(s);
3413 return 0;
3414 }
3415
3416 s->node[node] = n;
3417 init_kmem_cache_node(n);
3418 }
3419 return 1;
3420}
3421
3422static void set_min_partial(struct kmem_cache *s, unsigned long min)
3423{
3424 if (min < MIN_PARTIAL)
3425 min = MIN_PARTIAL;
3426 else if (min > MAX_PARTIAL)
3427 min = MAX_PARTIAL;
3428 s->min_partial = min;
3429}
3430
3431/*
3432 * calculate_sizes() determines the order and the distribution of data within
3433 * a slab object.
3434 */
3435static int calculate_sizes(struct kmem_cache *s, int forced_order)
3436{
3437 unsigned long flags = s->flags;
3438 size_t size = s->object_size;
3439 int order;
3440
3441 /*
3442 * Round up object size to the next word boundary. We can only
3443 * place the free pointer at word boundaries and this determines
3444 * the possible location of the free pointer.
3445 */
3446 size = ALIGN(size, sizeof(void *));
3447
3448#ifdef CONFIG_SLUB_DEBUG
3449 /*
3450 * Determine if we can poison the object itself. If the user of
3451 * the slab may touch the object after free or before allocation
3452 * then we should never poison the object itself.
3453 */
3454 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
3455 !s->ctor)
3456 s->flags |= __OBJECT_POISON;
3457 else
3458 s->flags &= ~__OBJECT_POISON;
3459
3460
3461 /*
3462 * If we are Redzoning then check if there is some space between the
3463 * end of the object and the free pointer. If not then add an
3464 * additional word to have some bytes to store Redzone information.
3465 */
3466 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3467 size += sizeof(void *);
3468#endif
3469
3470 /*
3471 * With that we have determined the number of bytes in actual use
3472 * by the object. This is the potential offset to the free pointer.
3473 */
3474 s->inuse = size;
3475
3476 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
3477 s->ctor)) {
3478 /*
3479 * Relocate free pointer after the object if it is not
3480 * permitted to overwrite the first word of the object on
3481 * kmem_cache_free.
3482 *
3483 * This is the case if we do RCU, have a constructor or
3484 * destructor or are poisoning the objects.
3485 */
3486 s->offset = size;
3487 size += sizeof(void *);
3488 }
3489
3490#ifdef CONFIG_SLUB_DEBUG
3491 if (flags & SLAB_STORE_USER)
3492 /*
3493 * Need to store information about allocs and frees after
3494 * the object.
3495 */
3496 size += 2 * sizeof(struct track);
3497#endif
3498
3499 kasan_cache_create(s, &size, &s->flags);
3500#ifdef CONFIG_SLUB_DEBUG
3501 if (flags & SLAB_RED_ZONE) {
3502 /*
3503 * Add some empty padding so that we can catch
3504 * overwrites from earlier objects rather than let
3505 * tracking information or the free pointer be
3506 * corrupted if a user writes before the start
3507 * of the object.
3508 */
3509 size += sizeof(void *);
3510
3511 s->red_left_pad = sizeof(void *);
3512 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3513 size += s->red_left_pad;
3514 }
3515#endif
3516
3517 /*
3518 * SLUB stores one object immediately after another beginning from
3519 * offset 0. In order to align the objects we have to simply size
3520 * each object to conform to the alignment.
3521 */
3522 size = ALIGN(size, s->align);
3523 s->size = size;
3524 if (forced_order >= 0)
3525 order = forced_order;
3526 else
3527 order = calculate_order(size, s->reserved);
3528
3529 if (order < 0)
3530 return 0;
3531
3532 s->allocflags = 0;
3533 if (order)
3534 s->allocflags |= __GFP_COMP;
3535
3536 if (s->flags & SLAB_CACHE_DMA)
3537 s->allocflags |= GFP_DMA;
3538
3539 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3540 s->allocflags |= __GFP_RECLAIMABLE;
3541
3542 /*
3543 * Determine the number of objects per slab
3544 */
3545 s->oo = oo_make(order, size, s->reserved);
3546 s->min = oo_make(get_order(size), size, s->reserved);
3547 if (oo_objects(s->oo) > oo_objects(s->max))
3548 s->max = s->oo;
3549
3550 return !!oo_objects(s->oo);
3551}
3552
3553#ifdef CONFIG_AMLOGIC_SLAB_TRACE
3554int get_cache_max_order(struct kmem_cache *s)
3555{
3556 return oo_order(s->oo);
3557}
3558#endif
3559
3560static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3561{
3562 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3563 s->reserved = 0;
3564
3565 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3566 s->reserved = sizeof(struct rcu_head);
3567
3568 if (!calculate_sizes(s, -1))
3569 goto error;
3570 if (disable_higher_order_debug) {
3571 /*
3572 * Disable debugging flags that store metadata if the min slab
3573 * order increased.
3574 */
3575 if (get_order(s->size) > get_order(s->object_size)) {
3576 s->flags &= ~DEBUG_METADATA_FLAGS;
3577 s->offset = 0;
3578 if (!calculate_sizes(s, -1))
3579 goto error;
3580 }
3581 }
3582
3583#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3584 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3585 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
3586 /* Enable fast mode */
3587 s->flags |= __CMPXCHG_DOUBLE;
3588#endif
3589
3590 /*
3591 * The larger the object size is, the more pages we want on the partial
3592 * list to avoid pounding the page allocator excessively.
3593 */
3594 set_min_partial(s, ilog2(s->size) / 2);
3595
3596 /*
3597 * cpu_partial determined the maximum number of objects kept in the
3598 * per cpu partial lists of a processor.
3599 *
3600 * Per cpu partial lists mainly contain slabs that just have one
3601 * object freed. If they are used for allocation then they can be
3602 * filled up again with minimal effort. The slab will never hit the
3603 * per node partial lists and therefore no locking will be required.
3604 *
3605 * This setting also determines
3606 *
3607 * A) The number of objects from per cpu partial slabs dumped to the
3608 * per node list when we reach the limit.
3609 * B) The number of objects in cpu partial slabs to extract from the
3610 * per node list when we run out of per cpu objects. We only fetch
3611 * 50% to keep some capacity around for frees.
3612 */
3613 if (!kmem_cache_has_cpu_partial(s))
3614 s->cpu_partial = 0;
3615 else if (s->size >= PAGE_SIZE)
3616 s->cpu_partial = 2;
3617 else if (s->size >= 1024)
3618 s->cpu_partial = 6;
3619 else if (s->size >= 256)
3620 s->cpu_partial = 13;
3621 else
3622 s->cpu_partial = 30;
3623
3624#ifdef CONFIG_NUMA
3625 s->remote_node_defrag_ratio = 1000;
3626#endif
3627
3628 /* Initialize the pre-computed randomized freelist if slab is up */
3629 if (slab_state >= UP) {
3630 if (init_cache_random_seq(s))
3631 goto error;
3632 }
3633
3634 if (!init_kmem_cache_nodes(s))
3635 goto error;
3636
3637 if (alloc_kmem_cache_cpus(s))
3638 return 0;
3639
3640 free_kmem_cache_nodes(s);
3641error:
3642 if (flags & SLAB_PANIC)
3643 panic("Cannot create slab %s size=%lu realsize=%u order=%u offset=%u flags=%lx\n",
3644 s->name, (unsigned long)s->size, s->size,
3645 oo_order(s->oo), s->offset, flags);
3646 return -EINVAL;
3647}
3648
3649static void list_slab_objects(struct kmem_cache *s, struct page *page,
3650 const char *text)
3651{
3652#ifdef CONFIG_SLUB_DEBUG
3653 void *addr = page_address(page);
3654 void *p;
3655 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3656 sizeof(long), GFP_ATOMIC);
3657 if (!map)
3658 return;
3659 slab_err(s, page, text, s->name);
3660 slab_lock(page);
3661
3662 get_map(s, page, map);
3663 for_each_object(p, s, addr, page->objects) {
3664
3665 if (!test_bit(slab_index(p, s, addr), map)) {
3666 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3667 print_tracking(s, p);
3668 }
3669 }
3670 slab_unlock(page);
3671 kfree(map);
3672#endif
3673}
3674
3675/*
3676 * Attempt to free all partial slabs on a node.
3677 * This is called from __kmem_cache_shutdown(). We must take list_lock
3678 * because sysfs file might still access partial list after the shutdowning.
3679 */
3680static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3681{
3682 LIST_HEAD(discard);
3683 struct page *page, *h;
3684
3685 BUG_ON(irqs_disabled());
3686 spin_lock_irq(&n->list_lock);
3687 list_for_each_entry_safe(page, h, &n->partial, lru) {
3688 if (!page->inuse) {
3689 remove_partial(n, page);
3690 list_add(&page->lru, &discard);
3691 } else {
3692 list_slab_objects(s, page,
3693 "Objects remaining in %s on __kmem_cache_shutdown()");
3694 }
3695 }
3696 spin_unlock_irq(&n->list_lock);
3697
3698 list_for_each_entry_safe(page, h, &discard, lru)
3699 discard_slab(s, page);
3700}
3701
3702/*
3703 * Release all resources used by a slab cache.
3704 */
3705int __kmem_cache_shutdown(struct kmem_cache *s)
3706{
3707 int node;
3708 struct kmem_cache_node *n;
3709
3710 flush_all(s);
3711 /* Attempt to free all objects */
3712 for_each_kmem_cache_node(s, node, n) {
3713 free_partial(s, n);
3714 if (n->nr_partial || slabs_node(s, node))
3715 return 1;
3716 }
3717 return 0;
3718}
3719
3720/********************************************************************
3721 * Kmalloc subsystem
3722 *******************************************************************/
3723
3724static int __init setup_slub_min_order(char *str)
3725{
3726 get_option(&str, &slub_min_order);
3727
3728 return 1;
3729}
3730
3731__setup("slub_min_order=", setup_slub_min_order);
3732
3733static int __init setup_slub_max_order(char *str)
3734{
3735 get_option(&str, &slub_max_order);
3736 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3737
3738 return 1;
3739}
3740
3741__setup("slub_max_order=", setup_slub_max_order);
3742
3743static int __init setup_slub_min_objects(char *str)
3744{
3745 get_option(&str, &slub_min_objects);
3746
3747 return 1;
3748}
3749
3750__setup("slub_min_objects=", setup_slub_min_objects);
3751
3752#ifdef CONFIG_AMLOGIC_MEMORY_EXTEND
3753static void aml_slub_free_large(struct page *page, const void *obj)
3754{
3755 unsigned int nr_pages, i;
3756
3757 if (page) {
3758 __ClearPageHead(page);
3759 ClearPageOwnerPriv1(page);
3760 nr_pages = page->index;
3761 pr_debug("%s, page:%p, pages:%d, obj:%p\n",
3762 __func__, page_address(page), nr_pages, obj);
3763 for (i = 0; i < nr_pages; i++) {
3764 __free_pages(page, 0);
3765 page++;
3766 }
3767 }
3768}
3769#endif
3770
3771void *__kmalloc(size_t size, gfp_t flags)
3772{
3773 struct kmem_cache *s;
3774 void *ret;
3775
3776 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3777 return kmalloc_large(size, flags);
3778
3779 s = kmalloc_slab(size, flags);
3780
3781 if (unlikely(ZERO_OR_NULL_PTR(s)))
3782 return s;
3783
3784 ret = slab_alloc(s, flags, _RET_IP_);
3785
3786 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3787
3788 kasan_kmalloc(s, ret, size, flags);
3789
3790 return ret;
3791}
3792EXPORT_SYMBOL(__kmalloc);
3793
3794#ifdef CONFIG_NUMA
3795static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3796{
3797 struct page *page;
3798 void *ptr = NULL;
3799
3800 flags |= __GFP_COMP | __GFP_NOTRACK;
3801 page = alloc_pages_node(node, flags, get_order(size));
3802 if (page)
3803 ptr = page_address(page);
3804
3805 kmalloc_large_node_hook(ptr, size, flags);
3806 return ptr;
3807}
3808
3809void *__kmalloc_node(size_t size, gfp_t flags, int node)
3810{
3811 struct kmem_cache *s;
3812 void *ret;
3813
3814 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3815 ret = kmalloc_large_node(size, flags, node);
3816
3817 trace_kmalloc_node(_RET_IP_, ret,
3818 size, PAGE_SIZE << get_order(size),
3819 flags, node);
3820
3821 return ret;
3822 }
3823
3824 s = kmalloc_slab(size, flags);
3825
3826 if (unlikely(ZERO_OR_NULL_PTR(s)))
3827 return s;
3828
3829 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3830
3831 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3832
3833 kasan_kmalloc(s, ret, size, flags);
3834
3835 return ret;
3836}
3837EXPORT_SYMBOL(__kmalloc_node);
3838#endif
3839
3840#ifdef CONFIG_HARDENED_USERCOPY
3841/*
3842 * Rejects objects that are incorrectly sized.
3843 *
3844 * Returns NULL if check passes, otherwise const char * to name of cache
3845 * to indicate an error.
3846 */
3847const char *__check_heap_object(const void *ptr, unsigned long n,
3848 struct page *page)
3849{
3850 struct kmem_cache *s;
3851 unsigned long offset;
3852 size_t object_size;
3853
3854 /* Find object and usable object size. */
3855 s = page->slab_cache;
3856 object_size = slab_ksize(s);
3857
3858 /* Reject impossible pointers. */
3859 if (ptr < page_address(page))
3860 return s->name;
3861
3862 /* Find offset within object. */
3863 offset = (ptr - page_address(page)) % s->size;
3864
3865 /* Adjust for redzone and reject if within the redzone. */
3866 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE) {
3867 if (offset < s->red_left_pad)
3868 return s->name;
3869 offset -= s->red_left_pad;
3870 }
3871
3872 /* Allow address range falling entirely within object size. */
3873 if (offset <= object_size && n <= object_size - offset)
3874 return NULL;
3875
3876 return s->name;
3877}
3878#endif /* CONFIG_HARDENED_USERCOPY */
3879
3880static size_t __ksize(const void *object)
3881{
3882 struct page *page;
3883
3884 if (unlikely(object == ZERO_SIZE_PTR))
3885 return 0;
3886
3887 page = virt_to_head_page(object);
3888
3889 if (unlikely(!PageSlab(page))) {
3890 WARN_ON(!PageCompound(page));
3891 #ifdef CONFIG_AMLOGIC_MEMORY_EXTEND
3892 if (unlikely(PageOwnerPriv1(page))) {
3893 pr_debug("%s, obj:%p, page:%p, index:%ld, size:%ld\n",
3894 __func__, object, page_address(page),
3895 page->index, PAGE_SIZE * page->index);
3896 return PAGE_SIZE * page->index;
3897 } else
3898 return PAGE_SIZE << compound_order(page);
3899 #else
3900 return PAGE_SIZE << compound_order(page);
3901 #endif
3902 }
3903
3904 return slab_ksize(page->slab_cache);
3905}
3906
3907size_t ksize(const void *object)
3908{
3909 size_t size = __ksize(object);
3910 /* We assume that ksize callers could use whole allocated area,
3911 * so we need to unpoison this area.
3912 */
3913 kasan_unpoison_shadow(object, size);
3914 return size;
3915}
3916EXPORT_SYMBOL(ksize);
3917
3918void kfree(const void *x)
3919{
3920 struct page *page;
3921 void *object = (void *)x;
3922
3923 trace_kfree(_RET_IP_, x);
3924
3925 if (unlikely(ZERO_OR_NULL_PTR(x)))
3926 return;
3927
3928 page = virt_to_head_page(x);
3929 if (unlikely(!PageSlab(page))) {
3930 BUG_ON(!PageCompound(page));
3931 #ifdef CONFIG_AMLOGIC_MEMORY_EXTEND
3932 kmemleak_free(x);
3933 if (unlikely(PageOwnerPriv1(page)))
3934 aml_slub_free_large(page, x);
3935 else {
3936 __free_pages(page, compound_order(page));
3937 kasan_kfree_large(x);
3938 }
3939 return;
3940 #else
3941 kfree_hook(x);
3942 __free_pages(page, compound_order(page));
3943 return;
3944 #endif /* CONFIG_AMLOGIC_MEMORY_EXTEND */
3945 }
3946 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
3947}
3948EXPORT_SYMBOL(kfree);
3949
3950#define SHRINK_PROMOTE_MAX 32
3951
3952/*
3953 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3954 * up most to the head of the partial lists. New allocations will then
3955 * fill those up and thus they can be removed from the partial lists.
3956 *
3957 * The slabs with the least items are placed last. This results in them
3958 * being allocated from last increasing the chance that the last objects
3959 * are freed in them.
3960 */
3961int __kmem_cache_shrink(struct kmem_cache *s)
3962{
3963 int node;
3964 int i;
3965 struct kmem_cache_node *n;
3966 struct page *page;
3967 struct page *t;
3968 struct list_head discard;
3969 struct list_head promote[SHRINK_PROMOTE_MAX];
3970 unsigned long flags;
3971 int ret = 0;
3972
3973 flush_all(s);
3974 for_each_kmem_cache_node(s, node, n) {
3975 INIT_LIST_HEAD(&discard);
3976 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3977 INIT_LIST_HEAD(promote + i);
3978
3979 spin_lock_irqsave(&n->list_lock, flags);
3980
3981 /*
3982 * Build lists of slabs to discard or promote.
3983 *
3984 * Note that concurrent frees may occur while we hold the
3985 * list_lock. page->inuse here is the upper limit.
3986 */
3987 list_for_each_entry_safe(page, t, &n->partial, lru) {
3988 int free = page->objects - page->inuse;
3989
3990 /* Do not reread page->inuse */
3991 barrier();
3992
3993 /* We do not keep full slabs on the list */
3994 BUG_ON(free <= 0);
3995
3996 if (free == page->objects) {
3997 list_move(&page->lru, &discard);
3998 n->nr_partial--;
3999 } else if (free <= SHRINK_PROMOTE_MAX)
4000 list_move(&page->lru, promote + free - 1);
4001 }
4002
4003 /*
4004 * Promote the slabs filled up most to the head of the
4005 * partial list.
4006 */
4007 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
4008 list_splice(promote + i, &n->partial);
4009
4010 spin_unlock_irqrestore(&n->list_lock, flags);
4011
4012 /* Release empty slabs */
4013 list_for_each_entry_safe(page, t, &discard, lru)
4014 discard_slab(s, page);
4015
4016 if (slabs_node(s, node))
4017 ret = 1;
4018 }
4019
4020 return ret;
4021}
4022
4023static int slab_mem_going_offline_callback(void *arg)
4024{
4025 struct kmem_cache *s;
4026
4027 mutex_lock(&slab_mutex);
4028 list_for_each_entry(s, &slab_caches, list)
4029 __kmem_cache_shrink(s);
4030 mutex_unlock(&slab_mutex);
4031
4032 return 0;
4033}
4034
4035static void slab_mem_offline_callback(void *arg)
4036{
4037 struct kmem_cache_node *n;
4038 struct kmem_cache *s;
4039 struct memory_notify *marg = arg;
4040 int offline_node;
4041
4042 offline_node = marg->status_change_nid_normal;
4043
4044 /*
4045 * If the node still has available memory. we need kmem_cache_node
4046 * for it yet.
4047 */
4048 if (offline_node < 0)
4049 return;
4050
4051 mutex_lock(&slab_mutex);
4052 list_for_each_entry(s, &slab_caches, list) {
4053 n = get_node(s, offline_node);
4054 if (n) {
4055 /*
4056 * if n->nr_slabs > 0, slabs still exist on the node
4057 * that is going down. We were unable to free them,
4058 * and offline_pages() function shouldn't call this
4059 * callback. So, we must fail.
4060 */
4061 BUG_ON(slabs_node(s, offline_node));
4062
4063 s->node[offline_node] = NULL;
4064 kmem_cache_free(kmem_cache_node, n);
4065 }
4066 }
4067 mutex_unlock(&slab_mutex);
4068}
4069
4070static int slab_mem_going_online_callback(void *arg)
4071{
4072 struct kmem_cache_node *n;
4073 struct kmem_cache *s;
4074 struct memory_notify *marg = arg;
4075 int nid = marg->status_change_nid_normal;
4076 int ret = 0;
4077
4078 /*
4079 * If the node's memory is already available, then kmem_cache_node is
4080 * already created. Nothing to do.
4081 */
4082 if (nid < 0)
4083 return 0;
4084
4085 /*
4086 * We are bringing a node online. No memory is available yet. We must
4087 * allocate a kmem_cache_node structure in order to bring the node
4088 * online.
4089 */
4090 mutex_lock(&slab_mutex);
4091 list_for_each_entry(s, &slab_caches, list) {
4092 /*
4093 * XXX: kmem_cache_alloc_node will fallback to other nodes
4094 * since memory is not yet available from the node that
4095 * is brought up.
4096 */
4097 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4098 if (!n) {
4099 ret = -ENOMEM;
4100 goto out;
4101 }
4102 init_kmem_cache_node(n);
4103 s->node[nid] = n;
4104 }
4105out:
4106 mutex_unlock(&slab_mutex);
4107 return ret;
4108}
4109
4110static int slab_memory_callback(struct notifier_block *self,
4111 unsigned long action, void *arg)
4112{
4113 int ret = 0;
4114
4115 switch (action) {
4116 case MEM_GOING_ONLINE:
4117 ret = slab_mem_going_online_callback(arg);
4118 break;
4119 case MEM_GOING_OFFLINE:
4120 ret = slab_mem_going_offline_callback(arg);
4121 break;
4122 case MEM_OFFLINE:
4123 case MEM_CANCEL_ONLINE:
4124 slab_mem_offline_callback(arg);
4125 break;
4126 case MEM_ONLINE:
4127 case MEM_CANCEL_OFFLINE:
4128 break;
4129 }
4130 if (ret)
4131 ret = notifier_from_errno(ret);
4132 else
4133 ret = NOTIFY_OK;
4134 return ret;
4135}
4136
4137static struct notifier_block slab_memory_callback_nb = {
4138 .notifier_call = slab_memory_callback,
4139 .priority = SLAB_CALLBACK_PRI,
4140};
4141
4142/********************************************************************
4143 * Basic setup of slabs
4144 *******************************************************************/
4145
4146/*
4147 * Used for early kmem_cache structures that were allocated using
4148 * the page allocator. Allocate them properly then fix up the pointers
4149 * that may be pointing to the wrong kmem_cache structure.
4150 */
4151
4152static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4153{
4154 int node;
4155 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4156 struct kmem_cache_node *n;
4157
4158 memcpy(s, static_cache, kmem_cache->object_size);
4159
4160 /*
4161 * This runs very early, and only the boot processor is supposed to be
4162 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4163 * IPIs around.
4164 */
4165 __flush_cpu_slab(s, smp_processor_id());
4166 for_each_kmem_cache_node(s, node, n) {
4167 struct page *p;
4168
4169 list_for_each_entry(p, &n->partial, lru)
4170 p->slab_cache = s;
4171
4172#ifdef CONFIG_SLUB_DEBUG
4173 list_for_each_entry(p, &n->full, lru)
4174 p->slab_cache = s;
4175#endif
4176 }
4177 slab_init_memcg_params(s);
4178 list_add(&s->list, &slab_caches);
4179 return s;
4180}
4181
4182void __init kmem_cache_init(void)
4183{
4184 static __initdata struct kmem_cache boot_kmem_cache,
4185 boot_kmem_cache_node;
4186
4187 if (debug_guardpage_minorder())
4188 slub_max_order = 0;
4189
4190 kmem_cache_node = &boot_kmem_cache_node;
4191 kmem_cache = &boot_kmem_cache;
4192
4193 create_boot_cache(kmem_cache_node, "kmem_cache_node",
4194 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
4195
4196 register_hotmemory_notifier(&slab_memory_callback_nb);
4197
4198 /* Able to allocate the per node structures */
4199 slab_state = PARTIAL;
4200
4201 create_boot_cache(kmem_cache, "kmem_cache",
4202 offsetof(struct kmem_cache, node) +
4203 nr_node_ids * sizeof(struct kmem_cache_node *),
4204 SLAB_HWCACHE_ALIGN);
4205
4206 kmem_cache = bootstrap(&boot_kmem_cache);
4207
4208 /*
4209 * Allocate kmem_cache_node properly from the kmem_cache slab.
4210 * kmem_cache_node is separately allocated so no need to
4211 * update any list pointers.
4212 */
4213 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
4214
4215 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4216 setup_kmalloc_cache_index_table();
4217 create_kmalloc_caches(0);
4218#ifdef CONFIG_AMLOGIC_SLAB_TRACE
4219 slab_trace_init();
4220#endif
4221
4222 /* Setup random freelists for each cache */
4223 init_freelist_randomization();
4224
4225 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
4226 slub_cpu_dead);
4227
4228 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
4229 cache_line_size(),
4230 slub_min_order, slub_max_order, slub_min_objects,
4231 nr_cpu_ids, nr_node_ids);
4232}
4233
4234void __init kmem_cache_init_late(void)
4235{
4236}
4237
4238struct kmem_cache *
4239__kmem_cache_alias(const char *name, size_t size, size_t align,
4240 unsigned long flags, void (*ctor)(void *))
4241{
4242 struct kmem_cache *s, *c;
4243
4244 s = find_mergeable(size, align, flags, name, ctor);
4245 if (s) {
4246 s->refcount++;
4247
4248 /*
4249 * Adjust the object sizes so that we clear
4250 * the complete object on kzalloc.
4251 */
4252 s->object_size = max(s->object_size, (int)size);
4253 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
4254
4255 for_each_memcg_cache(c, s) {
4256 c->object_size = s->object_size;
4257 c->inuse = max_t(int, c->inuse,
4258 ALIGN(size, sizeof(void *)));
4259 }
4260
4261 if (sysfs_slab_alias(s, name)) {
4262 s->refcount--;
4263 s = NULL;
4264 }
4265 }
4266
4267 return s;
4268}
4269
4270int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
4271{
4272 int err;
4273
4274 err = kmem_cache_open(s, flags);
4275 if (err)
4276 return err;
4277
4278 /* Mutex is not taken during early boot */
4279 if (slab_state <= UP)
4280 return 0;
4281
4282 memcg_propagate_slab_attrs(s);
4283 err = sysfs_slab_add(s);
4284 if (err)
4285 __kmem_cache_release(s);
4286
4287 return err;
4288}
4289
4290void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4291{
4292 struct kmem_cache *s;
4293 void *ret;
4294
4295 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4296 return kmalloc_large(size, gfpflags);
4297
4298 s = kmalloc_slab(size, gfpflags);
4299
4300 if (unlikely(ZERO_OR_NULL_PTR(s)))
4301 return s;
4302
4303 ret = slab_alloc(s, gfpflags, caller);
4304
4305 /* Honor the call site pointer we received. */
4306 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4307
4308 return ret;
4309}
4310
4311#ifdef CONFIG_NUMA
4312void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4313 int node, unsigned long caller)
4314{
4315 struct kmem_cache *s;
4316 void *ret;
4317
4318 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4319 ret = kmalloc_large_node(size, gfpflags, node);
4320
4321 trace_kmalloc_node(caller, ret,
4322 size, PAGE_SIZE << get_order(size),
4323 gfpflags, node);
4324
4325 return ret;
4326 }
4327
4328 s = kmalloc_slab(size, gfpflags);
4329
4330 if (unlikely(ZERO_OR_NULL_PTR(s)))
4331 return s;
4332
4333 ret = slab_alloc_node(s, gfpflags, node, caller);
4334
4335 /* Honor the call site pointer we received. */
4336 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4337
4338 return ret;
4339}
4340#endif
4341
4342#ifdef CONFIG_SYSFS
4343static int count_inuse(struct page *page)
4344{
4345 return page->inuse;
4346}
4347
4348static int count_total(struct page *page)
4349{
4350 return page->objects;
4351}
4352#endif
4353
4354#ifdef CONFIG_SLUB_DEBUG
4355static int validate_slab(struct kmem_cache *s, struct page *page,
4356 unsigned long *map)
4357{
4358 void *p;
4359 void *addr = page_address(page);
4360
4361 if (!check_slab(s, page) ||
4362 !on_freelist(s, page, NULL))
4363 return 0;
4364
4365 /* Now we know that a valid freelist exists */
4366 bitmap_zero(map, page->objects);
4367
4368 get_map(s, page, map);
4369 for_each_object(p, s, addr, page->objects) {
4370 if (test_bit(slab_index(p, s, addr), map))
4371 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4372 return 0;
4373 }
4374
4375 for_each_object(p, s, addr, page->objects)
4376 if (!test_bit(slab_index(p, s, addr), map))
4377 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4378 return 0;
4379 return 1;
4380}
4381
4382static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4383 unsigned long *map)
4384{
4385 slab_lock(page);
4386 validate_slab(s, page, map);
4387 slab_unlock(page);
4388}
4389
4390static int validate_slab_node(struct kmem_cache *s,
4391 struct kmem_cache_node *n, unsigned long *map)
4392{
4393 unsigned long count = 0;
4394 struct page *page;
4395 unsigned long flags;
4396
4397 spin_lock_irqsave(&n->list_lock, flags);
4398
4399 list_for_each_entry(page, &n->partial, lru) {
4400 validate_slab_slab(s, page, map);
4401 count++;
4402 }
4403 if (count != n->nr_partial)
4404 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4405 s->name, count, n->nr_partial);
4406
4407 if (!(s->flags & SLAB_STORE_USER))
4408 goto out;
4409
4410 list_for_each_entry(page, &n->full, lru) {
4411 validate_slab_slab(s, page, map);
4412 count++;
4413 }
4414 if (count != atomic_long_read(&n->nr_slabs))
4415 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4416 s->name, count, atomic_long_read(&n->nr_slabs));
4417
4418out:
4419 spin_unlock_irqrestore(&n->list_lock, flags);
4420 return count;
4421}
4422
4423static long validate_slab_cache(struct kmem_cache *s)
4424{
4425 int node;
4426 unsigned long count = 0;
4427 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4428 sizeof(unsigned long), GFP_KERNEL);
4429 struct kmem_cache_node *n;
4430
4431 if (!map)
4432 return -ENOMEM;
4433
4434 flush_all(s);
4435 for_each_kmem_cache_node(s, node, n)
4436 count += validate_slab_node(s, n, map);
4437 kfree(map);
4438 return count;
4439}
4440/*
4441 * Generate lists of code addresses where slabcache objects are allocated
4442 * and freed.
4443 */
4444
4445struct location {
4446 unsigned long count;
4447 unsigned long addr;
4448 long long sum_time;
4449 long min_time;
4450 long max_time;
4451 long min_pid;
4452 long max_pid;
4453 DECLARE_BITMAP(cpus, NR_CPUS);
4454 nodemask_t nodes;
4455};
4456
4457struct loc_track {
4458 unsigned long max;
4459 unsigned long count;
4460 struct location *loc;
4461};
4462
4463static void free_loc_track(struct loc_track *t)
4464{
4465 if (t->max)
4466 free_pages((unsigned long)t->loc,
4467 get_order(sizeof(struct location) * t->max));
4468}
4469
4470static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4471{
4472 struct location *l;
4473 int order;
4474
4475 order = get_order(sizeof(struct location) * max);
4476
4477 l = (void *)__get_free_pages(flags, order);
4478 if (!l)
4479 return 0;
4480
4481 if (t->count) {
4482 memcpy(l, t->loc, sizeof(struct location) * t->count);
4483 free_loc_track(t);
4484 }
4485 t->max = max;
4486 t->loc = l;
4487 return 1;
4488}
4489
4490static int add_location(struct loc_track *t, struct kmem_cache *s,
4491 const struct track *track)
4492{
4493 long start, end, pos;
4494 struct location *l;
4495 unsigned long caddr;
4496 unsigned long age = jiffies - track->when;
4497
4498 start = -1;
4499 end = t->count;
4500
4501 for ( ; ; ) {
4502 pos = start + (end - start + 1) / 2;
4503
4504 /*
4505 * There is nothing at "end". If we end up there
4506 * we need to add something to before end.
4507 */
4508 if (pos == end)
4509 break;
4510
4511 caddr = t->loc[pos].addr;
4512 if (track->addr == caddr) {
4513
4514 l = &t->loc[pos];
4515 l->count++;
4516 if (track->when) {
4517 l->sum_time += age;
4518 if (age < l->min_time)
4519 l->min_time = age;
4520 if (age > l->max_time)
4521 l->max_time = age;
4522
4523 if (track->pid < l->min_pid)
4524 l->min_pid = track->pid;
4525 if (track->pid > l->max_pid)
4526 l->max_pid = track->pid;
4527
4528 cpumask_set_cpu(track->cpu,
4529 to_cpumask(l->cpus));
4530 }
4531 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4532 return 1;
4533 }
4534
4535 if (track->addr < caddr)
4536 end = pos;
4537 else
4538 start = pos;
4539 }
4540
4541 /*
4542 * Not found. Insert new tracking element.
4543 */
4544 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4545 return 0;
4546
4547 l = t->loc + pos;
4548 if (pos < t->count)
4549 memmove(l + 1, l,
4550 (t->count - pos) * sizeof(struct location));
4551 t->count++;
4552 l->count = 1;
4553 l->addr = track->addr;
4554 l->sum_time = age;
4555 l->min_time = age;
4556 l->max_time = age;
4557 l->min_pid = track->pid;
4558 l->max_pid = track->pid;
4559 cpumask_clear(to_cpumask(l->cpus));
4560 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4561 nodes_clear(l->nodes);
4562 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4563 return 1;
4564}
4565
4566static void process_slab(struct loc_track *t, struct kmem_cache *s,
4567 struct page *page, enum track_item alloc,
4568 unsigned long *map)
4569{
4570 void *addr = page_address(page);
4571 void *p;
4572
4573 bitmap_zero(map, page->objects);
4574 get_map(s, page, map);
4575
4576 for_each_object(p, s, addr, page->objects)
4577 if (!test_bit(slab_index(p, s, addr), map))
4578 add_location(t, s, get_track(s, p, alloc));
4579}
4580
4581static int list_locations(struct kmem_cache *s, char *buf,
4582 enum track_item alloc)
4583{
4584 int len = 0;
4585 unsigned long i;
4586 struct loc_track t = { 0, 0, NULL };
4587 int node;
4588 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4589 sizeof(unsigned long), GFP_KERNEL);
4590 struct kmem_cache_node *n;
4591
4592 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4593 GFP_TEMPORARY)) {
4594 kfree(map);
4595 return sprintf(buf, "Out of memory\n");
4596 }
4597 /* Push back cpu slabs */
4598 flush_all(s);
4599
4600 for_each_kmem_cache_node(s, node, n) {
4601 unsigned long flags;
4602 struct page *page;
4603
4604 if (!atomic_long_read(&n->nr_slabs))
4605 continue;
4606
4607 spin_lock_irqsave(&n->list_lock, flags);
4608 list_for_each_entry(page, &n->partial, lru)
4609 process_slab(&t, s, page, alloc, map);
4610 list_for_each_entry(page, &n->full, lru)
4611 process_slab(&t, s, page, alloc, map);
4612 spin_unlock_irqrestore(&n->list_lock, flags);
4613 }
4614
4615 for (i = 0; i < t.count; i++) {
4616 struct location *l = &t.loc[i];
4617
4618 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4619 break;
4620 len += sprintf(buf + len, "%7ld ", l->count);
4621
4622 if (l->addr)
4623 len += sprintf(buf + len, "%pS", (void *)l->addr);
4624 else
4625 len += sprintf(buf + len, "<not-available>");
4626
4627 if (l->sum_time != l->min_time) {
4628 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4629 l->min_time,
4630 (long)div_u64(l->sum_time, l->count),
4631 l->max_time);
4632 } else
4633 len += sprintf(buf + len, " age=%ld",
4634 l->min_time);
4635
4636 if (l->min_pid != l->max_pid)
4637 len += sprintf(buf + len, " pid=%ld-%ld",
4638 l->min_pid, l->max_pid);
4639 else
4640 len += sprintf(buf + len, " pid=%ld",
4641 l->min_pid);
4642
4643 if (num_online_cpus() > 1 &&
4644 !cpumask_empty(to_cpumask(l->cpus)) &&
4645 len < PAGE_SIZE - 60)
4646 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4647 " cpus=%*pbl",
4648 cpumask_pr_args(to_cpumask(l->cpus)));
4649
4650 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4651 len < PAGE_SIZE - 60)
4652 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4653 " nodes=%*pbl",
4654 nodemask_pr_args(&l->nodes));
4655
4656 len += sprintf(buf + len, "\n");
4657 }
4658
4659 free_loc_track(&t);
4660 kfree(map);
4661 if (!t.count)
4662 len += sprintf(buf, "No data\n");
4663 return len;
4664}
4665#endif
4666
4667#ifdef SLUB_RESILIENCY_TEST
4668static void __init resiliency_test(void)
4669{
4670 u8 *p;
4671
4672 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4673
4674 pr_err("SLUB resiliency testing\n");
4675 pr_err("-----------------------\n");
4676 pr_err("A. Corruption after allocation\n");
4677
4678 p = kzalloc(16, GFP_KERNEL);
4679 p[16] = 0x12;
4680 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4681 p + 16);
4682
4683 validate_slab_cache(kmalloc_caches[4]);
4684
4685 /* Hmmm... The next two are dangerous */
4686 p = kzalloc(32, GFP_KERNEL);
4687 p[32 + sizeof(void *)] = 0x34;
4688 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4689 p);
4690 pr_err("If allocated object is overwritten then not detectable\n\n");
4691
4692 validate_slab_cache(kmalloc_caches[5]);
4693 p = kzalloc(64, GFP_KERNEL);
4694 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4695 *p = 0x56;
4696 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4697 p);
4698 pr_err("If allocated object is overwritten then not detectable\n\n");
4699 validate_slab_cache(kmalloc_caches[6]);
4700
4701 pr_err("\nB. Corruption after free\n");
4702 p = kzalloc(128, GFP_KERNEL);
4703 kfree(p);
4704 *p = 0x78;
4705 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4706 validate_slab_cache(kmalloc_caches[7]);
4707
4708 p = kzalloc(256, GFP_KERNEL);
4709 kfree(p);
4710 p[50] = 0x9a;
4711 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4712 validate_slab_cache(kmalloc_caches[8]);
4713
4714 p = kzalloc(512, GFP_KERNEL);
4715 kfree(p);
4716 p[512] = 0xab;
4717 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4718 validate_slab_cache(kmalloc_caches[9]);
4719}
4720#else
4721#ifdef CONFIG_SYSFS
4722static void resiliency_test(void) {};
4723#endif
4724#endif
4725
4726#ifdef CONFIG_SYSFS
4727enum slab_stat_type {
4728 SL_ALL, /* All slabs */
4729 SL_PARTIAL, /* Only partially allocated slabs */
4730 SL_CPU, /* Only slabs used for cpu caches */
4731 SL_OBJECTS, /* Determine allocated objects not slabs */
4732 SL_TOTAL /* Determine object capacity not slabs */
4733};
4734
4735#define SO_ALL (1 << SL_ALL)
4736#define SO_PARTIAL (1 << SL_PARTIAL)
4737#define SO_CPU (1 << SL_CPU)
4738#define SO_OBJECTS (1 << SL_OBJECTS)
4739#define SO_TOTAL (1 << SL_TOTAL)
4740
4741#ifdef CONFIG_MEMCG
4742static bool memcg_sysfs_enabled = IS_ENABLED(CONFIG_SLUB_MEMCG_SYSFS_ON);
4743
4744static int __init setup_slub_memcg_sysfs(char *str)
4745{
4746 int v;
4747
4748 if (get_option(&str, &v) > 0)
4749 memcg_sysfs_enabled = v;
4750
4751 return 1;
4752}
4753
4754__setup("slub_memcg_sysfs=", setup_slub_memcg_sysfs);
4755#endif
4756
4757static ssize_t show_slab_objects(struct kmem_cache *s,
4758 char *buf, unsigned long flags)
4759{
4760 unsigned long total = 0;
4761 int node;
4762 int x;
4763 unsigned long *nodes;
4764
4765 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4766 if (!nodes)
4767 return -ENOMEM;
4768
4769 if (flags & SO_CPU) {
4770 int cpu;
4771
4772 for_each_possible_cpu(cpu) {
4773 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4774 cpu);
4775 int node;
4776 struct page *page;
4777
4778 page = READ_ONCE(c->page);
4779 if (!page)
4780 continue;
4781
4782 node = page_to_nid(page);
4783 if (flags & SO_TOTAL)
4784 x = page->objects;
4785 else if (flags & SO_OBJECTS)
4786 x = page->inuse;
4787 else
4788 x = 1;
4789
4790 total += x;
4791 nodes[node] += x;
4792
4793 page = READ_ONCE(c->partial);
4794 if (page) {
4795 node = page_to_nid(page);
4796 if (flags & SO_TOTAL)
4797 WARN_ON_ONCE(1);
4798 else if (flags & SO_OBJECTS)
4799 WARN_ON_ONCE(1);
4800 else
4801 x = page->pages;
4802 total += x;
4803 nodes[node] += x;
4804 }
4805 }
4806 }
4807
4808 get_online_mems();
4809#ifdef CONFIG_SLUB_DEBUG
4810 if (flags & SO_ALL) {
4811 struct kmem_cache_node *n;
4812
4813 for_each_kmem_cache_node(s, node, n) {
4814
4815 if (flags & SO_TOTAL)
4816 x = atomic_long_read(&n->total_objects);
4817 else if (flags & SO_OBJECTS)
4818 x = atomic_long_read(&n->total_objects) -
4819 count_partial(n, count_free);
4820 else
4821 x = atomic_long_read(&n->nr_slabs);
4822 total += x;
4823 nodes[node] += x;
4824 }
4825
4826 } else
4827#endif
4828 if (flags & SO_PARTIAL) {
4829 struct kmem_cache_node *n;
4830
4831 for_each_kmem_cache_node(s, node, n) {
4832 if (flags & SO_TOTAL)
4833 x = count_partial(n, count_total);
4834 else if (flags & SO_OBJECTS)
4835 x = count_partial(n, count_inuse);
4836 else
4837 x = n->nr_partial;
4838 total += x;
4839 nodes[node] += x;
4840 }
4841 }
4842 x = sprintf(buf, "%lu", total);
4843#ifdef CONFIG_NUMA
4844 for (node = 0; node < nr_node_ids; node++)
4845 if (nodes[node])
4846 x += sprintf(buf + x, " N%d=%lu",
4847 node, nodes[node]);
4848#endif
4849 put_online_mems();
4850 kfree(nodes);
4851 return x + sprintf(buf + x, "\n");
4852}
4853
4854#ifdef CONFIG_SLUB_DEBUG
4855static int any_slab_objects(struct kmem_cache *s)
4856{
4857 int node;
4858 struct kmem_cache_node *n;
4859
4860 for_each_kmem_cache_node(s, node, n)
4861 if (atomic_long_read(&n->total_objects))
4862 return 1;
4863
4864 return 0;
4865}
4866#endif
4867
4868#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4869#define to_slab(n) container_of(n, struct kmem_cache, kobj)
4870
4871struct slab_attribute {
4872 struct attribute attr;
4873 ssize_t (*show)(struct kmem_cache *s, char *buf);
4874 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4875};
4876
4877#define SLAB_ATTR_RO(_name) \
4878 static struct slab_attribute _name##_attr = \
4879 __ATTR(_name, 0400, _name##_show, NULL)
4880
4881#define SLAB_ATTR(_name) \
4882 static struct slab_attribute _name##_attr = \
4883 __ATTR(_name, 0600, _name##_show, _name##_store)
4884
4885static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4886{
4887 return sprintf(buf, "%d\n", s->size);
4888}
4889SLAB_ATTR_RO(slab_size);
4890
4891static ssize_t align_show(struct kmem_cache *s, char *buf)
4892{
4893 return sprintf(buf, "%d\n", s->align);
4894}
4895SLAB_ATTR_RO(align);
4896
4897static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4898{
4899 return sprintf(buf, "%d\n", s->object_size);
4900}
4901SLAB_ATTR_RO(object_size);
4902
4903static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4904{
4905 return sprintf(buf, "%d\n", oo_objects(s->oo));
4906}
4907SLAB_ATTR_RO(objs_per_slab);
4908
4909static ssize_t order_store(struct kmem_cache *s,
4910 const char *buf, size_t length)
4911{
4912 unsigned long order;
4913 int err;
4914
4915 err = kstrtoul(buf, 10, &order);
4916 if (err)
4917 return err;
4918
4919 if (order > slub_max_order || order < slub_min_order)
4920 return -EINVAL;
4921
4922 calculate_sizes(s, order);
4923 return length;
4924}
4925
4926static ssize_t order_show(struct kmem_cache *s, char *buf)
4927{
4928 return sprintf(buf, "%d\n", oo_order(s->oo));
4929}
4930SLAB_ATTR(order);
4931
4932static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4933{
4934 return sprintf(buf, "%lu\n", s->min_partial);
4935}
4936
4937static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4938 size_t length)
4939{
4940 unsigned long min;
4941 int err;
4942
4943 err = kstrtoul(buf, 10, &min);
4944 if (err)
4945 return err;
4946
4947 set_min_partial(s, min);
4948 return length;
4949}
4950SLAB_ATTR(min_partial);
4951
4952static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4953{
4954 return sprintf(buf, "%u\n", s->cpu_partial);
4955}
4956
4957static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4958 size_t length)
4959{
4960 unsigned int objects;
4961 int err;
4962
4963 err = kstrtouint(buf, 10, &objects);
4964 if (err)
4965 return err;
4966 if (objects && !kmem_cache_has_cpu_partial(s))
4967 return -EINVAL;
4968
4969 s->cpu_partial = objects;
4970 flush_all(s);
4971 return length;
4972}
4973SLAB_ATTR(cpu_partial);
4974
4975static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4976{
4977 if (!s->ctor)
4978 return 0;
4979 return sprintf(buf, "%pS\n", s->ctor);
4980}
4981SLAB_ATTR_RO(ctor);
4982
4983static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4984{
4985 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
4986}
4987SLAB_ATTR_RO(aliases);
4988
4989static ssize_t partial_show(struct kmem_cache *s, char *buf)
4990{
4991 return show_slab_objects(s, buf, SO_PARTIAL);
4992}
4993SLAB_ATTR_RO(partial);
4994
4995static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4996{
4997 return show_slab_objects(s, buf, SO_CPU);
4998}
4999SLAB_ATTR_RO(cpu_slabs);
5000
5001static ssize_t objects_show(struct kmem_cache *s, char *buf)
5002{
5003 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5004}
5005SLAB_ATTR_RO(objects);
5006
5007static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5008{
5009 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5010}
5011SLAB_ATTR_RO(objects_partial);
5012
5013static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5014{
5015 int objects = 0;
5016 int pages = 0;
5017 int cpu;
5018 int len;
5019
5020 for_each_online_cpu(cpu) {
5021 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
5022
5023 if (page) {
5024 pages += page->pages;
5025 objects += page->pobjects;
5026 }
5027 }
5028
5029 len = sprintf(buf, "%d(%d)", objects, pages);
5030
5031#ifdef CONFIG_SMP
5032 for_each_online_cpu(cpu) {
5033 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
5034
5035 if (page && len < PAGE_SIZE - 20)
5036 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
5037 page->pobjects, page->pages);
5038 }
5039#endif
5040 return len + sprintf(buf + len, "\n");
5041}
5042SLAB_ATTR_RO(slabs_cpu_partial);
5043
5044static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5045{
5046 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5047}
5048
5049static ssize_t reclaim_account_store(struct kmem_cache *s,
5050 const char *buf, size_t length)
5051{
5052 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
5053 if (buf[0] == '1')
5054 s->flags |= SLAB_RECLAIM_ACCOUNT;
5055 return length;
5056}
5057SLAB_ATTR(reclaim_account);
5058
5059static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5060{
5061 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5062}
5063SLAB_ATTR_RO(hwcache_align);
5064
5065#ifdef CONFIG_ZONE_DMA
5066static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5067{
5068 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5069}
5070SLAB_ATTR_RO(cache_dma);
5071#endif
5072
5073static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5074{
5075 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
5076}
5077SLAB_ATTR_RO(destroy_by_rcu);
5078
5079static ssize_t reserved_show(struct kmem_cache *s, char *buf)
5080{
5081 return sprintf(buf, "%d\n", s->reserved);
5082}
5083SLAB_ATTR_RO(reserved);
5084
5085#ifdef CONFIG_SLUB_DEBUG
5086static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5087{
5088 return show_slab_objects(s, buf, SO_ALL);
5089}
5090SLAB_ATTR_RO(slabs);
5091
5092static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5093{
5094 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5095}
5096SLAB_ATTR_RO(total_objects);
5097
5098static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5099{
5100 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5101}
5102
5103static ssize_t sanity_checks_store(struct kmem_cache *s,
5104 const char *buf, size_t length)
5105{
5106 s->flags &= ~SLAB_CONSISTENCY_CHECKS;
5107 if (buf[0] == '1') {
5108 s->flags &= ~__CMPXCHG_DOUBLE;
5109 s->flags |= SLAB_CONSISTENCY_CHECKS;
5110 }
5111 return length;
5112}
5113SLAB_ATTR(sanity_checks);
5114
5115static ssize_t trace_show(struct kmem_cache *s, char *buf)
5116{
5117 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5118}
5119
5120static ssize_t trace_store(struct kmem_cache *s, const char *buf,
5121 size_t length)
5122{
5123 /*
5124 * Tracing a merged cache is going to give confusing results
5125 * as well as cause other issues like converting a mergeable
5126 * cache into an umergeable one.
5127 */
5128 if (s->refcount > 1)
5129 return -EINVAL;
5130
5131 s->flags &= ~SLAB_TRACE;
5132 if (buf[0] == '1') {
5133 s->flags &= ~__CMPXCHG_DOUBLE;
5134 s->flags |= SLAB_TRACE;
5135 }
5136 return length;
5137}
5138SLAB_ATTR(trace);
5139
5140static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5141{
5142 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5143}
5144
5145static ssize_t red_zone_store(struct kmem_cache *s,
5146 const char *buf, size_t length)
5147{
5148 if (any_slab_objects(s))
5149 return -EBUSY;
5150
5151 s->flags &= ~SLAB_RED_ZONE;
5152 if (buf[0] == '1') {
5153 s->flags |= SLAB_RED_ZONE;
5154 }
5155 calculate_sizes(s, -1);
5156 return length;
5157}
5158SLAB_ATTR(red_zone);
5159
5160static ssize_t poison_show(struct kmem_cache *s, char *buf)
5161{
5162 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
5163}
5164
5165static ssize_t poison_store(struct kmem_cache *s,
5166 const char *buf, size_t length)
5167{
5168 if (any_slab_objects(s))
5169 return -EBUSY;
5170
5171 s->flags &= ~SLAB_POISON;
5172 if (buf[0] == '1') {
5173 s->flags |= SLAB_POISON;
5174 }
5175 calculate_sizes(s, -1);
5176 return length;
5177}
5178SLAB_ATTR(poison);
5179
5180static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5181{
5182 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5183}
5184
5185static ssize_t store_user_store(struct kmem_cache *s,
5186 const char *buf, size_t length)
5187{
5188 if (any_slab_objects(s))
5189 return -EBUSY;
5190
5191 s->flags &= ~SLAB_STORE_USER;
5192 if (buf[0] == '1') {
5193 s->flags &= ~__CMPXCHG_DOUBLE;
5194 s->flags |= SLAB_STORE_USER;
5195 }
5196 calculate_sizes(s, -1);
5197 return length;
5198}
5199SLAB_ATTR(store_user);
5200
5201static ssize_t validate_show(struct kmem_cache *s, char *buf)
5202{
5203 return 0;
5204}
5205
5206static ssize_t validate_store(struct kmem_cache *s,
5207 const char *buf, size_t length)
5208{
5209 int ret = -EINVAL;
5210
5211 if (buf[0] == '1') {
5212 ret = validate_slab_cache(s);
5213 if (ret >= 0)
5214 ret = length;
5215 }
5216 return ret;
5217}
5218SLAB_ATTR(validate);
5219
5220static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
5221{
5222 if (!(s->flags & SLAB_STORE_USER))
5223 return -ENOSYS;
5224 return list_locations(s, buf, TRACK_ALLOC);
5225}
5226SLAB_ATTR_RO(alloc_calls);
5227
5228static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
5229{
5230 if (!(s->flags & SLAB_STORE_USER))
5231 return -ENOSYS;
5232 return list_locations(s, buf, TRACK_FREE);
5233}
5234SLAB_ATTR_RO(free_calls);
5235#endif /* CONFIG_SLUB_DEBUG */
5236
5237#ifdef CONFIG_FAILSLAB
5238static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5239{
5240 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5241}
5242
5243static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
5244 size_t length)
5245{
5246 if (s->refcount > 1)
5247 return -EINVAL;
5248
5249 s->flags &= ~SLAB_FAILSLAB;
5250 if (buf[0] == '1')
5251 s->flags |= SLAB_FAILSLAB;
5252 return length;
5253}
5254SLAB_ATTR(failslab);
5255#endif
5256
5257static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5258{
5259 return 0;
5260}
5261
5262static ssize_t shrink_store(struct kmem_cache *s,
5263 const char *buf, size_t length)
5264{
5265 if (buf[0] == '1')
5266 kmem_cache_shrink(s);
5267 else
5268 return -EINVAL;
5269 return length;
5270}
5271SLAB_ATTR(shrink);
5272
5273#ifdef CONFIG_NUMA
5274static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5275{
5276 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
5277}
5278
5279static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5280 const char *buf, size_t length)
5281{
5282 unsigned long ratio;
5283 int err;
5284
5285 err = kstrtoul(buf, 10, &ratio);
5286 if (err)
5287 return err;
5288
5289 if (ratio <= 100)
5290 s->remote_node_defrag_ratio = ratio * 10;
5291
5292 return length;
5293}
5294SLAB_ATTR(remote_node_defrag_ratio);
5295#endif
5296
5297#ifdef CONFIG_SLUB_STATS
5298static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5299{
5300 unsigned long sum = 0;
5301 int cpu;
5302 int len;
5303 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
5304
5305 if (!data)
5306 return -ENOMEM;
5307
5308 for_each_online_cpu(cpu) {
5309 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5310
5311 data[cpu] = x;
5312 sum += x;
5313 }
5314
5315 len = sprintf(buf, "%lu", sum);
5316
5317#ifdef CONFIG_SMP
5318 for_each_online_cpu(cpu) {
5319 if (data[cpu] && len < PAGE_SIZE - 20)
5320 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5321 }
5322#endif
5323 kfree(data);
5324 return len + sprintf(buf + len, "\n");
5325}
5326
5327static void clear_stat(struct kmem_cache *s, enum stat_item si)
5328{
5329 int cpu;
5330
5331 for_each_online_cpu(cpu)
5332 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5333}
5334
5335#define STAT_ATTR(si, text) \
5336static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5337{ \
5338 return show_stat(s, buf, si); \
5339} \
5340static ssize_t text##_store(struct kmem_cache *s, \
5341 const char *buf, size_t length) \
5342{ \
5343 if (buf[0] != '0') \
5344 return -EINVAL; \
5345 clear_stat(s, si); \
5346 return length; \
5347} \
5348SLAB_ATTR(text); \
5349
5350STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5351STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5352STAT_ATTR(FREE_FASTPATH, free_fastpath);
5353STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5354STAT_ATTR(FREE_FROZEN, free_frozen);
5355STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5356STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5357STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5358STAT_ATTR(ALLOC_SLAB, alloc_slab);
5359STAT_ATTR(ALLOC_REFILL, alloc_refill);
5360STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5361STAT_ATTR(FREE_SLAB, free_slab);
5362STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5363STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5364STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5365STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5366STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5367STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5368STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5369STAT_ATTR(ORDER_FALLBACK, order_fallback);
5370STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5371STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5372STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5373STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5374STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5375STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5376#endif
5377
5378static struct attribute *slab_attrs[] = {
5379 &slab_size_attr.attr,
5380 &object_size_attr.attr,
5381 &objs_per_slab_attr.attr,
5382 &order_attr.attr,
5383 &min_partial_attr.attr,
5384 &cpu_partial_attr.attr,
5385 &objects_attr.attr,
5386 &objects_partial_attr.attr,
5387 &partial_attr.attr,
5388 &cpu_slabs_attr.attr,
5389 &ctor_attr.attr,
5390 &aliases_attr.attr,
5391 &align_attr.attr,
5392 &hwcache_align_attr.attr,
5393 &reclaim_account_attr.attr,
5394 &destroy_by_rcu_attr.attr,
5395 &shrink_attr.attr,
5396 &reserved_attr.attr,
5397 &slabs_cpu_partial_attr.attr,
5398#ifdef CONFIG_SLUB_DEBUG
5399 &total_objects_attr.attr,
5400 &slabs_attr.attr,
5401 &sanity_checks_attr.attr,
5402 &trace_attr.attr,
5403 &red_zone_attr.attr,
5404 &poison_attr.attr,
5405 &store_user_attr.attr,
5406 &validate_attr.attr,
5407 &alloc_calls_attr.attr,
5408 &free_calls_attr.attr,
5409#endif
5410#ifdef CONFIG_ZONE_DMA
5411 &cache_dma_attr.attr,
5412#endif
5413#ifdef CONFIG_NUMA
5414 &remote_node_defrag_ratio_attr.attr,
5415#endif
5416#ifdef CONFIG_SLUB_STATS
5417 &alloc_fastpath_attr.attr,
5418 &alloc_slowpath_attr.attr,
5419 &free_fastpath_attr.attr,
5420 &free_slowpath_attr.attr,
5421 &free_frozen_attr.attr,
5422 &free_add_partial_attr.attr,
5423 &free_remove_partial_attr.attr,
5424 &alloc_from_partial_attr.attr,
5425 &alloc_slab_attr.attr,
5426 &alloc_refill_attr.attr,
5427 &alloc_node_mismatch_attr.attr,
5428 &free_slab_attr.attr,
5429 &cpuslab_flush_attr.attr,
5430 &deactivate_full_attr.attr,
5431 &deactivate_empty_attr.attr,
5432 &deactivate_to_head_attr.attr,
5433 &deactivate_to_tail_attr.attr,
5434 &deactivate_remote_frees_attr.attr,
5435 &deactivate_bypass_attr.attr,
5436 &order_fallback_attr.attr,
5437 &cmpxchg_double_fail_attr.attr,
5438 &cmpxchg_double_cpu_fail_attr.attr,
5439 &cpu_partial_alloc_attr.attr,
5440 &cpu_partial_free_attr.attr,
5441 &cpu_partial_node_attr.attr,
5442 &cpu_partial_drain_attr.attr,
5443#endif
5444#ifdef CONFIG_FAILSLAB
5445 &failslab_attr.attr,
5446#endif
5447
5448 NULL
5449};
5450
5451static struct attribute_group slab_attr_group = {
5452 .attrs = slab_attrs,
5453};
5454
5455static ssize_t slab_attr_show(struct kobject *kobj,
5456 struct attribute *attr,
5457 char *buf)
5458{
5459 struct slab_attribute *attribute;
5460 struct kmem_cache *s;
5461 int err;
5462
5463 attribute = to_slab_attr(attr);
5464 s = to_slab(kobj);
5465
5466 if (!attribute->show)
5467 return -EIO;
5468
5469 err = attribute->show(s, buf);
5470
5471 return err;
5472}
5473
5474static ssize_t slab_attr_store(struct kobject *kobj,
5475 struct attribute *attr,
5476 const char *buf, size_t len)
5477{
5478 struct slab_attribute *attribute;
5479 struct kmem_cache *s;
5480 int err;
5481
5482 attribute = to_slab_attr(attr);
5483 s = to_slab(kobj);
5484
5485 if (!attribute->store)
5486 return -EIO;
5487
5488 err = attribute->store(s, buf, len);
5489#ifdef CONFIG_MEMCG
5490 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5491 struct kmem_cache *c;
5492
5493 mutex_lock(&slab_mutex);
5494 if (s->max_attr_size < len)
5495 s->max_attr_size = len;
5496
5497 /*
5498 * This is a best effort propagation, so this function's return
5499 * value will be determined by the parent cache only. This is
5500 * basically because not all attributes will have a well
5501 * defined semantics for rollbacks - most of the actions will
5502 * have permanent effects.
5503 *
5504 * Returning the error value of any of the children that fail
5505 * is not 100 % defined, in the sense that users seeing the
5506 * error code won't be able to know anything about the state of
5507 * the cache.
5508 *
5509 * Only returning the error code for the parent cache at least
5510 * has well defined semantics. The cache being written to
5511 * directly either failed or succeeded, in which case we loop
5512 * through the descendants with best-effort propagation.
5513 */
5514 for_each_memcg_cache(c, s)
5515 attribute->store(c, buf, len);
5516 mutex_unlock(&slab_mutex);
5517 }
5518#endif
5519 return err;
5520}
5521
5522static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5523{
5524#ifdef CONFIG_MEMCG
5525 int i;
5526 char *buffer = NULL;
5527 struct kmem_cache *root_cache;
5528
5529 if (is_root_cache(s))
5530 return;
5531
5532 root_cache = s->memcg_params.root_cache;
5533
5534 /*
5535 * This mean this cache had no attribute written. Therefore, no point
5536 * in copying default values around
5537 */
5538 if (!root_cache->max_attr_size)
5539 return;
5540
5541 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5542 char mbuf[64];
5543 char *buf;
5544 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5545 ssize_t len;
5546
5547 if (!attr || !attr->store || !attr->show)
5548 continue;
5549
5550 /*
5551 * It is really bad that we have to allocate here, so we will
5552 * do it only as a fallback. If we actually allocate, though,
5553 * we can just use the allocated buffer until the end.
5554 *
5555 * Most of the slub attributes will tend to be very small in
5556 * size, but sysfs allows buffers up to a page, so they can
5557 * theoretically happen.
5558 */
5559 if (buffer)
5560 buf = buffer;
5561 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5562 buf = mbuf;
5563 else {
5564 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5565 if (WARN_ON(!buffer))
5566 continue;
5567 buf = buffer;
5568 }
5569
5570 len = attr->show(root_cache, buf);
5571 if (len > 0)
5572 attr->store(s, buf, len);
5573 }
5574
5575 if (buffer)
5576 free_page((unsigned long)buffer);
5577#endif
5578}
5579
5580static void kmem_cache_release(struct kobject *k)
5581{
5582 slab_kmem_cache_release(to_slab(k));
5583}
5584
5585static const struct sysfs_ops slab_sysfs_ops = {
5586 .show = slab_attr_show,
5587 .store = slab_attr_store,
5588};
5589
5590static struct kobj_type slab_ktype = {
5591 .sysfs_ops = &slab_sysfs_ops,
5592 .release = kmem_cache_release,
5593};
5594
5595static int uevent_filter(struct kset *kset, struct kobject *kobj)
5596{
5597 struct kobj_type *ktype = get_ktype(kobj);
5598
5599 if (ktype == &slab_ktype)
5600 return 1;
5601 return 0;
5602}
5603
5604static const struct kset_uevent_ops slab_uevent_ops = {
5605 .filter = uevent_filter,
5606};
5607
5608static struct kset *slab_kset;
5609
5610static inline struct kset *cache_kset(struct kmem_cache *s)
5611{
5612#ifdef CONFIG_MEMCG
5613 if (!is_root_cache(s))
5614 return s->memcg_params.root_cache->memcg_kset;
5615#endif
5616 return slab_kset;
5617}
5618
5619#define ID_STR_LENGTH 64
5620
5621/* Create a unique string id for a slab cache:
5622 *
5623 * Format :[flags-]size
5624 */
5625static char *create_unique_id(struct kmem_cache *s)
5626{
5627 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5628 char *p = name;
5629
5630 BUG_ON(!name);
5631
5632 *p++ = ':';
5633 /*
5634 * First flags affecting slabcache operations. We will only
5635 * get here for aliasable slabs so we do not need to support
5636 * too many flags. The flags here must cover all flags that
5637 * are matched during merging to guarantee that the id is
5638 * unique.
5639 */
5640 if (s->flags & SLAB_CACHE_DMA)
5641 *p++ = 'd';
5642 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5643 *p++ = 'a';
5644 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5645 *p++ = 'F';
5646 if (!(s->flags & SLAB_NOTRACK))
5647 *p++ = 't';
5648 if (s->flags & SLAB_ACCOUNT)
5649 *p++ = 'A';
5650 if (p != name + 1)
5651 *p++ = '-';
5652 p += sprintf(p, "%07d", s->size);
5653
5654 BUG_ON(p > name + ID_STR_LENGTH - 1);
5655 return name;
5656}
5657
5658static int sysfs_slab_add(struct kmem_cache *s)
5659{
5660 int err;
5661 const char *name;
5662 struct kset *kset = cache_kset(s);
5663 int unmergeable = slab_unmergeable(s);
5664
5665 if (!kset) {
5666 kobject_init(&s->kobj, &slab_ktype);
5667 return 0;
5668 }
5669
5670 if (unmergeable) {
5671 /*
5672 * Slabcache can never be merged so we can use the name proper.
5673 * This is typically the case for debug situations. In that
5674 * case we can catch duplicate names easily.
5675 */
5676 sysfs_remove_link(&slab_kset->kobj, s->name);
5677 name = s->name;
5678 } else {
5679 /*
5680 * Create a unique name for the slab as a target
5681 * for the symlinks.
5682 */
5683 name = create_unique_id(s);
5684 }
5685
5686 s->kobj.kset = kset;
5687 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5688 if (err)
5689 goto out;
5690
5691 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5692 if (err)
5693 goto out_del_kobj;
5694
5695#ifdef CONFIG_MEMCG
5696 if (is_root_cache(s) && memcg_sysfs_enabled) {
5697 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5698 if (!s->memcg_kset) {
5699 err = -ENOMEM;
5700 goto out_del_kobj;
5701 }
5702 }
5703#endif
5704
5705 kobject_uevent(&s->kobj, KOBJ_ADD);
5706 if (!unmergeable) {
5707 /* Setup first alias */
5708 sysfs_slab_alias(s, s->name);
5709 }
5710out:
5711 if (!unmergeable)
5712 kfree(name);
5713 return err;
5714out_del_kobj:
5715 kobject_del(&s->kobj);
5716 goto out;
5717}
5718
5719void sysfs_slab_remove(struct kmem_cache *s)
5720{
5721 if (slab_state < FULL)
5722 /*
5723 * Sysfs has not been setup yet so no need to remove the
5724 * cache from sysfs.
5725 */
5726 return;
5727
5728#ifdef CONFIG_MEMCG
5729 kset_unregister(s->memcg_kset);
5730#endif
5731 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5732 kobject_del(&s->kobj);
5733 kobject_put(&s->kobj);
5734}
5735
5736/*
5737 * Need to buffer aliases during bootup until sysfs becomes
5738 * available lest we lose that information.
5739 */
5740struct saved_alias {
5741 struct kmem_cache *s;
5742 const char *name;
5743 struct saved_alias *next;
5744};
5745
5746static struct saved_alias *alias_list;
5747
5748static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5749{
5750 struct saved_alias *al;
5751
5752 if (slab_state == FULL) {
5753 /*
5754 * If we have a leftover link then remove it.
5755 */
5756 sysfs_remove_link(&slab_kset->kobj, name);
5757 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5758 }
5759
5760 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5761 if (!al)
5762 return -ENOMEM;
5763
5764 al->s = s;
5765 al->name = name;
5766 al->next = alias_list;
5767 alias_list = al;
5768 return 0;
5769}
5770
5771static int __init slab_sysfs_init(void)
5772{
5773 struct kmem_cache *s;
5774 int err;
5775
5776 mutex_lock(&slab_mutex);
5777
5778 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5779 if (!slab_kset) {
5780 mutex_unlock(&slab_mutex);
5781 pr_err("Cannot register slab subsystem.\n");
5782 return -ENOSYS;
5783 }
5784
5785 slab_state = FULL;
5786
5787 list_for_each_entry(s, &slab_caches, list) {
5788 err = sysfs_slab_add(s);
5789 if (err)
5790 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5791 s->name);
5792 }
5793
5794 while (alias_list) {
5795 struct saved_alias *al = alias_list;
5796
5797 alias_list = alias_list->next;
5798 err = sysfs_slab_alias(al->s, al->name);
5799 if (err)
5800 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5801 al->name);
5802 kfree(al);
5803 }
5804
5805 mutex_unlock(&slab_mutex);
5806 resiliency_test();
5807 return 0;
5808}
5809
5810__initcall(slab_sysfs_init);
5811#endif /* CONFIG_SYSFS */
5812
5813/*
5814 * The /proc/slabinfo ABI
5815 */
5816#ifdef CONFIG_SLABINFO
5817void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5818{
5819 unsigned long nr_slabs = 0;
5820 unsigned long nr_objs = 0;
5821 unsigned long nr_free = 0;
5822 int node;
5823 struct kmem_cache_node *n;
5824
5825 for_each_kmem_cache_node(s, node, n) {
5826 nr_slabs += node_nr_slabs(n);
5827 nr_objs += node_nr_objs(n);
5828 nr_free += count_partial(n, count_free);
5829 }
5830
5831 sinfo->active_objs = nr_objs - nr_free;
5832 sinfo->num_objs = nr_objs;
5833 sinfo->active_slabs = nr_slabs;
5834 sinfo->num_slabs = nr_slabs;
5835 sinfo->objects_per_slab = oo_objects(s->oo);
5836 sinfo->cache_order = oo_order(s->oo);
5837}
5838
5839void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5840{
5841}
5842
5843ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5844 size_t count, loff_t *ppos)
5845{
5846 return -EIO;
5847}
5848#endif /* CONFIG_SLABINFO */
5849
generated by cgit at 2024-05-18 18:16:55 (GMT)