Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * SLUB: A slab allocator that limits cache line use instead of queuing
4 * objects in per cpu and per node lists.
5 *
6 * The allocator synchronizes using per slab locks or atomic operations
7 * and only uses a centralized lock to manage a pool of partial slabs.
8 *
9 * (C) 2007 SGI, Christoph Lameter
10 * (C) 2011 Linux Foundation, Christoph Lameter
11 */
12
13#include <linux/mm.h>
14#include <linux/swap.h> /* struct reclaim_state */
15#include <linux/module.h>
16#include <linux/bit_spinlock.h>
17#include <linux/interrupt.h>
18#include <linux/swab.h>
19#include <linux/bitops.h>
20#include <linux/slab.h>
21#include "slab.h"
22#include <linux/proc_fs.h>
23#include <linux/seq_file.h>
24#include <linux/kasan.h>
25#include <linux/kmsan.h>
26#include <linux/cpu.h>
27#include <linux/cpuset.h>
28#include <linux/mempolicy.h>
29#include <linux/ctype.h>
30#include <linux/stackdepot.h>
31#include <linux/debugobjects.h>
32#include <linux/kallsyms.h>
33#include <linux/kfence.h>
34#include <linux/memory.h>
35#include <linux/math64.h>
36#include <linux/fault-inject.h>
37#include <linux/stacktrace.h>
38#include <linux/prefetch.h>
39#include <linux/memcontrol.h>
40#include <linux/random.h>
41#include <kunit/test.h>
42#include <kunit/test-bug.h>
43#include <linux/sort.h>
44
45#include <linux/debugfs.h>
46#include <trace/events/kmem.h>
47
48#include "internal.h"
49
50/*
51 * Lock order:
52 * 1. slab_mutex (Global Mutex)
53 * 2. node->list_lock (Spinlock)
54 * 3. kmem_cache->cpu_slab->lock (Local lock)
55 * 4. slab_lock(slab) (Only on some arches)
56 * 5. object_map_lock (Only for debugging)
57 *
58 * slab_mutex
59 *
60 * The role of the slab_mutex is to protect the list of all the slabs
61 * and to synchronize major metadata changes to slab cache structures.
62 * Also synchronizes memory hotplug callbacks.
63 *
64 * slab_lock
65 *
66 * The slab_lock is a wrapper around the page lock, thus it is a bit
67 * spinlock.
68 *
69 * The slab_lock is only used on arches that do not have the ability
70 * to do a cmpxchg_double. It only protects:
71 *
72 * A. slab->freelist -> List of free objects in a slab
73 * B. slab->inuse -> Number of objects in use
74 * C. slab->objects -> Number of objects in slab
75 * D. slab->frozen -> frozen state
76 *
77 * Frozen slabs
78 *
79 * If a slab is frozen then it is exempt from list management. It is not
80 * on any list except per cpu partial list. The processor that froze the
81 * slab is the one who can perform list operations on the slab. Other
82 * processors may put objects onto the freelist but the processor that
83 * froze the slab is the only one that can retrieve the objects from the
84 * slab's freelist.
85 *
86 * list_lock
87 *
88 * The list_lock protects the partial and full list on each node and
89 * the partial slab counter. If taken then no new slabs may be added or
90 * removed from the lists nor make the number of partial slabs be modified.
91 * (Note that the total number of slabs is an atomic value that may be
92 * modified without taking the list lock).
93 *
94 * The list_lock is a centralized lock and thus we avoid taking it as
95 * much as possible. As long as SLUB does not have to handle partial
96 * slabs, operations can continue without any centralized lock. F.e.
97 * allocating a long series of objects that fill up slabs does not require
98 * the list lock.
99 *
100 * For debug caches, all allocations are forced to go through a list_lock
101 * protected region to serialize against concurrent validation.
102 *
103 * cpu_slab->lock local lock
104 *
105 * This locks protect slowpath manipulation of all kmem_cache_cpu fields
106 * except the stat counters. This is a percpu structure manipulated only by
107 * the local cpu, so the lock protects against being preempted or interrupted
108 * by an irq. Fast path operations rely on lockless operations instead.
109 *
110 * On PREEMPT_RT, the local lock neither disables interrupts nor preemption
111 * which means the lockless fastpath cannot be used as it might interfere with
112 * an in-progress slow path operations. In this case the local lock is always
113 * taken but it still utilizes the freelist for the common operations.
114 *
115 * lockless fastpaths
116 *
117 * The fast path allocation (slab_alloc_node()) and freeing (do_slab_free())
118 * are fully lockless when satisfied from the percpu slab (and when
119 * cmpxchg_double is possible to use, otherwise slab_lock is taken).
120 * They also don't disable preemption or migration or irqs. They rely on
121 * the transaction id (tid) field to detect being preempted or moved to
122 * another cpu.
123 *
124 * irq, preemption, migration considerations
125 *
126 * Interrupts are disabled as part of list_lock or local_lock operations, or
127 * around the slab_lock operation, in order to make the slab allocator safe
128 * to use in the context of an irq.
129 *
130 * In addition, preemption (or migration on PREEMPT_RT) is disabled in the
131 * allocation slowpath, bulk allocation, and put_cpu_partial(), so that the
132 * local cpu doesn't change in the process and e.g. the kmem_cache_cpu pointer
133 * doesn't have to be revalidated in each section protected by the local lock.
134 *
135 * SLUB assigns one slab for allocation to each processor.
136 * Allocations only occur from these slabs called cpu slabs.
137 *
138 * Slabs with free elements are kept on a partial list and during regular
139 * operations no list for full slabs is used. If an object in a full slab is
140 * freed then the slab will show up again on the partial lists.
141 * We track full slabs for debugging purposes though because otherwise we
142 * cannot scan all objects.
143 *
144 * Slabs are freed when they become empty. Teardown and setup is
145 * minimal so we rely on the page allocators per cpu caches for
146 * fast frees and allocs.
147 *
148 * slab->frozen The slab is frozen and exempt from list processing.
149 * This means that the slab is dedicated to a purpose
150 * such as satisfying allocations for a specific
151 * processor. Objects may be freed in the slab while
152 * it is frozen but slab_free will then skip the usual
153 * list operations. It is up to the processor holding
154 * the slab to integrate the slab into the slab lists
155 * when the slab is no longer needed.
156 *
157 * One use of this flag is to mark slabs that are
158 * used for allocations. Then such a slab becomes a cpu
159 * slab. The cpu slab may be equipped with an additional
160 * freelist that allows lockless access to
161 * free objects in addition to the regular freelist
162 * that requires the slab lock.
163 *
164 * SLAB_DEBUG_FLAGS Slab requires special handling due to debug
165 * options set. This moves slab handling out of
166 * the fast path and disables lockless freelists.
167 */
168
169/*
170 * We could simply use migrate_disable()/enable() but as long as it's a
171 * function call even on !PREEMPT_RT, use inline preempt_disable() there.
172 */
173#ifndef CONFIG_PREEMPT_RT
174#define slub_get_cpu_ptr(var) get_cpu_ptr(var)
175#define slub_put_cpu_ptr(var) put_cpu_ptr(var)
176#define USE_LOCKLESS_FAST_PATH() (true)
177#else
178#define slub_get_cpu_ptr(var) \
179({ \
180 migrate_disable(); \
181 this_cpu_ptr(var); \
182})
183#define slub_put_cpu_ptr(var) \
184do { \
185 (void)(var); \
186 migrate_enable(); \
187} while (0)
188#define USE_LOCKLESS_FAST_PATH() (false)
189#endif
190
191#ifndef CONFIG_SLUB_TINY
192#define __fastpath_inline __always_inline
193#else
194#define __fastpath_inline
195#endif
196
197#ifdef CONFIG_SLUB_DEBUG
198#ifdef CONFIG_SLUB_DEBUG_ON
199DEFINE_STATIC_KEY_TRUE(slub_debug_enabled);
200#else
201DEFINE_STATIC_KEY_FALSE(slub_debug_enabled);
202#endif
203#endif /* CONFIG_SLUB_DEBUG */
204
205/* Structure holding parameters for get_partial() call chain */
206struct partial_context {
207 struct slab **slab;
208 gfp_t flags;
209 unsigned int orig_size;
210};
211
212static inline bool kmem_cache_debug(struct kmem_cache *s)
213{
214 return kmem_cache_debug_flags(s, SLAB_DEBUG_FLAGS);
215}
216
217static inline bool slub_debug_orig_size(struct kmem_cache *s)
218{
219 return (kmem_cache_debug_flags(s, SLAB_STORE_USER) &&
220 (s->flags & SLAB_KMALLOC));
221}
222
223void *fixup_red_left(struct kmem_cache *s, void *p)
224{
225 if (kmem_cache_debug_flags(s, SLAB_RED_ZONE))
226 p += s->red_left_pad;
227
228 return p;
229}
230
231static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
232{
233#ifdef CONFIG_SLUB_CPU_PARTIAL
234 return !kmem_cache_debug(s);
235#else
236 return false;
237#endif
238}
239
240/*
241 * Issues still to be resolved:
242 *
243 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
244 *
245 * - Variable sizing of the per node arrays
246 */
247
248/* Enable to log cmpxchg failures */
249#undef SLUB_DEBUG_CMPXCHG
250
251#ifndef CONFIG_SLUB_TINY
252/*
253 * Minimum number of partial slabs. These will be left on the partial
254 * lists even if they are empty. kmem_cache_shrink may reclaim them.
255 */
256#define MIN_PARTIAL 5
257
258/*
259 * Maximum number of desirable partial slabs.
260 * The existence of more partial slabs makes kmem_cache_shrink
261 * sort the partial list by the number of objects in use.
262 */
263#define MAX_PARTIAL 10
264#else
265#define MIN_PARTIAL 0
266#define MAX_PARTIAL 0
267#endif
268
269#define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
270 SLAB_POISON | SLAB_STORE_USER)
271
272/*
273 * These debug flags cannot use CMPXCHG because there might be consistency
274 * issues when checking or reading debug information
275 */
276#define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
277 SLAB_TRACE)
278
279
280/*
281 * Debugging flags that require metadata to be stored in the slab. These get
282 * disabled when slub_debug=O is used and a cache's min order increases with
283 * metadata.
284 */
285#define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
286
287#define OO_SHIFT 16
288#define OO_MASK ((1 << OO_SHIFT) - 1)
289#define MAX_OBJS_PER_PAGE 32767 /* since slab.objects is u15 */
290
291/* Internal SLUB flags */
292/* Poison object */
293#define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
294/* Use cmpxchg_double */
295#define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
296
297/*
298 * Tracking user of a slab.
299 */
300#define TRACK_ADDRS_COUNT 16
301struct track {
302 unsigned long addr; /* Called from address */
303#ifdef CONFIG_STACKDEPOT
304 depot_stack_handle_t handle;
305#endif
306 int cpu; /* Was running on cpu */
307 int pid; /* Pid context */
308 unsigned long when; /* When did the operation occur */
309};
310
311enum track_item { TRACK_ALLOC, TRACK_FREE };
312
313#ifdef SLAB_SUPPORTS_SYSFS
314static int sysfs_slab_add(struct kmem_cache *);
315static int sysfs_slab_alias(struct kmem_cache *, const char *);
316#else
317static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
318static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
319 { return 0; }
320#endif
321
322#if defined(CONFIG_DEBUG_FS) && defined(CONFIG_SLUB_DEBUG)
323static void debugfs_slab_add(struct kmem_cache *);
324#else
325static inline void debugfs_slab_add(struct kmem_cache *s) { }
326#endif
327
328static inline void stat(const struct kmem_cache *s, enum stat_item si)
329{
330#ifdef CONFIG_SLUB_STATS
331 /*
332 * The rmw is racy on a preemptible kernel but this is acceptable, so
333 * avoid this_cpu_add()'s irq-disable overhead.
334 */
335 raw_cpu_inc(s->cpu_slab->stat[si]);
336#endif
337}
338
339/*
340 * Tracks for which NUMA nodes we have kmem_cache_nodes allocated.
341 * Corresponds to node_state[N_NORMAL_MEMORY], but can temporarily
342 * differ during memory hotplug/hotremove operations.
343 * Protected by slab_mutex.
344 */
345static nodemask_t slab_nodes;
346
347#ifndef CONFIG_SLUB_TINY
348/*
349 * Workqueue used for flush_cpu_slab().
350 */
351static struct workqueue_struct *flushwq;
352#endif
353
354/********************************************************************
355 * Core slab cache functions
356 *******************************************************************/
357
358/*
359 * Returns freelist pointer (ptr). With hardening, this is obfuscated
360 * with an XOR of the address where the pointer is held and a per-cache
361 * random number.
362 */
363static inline void *freelist_ptr(const struct kmem_cache *s, void *ptr,
364 unsigned long ptr_addr)
365{
366#ifdef CONFIG_SLAB_FREELIST_HARDENED
367 /*
368 * When CONFIG_KASAN_SW/HW_TAGS is enabled, ptr_addr might be tagged.
369 * Normally, this doesn't cause any issues, as both set_freepointer()
370 * and get_freepointer() are called with a pointer with the same tag.
371 * However, there are some issues with CONFIG_SLUB_DEBUG code. For
372 * example, when __free_slub() iterates over objects in a cache, it
373 * passes untagged pointers to check_object(). check_object() in turns
374 * calls get_freepointer() with an untagged pointer, which causes the
375 * freepointer to be restored incorrectly.
376 */
377 return (void *)((unsigned long)ptr ^ s->random ^
378 swab((unsigned long)kasan_reset_tag((void *)ptr_addr)));
379#else
380 return ptr;
381#endif
382}
383
384/* Returns the freelist pointer recorded at location ptr_addr. */
385static inline void *freelist_dereference(const struct kmem_cache *s,
386 void *ptr_addr)
387{
388 return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr),
389 (unsigned long)ptr_addr);
390}
391
392static inline void *get_freepointer(struct kmem_cache *s, void *object)
393{
394 object = kasan_reset_tag(object);
395 return freelist_dereference(s, object + s->offset);
396}
397
398#ifndef CONFIG_SLUB_TINY
399static void prefetch_freepointer(const struct kmem_cache *s, void *object)
400{
401 prefetchw(object + s->offset);
402}
403#endif
404
405/*
406 * When running under KMSAN, get_freepointer_safe() may return an uninitialized
407 * pointer value in the case the current thread loses the race for the next
408 * memory chunk in the freelist. In that case this_cpu_cmpxchg_double() in
409 * slab_alloc_node() will fail, so the uninitialized value won't be used, but
410 * KMSAN will still check all arguments of cmpxchg because of imperfect
411 * handling of inline assembly.
412 * To work around this problem, we apply __no_kmsan_checks to ensure that
413 * get_freepointer_safe() returns initialized memory.
414 */
415__no_kmsan_checks
416static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
417{
418 unsigned long freepointer_addr;
419 void *p;
420
421 if (!debug_pagealloc_enabled_static())
422 return get_freepointer(s, object);
423
424 object = kasan_reset_tag(object);
425 freepointer_addr = (unsigned long)object + s->offset;
426 copy_from_kernel_nofault(&p, (void **)freepointer_addr, sizeof(p));
427 return freelist_ptr(s, p, freepointer_addr);
428}
429
430static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
431{
432 unsigned long freeptr_addr = (unsigned long)object + s->offset;
433
434#ifdef CONFIG_SLAB_FREELIST_HARDENED
435 BUG_ON(object == fp); /* naive detection of double free or corruption */
436#endif
437
438 freeptr_addr = (unsigned long)kasan_reset_tag((void *)freeptr_addr);
439 *(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr);
440}
441
442/* Loop over all objects in a slab */
443#define for_each_object(__p, __s, __addr, __objects) \
444 for (__p = fixup_red_left(__s, __addr); \
445 __p < (__addr) + (__objects) * (__s)->size; \
446 __p += (__s)->size)
447
448static inline unsigned int order_objects(unsigned int order, unsigned int size)
449{
450 return ((unsigned int)PAGE_SIZE << order) / size;
451}
452
453static inline struct kmem_cache_order_objects oo_make(unsigned int order,
454 unsigned int size)
455{
456 struct kmem_cache_order_objects x = {
457 (order << OO_SHIFT) + order_objects(order, size)
458 };
459
460 return x;
461}
462
463static inline unsigned int oo_order(struct kmem_cache_order_objects x)
464{
465 return x.x >> OO_SHIFT;
466}
467
468static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
469{
470 return x.x & OO_MASK;
471}
472
473#ifdef CONFIG_SLUB_CPU_PARTIAL
474static void slub_set_cpu_partial(struct kmem_cache *s, unsigned int nr_objects)
475{
476 unsigned int nr_slabs;
477
478 s->cpu_partial = nr_objects;
479
480 /*
481 * We take the number of objects but actually limit the number of
482 * slabs on the per cpu partial list, in order to limit excessive
483 * growth of the list. For simplicity we assume that the slabs will
484 * be half-full.
485 */
486 nr_slabs = DIV_ROUND_UP(nr_objects * 2, oo_objects(s->oo));
487 s->cpu_partial_slabs = nr_slabs;
488}
489#else
490static inline void
491slub_set_cpu_partial(struct kmem_cache *s, unsigned int nr_objects)
492{
493}
494#endif /* CONFIG_SLUB_CPU_PARTIAL */
495
496/*
497 * Per slab locking using the pagelock
498 */
499static __always_inline void slab_lock(struct slab *slab)
500{
501 struct page *page = slab_page(slab);
502
503 VM_BUG_ON_PAGE(PageTail(page), page);
504 bit_spin_lock(PG_locked, &page->flags);
505}
506
507static __always_inline void slab_unlock(struct slab *slab)
508{
509 struct page *page = slab_page(slab);
510
511 VM_BUG_ON_PAGE(PageTail(page), page);
512 __bit_spin_unlock(PG_locked, &page->flags);
513}
514
515/*
516 * Interrupts must be disabled (for the fallback code to work right), typically
517 * by an _irqsave() lock variant. On PREEMPT_RT the preempt_disable(), which is
518 * part of bit_spin_lock(), is sufficient because the policy is not to allow any
519 * allocation/ free operation in hardirq context. Therefore nothing can
520 * interrupt the operation.
521 */
522static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct slab *slab,
523 void *freelist_old, unsigned long counters_old,
524 void *freelist_new, unsigned long counters_new,
525 const char *n)
526{
527 if (USE_LOCKLESS_FAST_PATH())
528 lockdep_assert_irqs_disabled();
529#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
530 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
531 if (s->flags & __CMPXCHG_DOUBLE) {
532 if (cmpxchg_double(&slab->freelist, &slab->counters,
533 freelist_old, counters_old,
534 freelist_new, counters_new))
535 return true;
536 } else
537#endif
538 {
539 slab_lock(slab);
540 if (slab->freelist == freelist_old &&
541 slab->counters == counters_old) {
542 slab->freelist = freelist_new;
543 slab->counters = counters_new;
544 slab_unlock(slab);
545 return true;
546 }
547 slab_unlock(slab);
548 }
549
550 cpu_relax();
551 stat(s, CMPXCHG_DOUBLE_FAIL);
552
553#ifdef SLUB_DEBUG_CMPXCHG
554 pr_info("%s %s: cmpxchg double redo ", n, s->name);
555#endif
556
557 return false;
558}
559
560static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct slab *slab,
561 void *freelist_old, unsigned long counters_old,
562 void *freelist_new, unsigned long counters_new,
563 const char *n)
564{
565#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
566 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
567 if (s->flags & __CMPXCHG_DOUBLE) {
568 if (cmpxchg_double(&slab->freelist, &slab->counters,
569 freelist_old, counters_old,
570 freelist_new, counters_new))
571 return true;
572 } else
573#endif
574 {
575 unsigned long flags;
576
577 local_irq_save(flags);
578 slab_lock(slab);
579 if (slab->freelist == freelist_old &&
580 slab->counters == counters_old) {
581 slab->freelist = freelist_new;
582 slab->counters = counters_new;
583 slab_unlock(slab);
584 local_irq_restore(flags);
585 return true;
586 }
587 slab_unlock(slab);
588 local_irq_restore(flags);
589 }
590
591 cpu_relax();
592 stat(s, CMPXCHG_DOUBLE_FAIL);
593
594#ifdef SLUB_DEBUG_CMPXCHG
595 pr_info("%s %s: cmpxchg double redo ", n, s->name);
596#endif
597
598 return false;
599}
600
601#ifdef CONFIG_SLUB_DEBUG
602static unsigned long object_map[BITS_TO_LONGS(MAX_OBJS_PER_PAGE)];
603static DEFINE_SPINLOCK(object_map_lock);
604
605static void __fill_map(unsigned long *obj_map, struct kmem_cache *s,
606 struct slab *slab)
607{
608 void *addr = slab_address(slab);
609 void *p;
610
611 bitmap_zero(obj_map, slab->objects);
612
613 for (p = slab->freelist; p; p = get_freepointer(s, p))
614 set_bit(__obj_to_index(s, addr, p), obj_map);
615}
616
617#if IS_ENABLED(CONFIG_KUNIT)
618static bool slab_add_kunit_errors(void)
619{
620 struct kunit_resource *resource;
621
622 if (!kunit_get_current_test())
623 return false;
624
625 resource = kunit_find_named_resource(current->kunit_test, "slab_errors");
626 if (!resource)
627 return false;
628
629 (*(int *)resource->data)++;
630 kunit_put_resource(resource);
631 return true;
632}
633#else
634static inline bool slab_add_kunit_errors(void) { return false; }
635#endif
636
637static inline unsigned int size_from_object(struct kmem_cache *s)
638{
639 if (s->flags & SLAB_RED_ZONE)
640 return s->size - s->red_left_pad;
641
642 return s->size;
643}
644
645static inline void *restore_red_left(struct kmem_cache *s, void *p)
646{
647 if (s->flags & SLAB_RED_ZONE)
648 p -= s->red_left_pad;
649
650 return p;
651}
652
653/*
654 * Debug settings:
655 */
656#if defined(CONFIG_SLUB_DEBUG_ON)
657static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
658#else
659static slab_flags_t slub_debug;
660#endif
661
662static char *slub_debug_string;
663static int disable_higher_order_debug;
664
665/*
666 * slub is about to manipulate internal object metadata. This memory lies
667 * outside the range of the allocated object, so accessing it would normally
668 * be reported by kasan as a bounds error. metadata_access_enable() is used
669 * to tell kasan that these accesses are OK.
670 */
671static inline void metadata_access_enable(void)
672{
673 kasan_disable_current();
674}
675
676static inline void metadata_access_disable(void)
677{
678 kasan_enable_current();
679}
680
681/*
682 * Object debugging
683 */
684
685/* Verify that a pointer has an address that is valid within a slab page */
686static inline int check_valid_pointer(struct kmem_cache *s,
687 struct slab *slab, void *object)
688{
689 void *base;
690
691 if (!object)
692 return 1;
693
694 base = slab_address(slab);
695 object = kasan_reset_tag(object);
696 object = restore_red_left(s, object);
697 if (object < base || object >= base + slab->objects * s->size ||
698 (object - base) % s->size) {
699 return 0;
700 }
701
702 return 1;
703}
704
705static void print_section(char *level, char *text, u8 *addr,
706 unsigned int length)
707{
708 metadata_access_enable();
709 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS,
710 16, 1, kasan_reset_tag((void *)addr), length, 1);
711 metadata_access_disable();
712}
713
714/*
715 * See comment in calculate_sizes().
716 */
717static inline bool freeptr_outside_object(struct kmem_cache *s)
718{
719 return s->offset >= s->inuse;
720}
721
722/*
723 * Return offset of the end of info block which is inuse + free pointer if
724 * not overlapping with object.
725 */
726static inline unsigned int get_info_end(struct kmem_cache *s)
727{
728 if (freeptr_outside_object(s))
729 return s->inuse + sizeof(void *);
730 else
731 return s->inuse;
732}
733
734static struct track *get_track(struct kmem_cache *s, void *object,
735 enum track_item alloc)
736{
737 struct track *p;
738
739 p = object + get_info_end(s);
740
741 return kasan_reset_tag(p + alloc);
742}
743
744#ifdef CONFIG_STACKDEPOT
745static noinline depot_stack_handle_t set_track_prepare(void)
746{
747 depot_stack_handle_t handle;
748 unsigned long entries[TRACK_ADDRS_COUNT];
749 unsigned int nr_entries;
750
751 nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 3);
752 handle = stack_depot_save(entries, nr_entries, GFP_NOWAIT);
753
754 return handle;
755}
756#else
757static inline depot_stack_handle_t set_track_prepare(void)
758{
759 return 0;
760}
761#endif
762
763static void set_track_update(struct kmem_cache *s, void *object,
764 enum track_item alloc, unsigned long addr,
765 depot_stack_handle_t handle)
766{
767 struct track *p = get_track(s, object, alloc);
768
769#ifdef CONFIG_STACKDEPOT
770 p->handle = handle;
771#endif
772 p->addr = addr;
773 p->cpu = smp_processor_id();
774 p->pid = current->pid;
775 p->when = jiffies;
776}
777
778static __always_inline void set_track(struct kmem_cache *s, void *object,
779 enum track_item alloc, unsigned long addr)
780{
781 depot_stack_handle_t handle = set_track_prepare();
782
783 set_track_update(s, object, alloc, addr, handle);
784}
785
786static void init_tracking(struct kmem_cache *s, void *object)
787{
788 struct track *p;
789
790 if (!(s->flags & SLAB_STORE_USER))
791 return;
792
793 p = get_track(s, object, TRACK_ALLOC);
794 memset(p, 0, 2*sizeof(struct track));
795}
796
797static void print_track(const char *s, struct track *t, unsigned long pr_time)
798{
799 depot_stack_handle_t handle __maybe_unused;
800
801 if (!t->addr)
802 return;
803
804 pr_err("%s in %pS age=%lu cpu=%u pid=%d\n",
805 s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
806#ifdef CONFIG_STACKDEPOT
807 handle = READ_ONCE(t->handle);
808 if (handle)
809 stack_depot_print(handle);
810 else
811 pr_err("object allocation/free stack trace missing\n");
812#endif
813}
814
815void print_tracking(struct kmem_cache *s, void *object)
816{
817 unsigned long pr_time = jiffies;
818 if (!(s->flags & SLAB_STORE_USER))
819 return;
820
821 print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
822 print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
823}
824
825static void print_slab_info(const struct slab *slab)
826{
827 struct folio *folio = (struct folio *)slab_folio(slab);
828
829 pr_err("Slab 0x%p objects=%u used=%u fp=0x%p flags=%pGp\n",
830 slab, slab->objects, slab->inuse, slab->freelist,
831 folio_flags(folio, 0));
832}
833
834/*
835 * kmalloc caches has fixed sizes (mostly power of 2), and kmalloc() API
836 * family will round up the real request size to these fixed ones, so
837 * there could be an extra area than what is requested. Save the original
838 * request size in the meta data area, for better debug and sanity check.
839 */
840static inline void set_orig_size(struct kmem_cache *s,
841 void *object, unsigned int orig_size)
842{
843 void *p = kasan_reset_tag(object);
844
845 if (!slub_debug_orig_size(s))
846 return;
847
848#ifdef CONFIG_KASAN_GENERIC
849 /*
850 * KASAN could save its free meta data in object's data area at
851 * offset 0, if the size is larger than 'orig_size', it will
852 * overlap the data redzone in [orig_size+1, object_size], and
853 * the check should be skipped.
854 */
855 if (kasan_metadata_size(s, true) > orig_size)
856 orig_size = s->object_size;
857#endif
858
859 p += get_info_end(s);
860 p += sizeof(struct track) * 2;
861
862 *(unsigned int *)p = orig_size;
863}
864
865static inline unsigned int get_orig_size(struct kmem_cache *s, void *object)
866{
867 void *p = kasan_reset_tag(object);
868
869 if (!slub_debug_orig_size(s))
870 return s->object_size;
871
872 p += get_info_end(s);
873 p += sizeof(struct track) * 2;
874
875 return *(unsigned int *)p;
876}
877
878void skip_orig_size_check(struct kmem_cache *s, const void *object)
879{
880 set_orig_size(s, (void *)object, s->object_size);
881}
882
883static void slab_bug(struct kmem_cache *s, char *fmt, ...)
884{
885 struct va_format vaf;
886 va_list args;
887
888 va_start(args, fmt);
889 vaf.fmt = fmt;
890 vaf.va = &args;
891 pr_err("=============================================================================\n");
892 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
893 pr_err("-----------------------------------------------------------------------------\n\n");
894 va_end(args);
895}
896
897__printf(2, 3)
898static void slab_fix(struct kmem_cache *s, char *fmt, ...)
899{
900 struct va_format vaf;
901 va_list args;
902
903 if (slab_add_kunit_errors())
904 return;
905
906 va_start(args, fmt);
907 vaf.fmt = fmt;
908 vaf.va = &args;
909 pr_err("FIX %s: %pV\n", s->name, &vaf);
910 va_end(args);
911}
912
913static void print_trailer(struct kmem_cache *s, struct slab *slab, u8 *p)
914{
915 unsigned int off; /* Offset of last byte */
916 u8 *addr = slab_address(slab);
917
918 print_tracking(s, p);
919
920 print_slab_info(slab);
921
922 pr_err("Object 0x%p @offset=%tu fp=0x%p\n\n",
923 p, p - addr, get_freepointer(s, p));
924
925 if (s->flags & SLAB_RED_ZONE)
926 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
927 s->red_left_pad);
928 else if (p > addr + 16)
929 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
930
931 print_section(KERN_ERR, "Object ", p,
932 min_t(unsigned int, s->object_size, PAGE_SIZE));
933 if (s->flags & SLAB_RED_ZONE)
934 print_section(KERN_ERR, "Redzone ", p + s->object_size,
935 s->inuse - s->object_size);
936
937 off = get_info_end(s);
938
939 if (s->flags & SLAB_STORE_USER)
940 off += 2 * sizeof(struct track);
941
942 if (slub_debug_orig_size(s))
943 off += sizeof(unsigned int);
944
945 off += kasan_metadata_size(s, false);
946
947 if (off != size_from_object(s))
948 /* Beginning of the filler is the free pointer */
949 print_section(KERN_ERR, "Padding ", p + off,
950 size_from_object(s) - off);
951
952 dump_stack();
953}
954
955static void object_err(struct kmem_cache *s, struct slab *slab,
956 u8 *object, char *reason)
957{
958 if (slab_add_kunit_errors())
959 return;
960
961 slab_bug(s, "%s", reason);
962 print_trailer(s, slab, object);
963 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
964}
965
966static bool freelist_corrupted(struct kmem_cache *s, struct slab *slab,
967 void **freelist, void *nextfree)
968{
969 if ((s->flags & SLAB_CONSISTENCY_CHECKS) &&
970 !check_valid_pointer(s, slab, nextfree) && freelist) {
971 object_err(s, slab, *freelist, "Freechain corrupt");
972 *freelist = NULL;
973 slab_fix(s, "Isolate corrupted freechain");
974 return true;
975 }
976
977 return false;
978}
979
980static __printf(3, 4) void slab_err(struct kmem_cache *s, struct slab *slab,
981 const char *fmt, ...)
982{
983 va_list args;
984 char buf[100];
985
986 if (slab_add_kunit_errors())
987 return;
988
989 va_start(args, fmt);
990 vsnprintf(buf, sizeof(buf), fmt, args);
991 va_end(args);
992 slab_bug(s, "%s", buf);
993 print_slab_info(slab);
994 dump_stack();
995 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
996}
997
998static void init_object(struct kmem_cache *s, void *object, u8 val)
999{
1000 u8 *p = kasan_reset_tag(object);
1001 unsigned int poison_size = s->object_size;
1002
1003 if (s->flags & SLAB_RED_ZONE) {
1004 memset(p - s->red_left_pad, val, s->red_left_pad);
1005
1006 if (slub_debug_orig_size(s) && val == SLUB_RED_ACTIVE) {
1007 /*
1008 * Redzone the extra allocated space by kmalloc than
1009 * requested, and the poison size will be limited to
1010 * the original request size accordingly.
1011 */
1012 poison_size = get_orig_size(s, object);
1013 }
1014 }
1015
1016 if (s->flags & __OBJECT_POISON) {
1017 memset(p, POISON_FREE, poison_size - 1);
1018 p[poison_size - 1] = POISON_END;
1019 }
1020
1021 if (s->flags & SLAB_RED_ZONE)
1022 memset(p + poison_size, val, s->inuse - poison_size);
1023}
1024
1025static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
1026 void *from, void *to)
1027{
1028 slab_fix(s, "Restoring %s 0x%p-0x%p=0x%x", message, from, to - 1, data);
1029 memset(from, data, to - from);
1030}
1031
1032static int check_bytes_and_report(struct kmem_cache *s, struct slab *slab,
1033 u8 *object, char *what,
1034 u8 *start, unsigned int value, unsigned int bytes)
1035{
1036 u8 *fault;
1037 u8 *end;
1038 u8 *addr = slab_address(slab);
1039
1040 metadata_access_enable();
1041 fault = memchr_inv(kasan_reset_tag(start), value, bytes);
1042 metadata_access_disable();
1043 if (!fault)
1044 return 1;
1045
1046 end = start + bytes;
1047 while (end > fault && end[-1] == value)
1048 end--;
1049
1050 if (slab_add_kunit_errors())
1051 goto skip_bug_print;
1052
1053 slab_bug(s, "%s overwritten", what);
1054 pr_err("0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
1055 fault, end - 1, fault - addr,
1056 fault[0], value);
1057 print_trailer(s, slab, object);
1058 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
1059
1060skip_bug_print:
1061 restore_bytes(s, what, value, fault, end);
1062 return 0;
1063}
1064
1065/*
1066 * Object layout:
1067 *
1068 * object address
1069 * Bytes of the object to be managed.
1070 * If the freepointer may overlay the object then the free
1071 * pointer is at the middle of the object.
1072 *
1073 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
1074 * 0xa5 (POISON_END)
1075 *
1076 * object + s->object_size
1077 * Padding to reach word boundary. This is also used for Redzoning.
1078 * Padding is extended by another word if Redzoning is enabled and
1079 * object_size == inuse.
1080 *
1081 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
1082 * 0xcc (RED_ACTIVE) for objects in use.
1083 *
1084 * object + s->inuse
1085 * Meta data starts here.
1086 *
1087 * A. Free pointer (if we cannot overwrite object on free)
1088 * B. Tracking data for SLAB_STORE_USER
1089 * C. Original request size for kmalloc object (SLAB_STORE_USER enabled)
1090 * D. Padding to reach required alignment boundary or at minimum
1091 * one word if debugging is on to be able to detect writes
1092 * before the word boundary.
1093 *
1094 * Padding is done using 0x5a (POISON_INUSE)
1095 *
1096 * object + s->size
1097 * Nothing is used beyond s->size.
1098 *
1099 * If slabcaches are merged then the object_size and inuse boundaries are mostly
1100 * ignored. And therefore no slab options that rely on these boundaries
1101 * may be used with merged slabcaches.
1102 */
1103
1104static int check_pad_bytes(struct kmem_cache *s, struct slab *slab, u8 *p)
1105{
1106 unsigned long off = get_info_end(s); /* The end of info */
1107
1108 if (s->flags & SLAB_STORE_USER) {
1109 /* We also have user information there */
1110 off += 2 * sizeof(struct track);
1111
1112 if (s->flags & SLAB_KMALLOC)
1113 off += sizeof(unsigned int);
1114 }
1115
1116 off += kasan_metadata_size(s, false);
1117
1118 if (size_from_object(s) == off)
1119 return 1;
1120
1121 return check_bytes_and_report(s, slab, p, "Object padding",
1122 p + off, POISON_INUSE, size_from_object(s) - off);
1123}
1124
1125/* Check the pad bytes at the end of a slab page */
1126static void slab_pad_check(struct kmem_cache *s, struct slab *slab)
1127{
1128 u8 *start;
1129 u8 *fault;
1130 u8 *end;
1131 u8 *pad;
1132 int length;
1133 int remainder;
1134
1135 if (!(s->flags & SLAB_POISON))
1136 return;
1137
1138 start = slab_address(slab);
1139 length = slab_size(slab);
1140 end = start + length;
1141 remainder = length % s->size;
1142 if (!remainder)
1143 return;
1144
1145 pad = end - remainder;
1146 metadata_access_enable();
1147 fault = memchr_inv(kasan_reset_tag(pad), POISON_INUSE, remainder);
1148 metadata_access_disable();
1149 if (!fault)
1150 return;
1151 while (end > fault && end[-1] == POISON_INUSE)
1152 end--;
1153
1154 slab_err(s, slab, "Padding overwritten. 0x%p-0x%p @offset=%tu",
1155 fault, end - 1, fault - start);
1156 print_section(KERN_ERR, "Padding ", pad, remainder);
1157
1158 restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
1159}
1160
1161static int check_object(struct kmem_cache *s, struct slab *slab,
1162 void *object, u8 val)
1163{
1164 u8 *p = object;
1165 u8 *endobject = object + s->object_size;
1166 unsigned int orig_size;
1167
1168 if (s->flags & SLAB_RED_ZONE) {
1169 if (!check_bytes_and_report(s, slab, object, "Left Redzone",
1170 object - s->red_left_pad, val, s->red_left_pad))
1171 return 0;
1172
1173 if (!check_bytes_and_report(s, slab, object, "Right Redzone",
1174 endobject, val, s->inuse - s->object_size))
1175 return 0;
1176
1177 if (slub_debug_orig_size(s) && val == SLUB_RED_ACTIVE) {
1178 orig_size = get_orig_size(s, object);
1179
1180 if (s->object_size > orig_size &&
1181 !check_bytes_and_report(s, slab, object,
1182 "kmalloc Redzone", p + orig_size,
1183 val, s->object_size - orig_size)) {
1184 return 0;
1185 }
1186 }
1187 } else {
1188 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
1189 check_bytes_and_report(s, slab, p, "Alignment padding",
1190 endobject, POISON_INUSE,
1191 s->inuse - s->object_size);
1192 }
1193 }
1194
1195 if (s->flags & SLAB_POISON) {
1196 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
1197 (!check_bytes_and_report(s, slab, p, "Poison", p,
1198 POISON_FREE, s->object_size - 1) ||
1199 !check_bytes_and_report(s, slab, p, "End Poison",
1200 p + s->object_size - 1, POISON_END, 1)))
1201 return 0;
1202 /*
1203 * check_pad_bytes cleans up on its own.
1204 */
1205 check_pad_bytes(s, slab, p);
1206 }
1207
1208 if (!freeptr_outside_object(s) && val == SLUB_RED_ACTIVE)
1209 /*
1210 * Object and freepointer overlap. Cannot check
1211 * freepointer while object is allocated.
1212 */
1213 return 1;
1214
1215 /* Check free pointer validity */
1216 if (!check_valid_pointer(s, slab, get_freepointer(s, p))) {
1217 object_err(s, slab, p, "Freepointer corrupt");
1218 /*
1219 * No choice but to zap it and thus lose the remainder
1220 * of the free objects in this slab. May cause
1221 * another error because the object count is now wrong.
1222 */
1223 set_freepointer(s, p, NULL);
1224 return 0;
1225 }
1226 return 1;
1227}
1228
1229static int check_slab(struct kmem_cache *s, struct slab *slab)
1230{
1231 int maxobj;
1232
1233 if (!folio_test_slab(slab_folio(slab))) {
1234 slab_err(s, slab, "Not a valid slab page");
1235 return 0;
1236 }
1237
1238 maxobj = order_objects(slab_order(slab), s->size);
1239 if (slab->objects > maxobj) {
1240 slab_err(s, slab, "objects %u > max %u",
1241 slab->objects, maxobj);
1242 return 0;
1243 }
1244 if (slab->inuse > slab->objects) {
1245 slab_err(s, slab, "inuse %u > max %u",
1246 slab->inuse, slab->objects);
1247 return 0;
1248 }
1249 /* Slab_pad_check fixes things up after itself */
1250 slab_pad_check(s, slab);
1251 return 1;
1252}
1253
1254/*
1255 * Determine if a certain object in a slab is on the freelist. Must hold the
1256 * slab lock to guarantee that the chains are in a consistent state.
1257 */
1258static int on_freelist(struct kmem_cache *s, struct slab *slab, void *search)
1259{
1260 int nr = 0;
1261 void *fp;
1262 void *object = NULL;
1263 int max_objects;
1264
1265 fp = slab->freelist;
1266 while (fp && nr <= slab->objects) {
1267 if (fp == search)
1268 return 1;
1269 if (!check_valid_pointer(s, slab, fp)) {
1270 if (object) {
1271 object_err(s, slab, object,
1272 "Freechain corrupt");
1273 set_freepointer(s, object, NULL);
1274 } else {
1275 slab_err(s, slab, "Freepointer corrupt");
1276 slab->freelist = NULL;
1277 slab->inuse = slab->objects;
1278 slab_fix(s, "Freelist cleared");
1279 return 0;
1280 }
1281 break;
1282 }
1283 object = fp;
1284 fp = get_freepointer(s, object);
1285 nr++;
1286 }
1287
1288 max_objects = order_objects(slab_order(slab), s->size);
1289 if (max_objects > MAX_OBJS_PER_PAGE)
1290 max_objects = MAX_OBJS_PER_PAGE;
1291
1292 if (slab->objects != max_objects) {
1293 slab_err(s, slab, "Wrong number of objects. Found %d but should be %d",
1294 slab->objects, max_objects);
1295 slab->objects = max_objects;
1296 slab_fix(s, "Number of objects adjusted");
1297 }
1298 if (slab->inuse != slab->objects - nr) {
1299 slab_err(s, slab, "Wrong object count. Counter is %d but counted were %d",
1300 slab->inuse, slab->objects - nr);
1301 slab->inuse = slab->objects - nr;
1302 slab_fix(s, "Object count adjusted");
1303 }
1304 return search == NULL;
1305}
1306
1307static void trace(struct kmem_cache *s, struct slab *slab, void *object,
1308 int alloc)
1309{
1310 if (s->flags & SLAB_TRACE) {
1311 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1312 s->name,
1313 alloc ? "alloc" : "free",
1314 object, slab->inuse,
1315 slab->freelist);
1316
1317 if (!alloc)
1318 print_section(KERN_INFO, "Object ", (void *)object,
1319 s->object_size);
1320
1321 dump_stack();
1322 }
1323}
1324
1325/*
1326 * Tracking of fully allocated slabs for debugging purposes.
1327 */
1328static void add_full(struct kmem_cache *s,
1329 struct kmem_cache_node *n, struct slab *slab)
1330{
1331 if (!(s->flags & SLAB_STORE_USER))
1332 return;
1333
1334 lockdep_assert_held(&n->list_lock);
1335 list_add(&slab->slab_list, &n->full);
1336}
1337
1338static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct slab *slab)
1339{
1340 if (!(s->flags & SLAB_STORE_USER))
1341 return;
1342
1343 lockdep_assert_held(&n->list_lock);
1344 list_del(&slab->slab_list);
1345}
1346
1347/* Tracking of the number of slabs for debugging purposes */
1348static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1349{
1350 struct kmem_cache_node *n = get_node(s, node);
1351
1352 return atomic_long_read(&n->nr_slabs);
1353}
1354
1355static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1356{
1357 return atomic_long_read(&n->nr_slabs);
1358}
1359
1360static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1361{
1362 struct kmem_cache_node *n = get_node(s, node);
1363
1364 /*
1365 * May be called early in order to allocate a slab for the
1366 * kmem_cache_node structure. Solve the chicken-egg
1367 * dilemma by deferring the increment of the count during
1368 * bootstrap (see early_kmem_cache_node_alloc).
1369 */
1370 if (likely(n)) {
1371 atomic_long_inc(&n->nr_slabs);
1372 atomic_long_add(objects, &n->total_objects);
1373 }
1374}
1375static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1376{
1377 struct kmem_cache_node *n = get_node(s, node);
1378
1379 atomic_long_dec(&n->nr_slabs);
1380 atomic_long_sub(objects, &n->total_objects);
1381}
1382
1383/* Object debug checks for alloc/free paths */
1384static void setup_object_debug(struct kmem_cache *s, void *object)
1385{
1386 if (!kmem_cache_debug_flags(s, SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))
1387 return;
1388
1389 init_object(s, object, SLUB_RED_INACTIVE);
1390 init_tracking(s, object);
1391}
1392
1393static
1394void setup_slab_debug(struct kmem_cache *s, struct slab *slab, void *addr)
1395{
1396 if (!kmem_cache_debug_flags(s, SLAB_POISON))
1397 return;
1398
1399 metadata_access_enable();
1400 memset(kasan_reset_tag(addr), POISON_INUSE, slab_size(slab));
1401 metadata_access_disable();
1402}
1403
1404static inline int alloc_consistency_checks(struct kmem_cache *s,
1405 struct slab *slab, void *object)
1406{
1407 if (!check_slab(s, slab))
1408 return 0;
1409
1410 if (!check_valid_pointer(s, slab, object)) {
1411 object_err(s, slab, object, "Freelist Pointer check fails");
1412 return 0;
1413 }
1414
1415 if (!check_object(s, slab, object, SLUB_RED_INACTIVE))
1416 return 0;
1417
1418 return 1;
1419}
1420
1421static noinline bool alloc_debug_processing(struct kmem_cache *s,
1422 struct slab *slab, void *object, int orig_size)
1423{
1424 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1425 if (!alloc_consistency_checks(s, slab, object))
1426 goto bad;
1427 }
1428
1429 /* Success. Perform special debug activities for allocs */
1430 trace(s, slab, object, 1);
1431 set_orig_size(s, object, orig_size);
1432 init_object(s, object, SLUB_RED_ACTIVE);
1433 return true;
1434
1435bad:
1436 if (folio_test_slab(slab_folio(slab))) {
1437 /*
1438 * If this is a slab page then lets do the best we can
1439 * to avoid issues in the future. Marking all objects
1440 * as used avoids touching the remaining objects.
1441 */
1442 slab_fix(s, "Marking all objects used");
1443 slab->inuse = slab->objects;
1444 slab->freelist = NULL;
1445 }
1446 return false;
1447}
1448
1449static inline int free_consistency_checks(struct kmem_cache *s,
1450 struct slab *slab, void *object, unsigned long addr)
1451{
1452 if (!check_valid_pointer(s, slab, object)) {
1453 slab_err(s, slab, "Invalid object pointer 0x%p", object);
1454 return 0;
1455 }
1456
1457 if (on_freelist(s, slab, object)) {
1458 object_err(s, slab, object, "Object already free");
1459 return 0;
1460 }
1461
1462 if (!check_object(s, slab, object, SLUB_RED_ACTIVE))
1463 return 0;
1464
1465 if (unlikely(s != slab->slab_cache)) {
1466 if (!folio_test_slab(slab_folio(slab))) {
1467 slab_err(s, slab, "Attempt to free object(0x%p) outside of slab",
1468 object);
1469 } else if (!slab->slab_cache) {
1470 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1471 object);
1472 dump_stack();
1473 } else
1474 object_err(s, slab, object,
1475 "page slab pointer corrupt.");
1476 return 0;
1477 }
1478 return 1;
1479}
1480
1481/*
1482 * Parse a block of slub_debug options. Blocks are delimited by ';'
1483 *
1484 * @str: start of block
1485 * @flags: returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified
1486 * @slabs: return start of list of slabs, or NULL when there's no list
1487 * @init: assume this is initial parsing and not per-kmem-create parsing
1488 *
1489 * returns the start of next block if there's any, or NULL
1490 */
1491static char *
1492parse_slub_debug_flags(char *str, slab_flags_t *flags, char **slabs, bool init)
1493{
1494 bool higher_order_disable = false;
1495
1496 /* Skip any completely empty blocks */
1497 while (*str && *str == ';')
1498 str++;
1499
1500 if (*str == ',') {
1501 /*
1502 * No options but restriction on slabs. This means full
1503 * debugging for slabs matching a pattern.
1504 */
1505 *flags = DEBUG_DEFAULT_FLAGS;
1506 goto check_slabs;
1507 }
1508 *flags = 0;
1509
1510 /* Determine which debug features should be switched on */
1511 for (; *str && *str != ',' && *str != ';'; str++) {
1512 switch (tolower(*str)) {
1513 case '-':
1514 *flags = 0;
1515 break;
1516 case 'f':
1517 *flags |= SLAB_CONSISTENCY_CHECKS;
1518 break;
1519 case 'z':
1520 *flags |= SLAB_RED_ZONE;
1521 break;
1522 case 'p':
1523 *flags |= SLAB_POISON;
1524 break;
1525 case 'u':
1526 *flags |= SLAB_STORE_USER;
1527 break;
1528 case 't':
1529 *flags |= SLAB_TRACE;
1530 break;
1531 case 'a':
1532 *flags |= SLAB_FAILSLAB;
1533 break;
1534 case 'o':
1535 /*
1536 * Avoid enabling debugging on caches if its minimum
1537 * order would increase as a result.
1538 */
1539 higher_order_disable = true;
1540 break;
1541 default:
1542 if (init)
1543 pr_err("slub_debug option '%c' unknown. skipped\n", *str);
1544 }
1545 }
1546check_slabs:
1547 if (*str == ',')
1548 *slabs = ++str;
1549 else
1550 *slabs = NULL;
1551
1552 /* Skip over the slab list */
1553 while (*str && *str != ';')
1554 str++;
1555
1556 /* Skip any completely empty blocks */
1557 while (*str && *str == ';')
1558 str++;
1559
1560 if (init && higher_order_disable)
1561 disable_higher_order_debug = 1;
1562
1563 if (*str)
1564 return str;
1565 else
1566 return NULL;
1567}
1568
1569static int __init setup_slub_debug(char *str)
1570{
1571 slab_flags_t flags;
1572 slab_flags_t global_flags;
1573 char *saved_str;
1574 char *slab_list;
1575 bool global_slub_debug_changed = false;
1576 bool slab_list_specified = false;
1577
1578 global_flags = DEBUG_DEFAULT_FLAGS;
1579 if (*str++ != '=' || !*str)
1580 /*
1581 * No options specified. Switch on full debugging.
1582 */
1583 goto out;
1584
1585 saved_str = str;
1586 while (str) {
1587 str = parse_slub_debug_flags(str, &flags, &slab_list, true);
1588
1589 if (!slab_list) {
1590 global_flags = flags;
1591 global_slub_debug_changed = true;
1592 } else {
1593 slab_list_specified = true;
1594 if (flags & SLAB_STORE_USER)
1595 stack_depot_want_early_init();
1596 }
1597 }
1598
1599 /*
1600 * For backwards compatibility, a single list of flags with list of
1601 * slabs means debugging is only changed for those slabs, so the global
1602 * slub_debug should be unchanged (0 or DEBUG_DEFAULT_FLAGS, depending
1603 * on CONFIG_SLUB_DEBUG_ON). We can extended that to multiple lists as
1604 * long as there is no option specifying flags without a slab list.
1605 */
1606 if (slab_list_specified) {
1607 if (!global_slub_debug_changed)
1608 global_flags = slub_debug;
1609 slub_debug_string = saved_str;
1610 }
1611out:
1612 slub_debug = global_flags;
1613 if (slub_debug & SLAB_STORE_USER)
1614 stack_depot_want_early_init();
1615 if (slub_debug != 0 || slub_debug_string)
1616 static_branch_enable(&slub_debug_enabled);
1617 else
1618 static_branch_disable(&slub_debug_enabled);
1619 if ((static_branch_unlikely(&init_on_alloc) ||
1620 static_branch_unlikely(&init_on_free)) &&
1621 (slub_debug & SLAB_POISON))
1622 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1623 return 1;
1624}
1625
1626__setup("slub_debug", setup_slub_debug);
1627
1628/*
1629 * kmem_cache_flags - apply debugging options to the cache
1630 * @object_size: the size of an object without meta data
1631 * @flags: flags to set
1632 * @name: name of the cache
1633 *
1634 * Debug option(s) are applied to @flags. In addition to the debug
1635 * option(s), if a slab name (or multiple) is specified i.e.
1636 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1637 * then only the select slabs will receive the debug option(s).
1638 */
1639slab_flags_t kmem_cache_flags(unsigned int object_size,
1640 slab_flags_t flags, const char *name)
1641{
1642 char *iter;
1643 size_t len;
1644 char *next_block;
1645 slab_flags_t block_flags;
1646 slab_flags_t slub_debug_local = slub_debug;
1647
1648 if (flags & SLAB_NO_USER_FLAGS)
1649 return flags;
1650
1651 /*
1652 * If the slab cache is for debugging (e.g. kmemleak) then
1653 * don't store user (stack trace) information by default,
1654 * but let the user enable it via the command line below.
1655 */
1656 if (flags & SLAB_NOLEAKTRACE)
1657 slub_debug_local &= ~SLAB_STORE_USER;
1658
1659 len = strlen(name);
1660 next_block = slub_debug_string;
1661 /* Go through all blocks of debug options, see if any matches our slab's name */
1662 while (next_block) {
1663 next_block = parse_slub_debug_flags(next_block, &block_flags, &iter, false);
1664 if (!iter)
1665 continue;
1666 /* Found a block that has a slab list, search it */
1667 while (*iter) {
1668 char *end, *glob;
1669 size_t cmplen;
1670
1671 end = strchrnul(iter, ',');
1672 if (next_block && next_block < end)
1673 end = next_block - 1;
1674
1675 glob = strnchr(iter, end - iter, '*');
1676 if (glob)
1677 cmplen = glob - iter;
1678 else
1679 cmplen = max_t(size_t, len, (end - iter));
1680
1681 if (!strncmp(name, iter, cmplen)) {
1682 flags |= block_flags;
1683 return flags;
1684 }
1685
1686 if (!*end || *end == ';')
1687 break;
1688 iter = end + 1;
1689 }
1690 }
1691
1692 return flags | slub_debug_local;
1693}
1694#else /* !CONFIG_SLUB_DEBUG */
1695static inline void setup_object_debug(struct kmem_cache *s, void *object) {}
1696static inline
1697void setup_slab_debug(struct kmem_cache *s, struct slab *slab, void *addr) {}
1698
1699static inline bool alloc_debug_processing(struct kmem_cache *s,
1700 struct slab *slab, void *object, int orig_size) { return true; }
1701
1702static inline bool free_debug_processing(struct kmem_cache *s,
1703 struct slab *slab, void *head, void *tail, int *bulk_cnt,
1704 unsigned long addr, depot_stack_handle_t handle) { return true; }
1705
1706static inline void slab_pad_check(struct kmem_cache *s, struct slab *slab) {}
1707static inline int check_object(struct kmem_cache *s, struct slab *slab,
1708 void *object, u8 val) { return 1; }
1709static inline depot_stack_handle_t set_track_prepare(void) { return 0; }
1710static inline void set_track(struct kmem_cache *s, void *object,
1711 enum track_item alloc, unsigned long addr) {}
1712static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1713 struct slab *slab) {}
1714static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1715 struct slab *slab) {}
1716slab_flags_t kmem_cache_flags(unsigned int object_size,
1717 slab_flags_t flags, const char *name)
1718{
1719 return flags;
1720}
1721#define slub_debug 0
1722
1723#define disable_higher_order_debug 0
1724
1725static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1726 { return 0; }
1727static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1728 { return 0; }
1729static inline void inc_slabs_node(struct kmem_cache *s, int node,
1730 int objects) {}
1731static inline void dec_slabs_node(struct kmem_cache *s, int node,
1732 int objects) {}
1733
1734#ifndef CONFIG_SLUB_TINY
1735static bool freelist_corrupted(struct kmem_cache *s, struct slab *slab,
1736 void **freelist, void *nextfree)
1737{
1738 return false;
1739}
1740#endif
1741#endif /* CONFIG_SLUB_DEBUG */
1742
1743/*
1744 * Hooks for other subsystems that check memory allocations. In a typical
1745 * production configuration these hooks all should produce no code at all.
1746 */
1747static __always_inline bool slab_free_hook(struct kmem_cache *s,
1748 void *x, bool init)
1749{
1750 kmemleak_free_recursive(x, s->flags);
1751 kmsan_slab_free(s, x);
1752
1753 debug_check_no_locks_freed(x, s->object_size);
1754
1755 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1756 debug_check_no_obj_freed(x, s->object_size);
1757
1758 /* Use KCSAN to help debug racy use-after-free. */
1759 if (!(s->flags & SLAB_TYPESAFE_BY_RCU))
1760 __kcsan_check_access(x, s->object_size,
1761 KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT);
1762
1763 /*
1764 * As memory initialization might be integrated into KASAN,
1765 * kasan_slab_free and initialization memset's must be
1766 * kept together to avoid discrepancies in behavior.
1767 *
1768 * The initialization memset's clear the object and the metadata,
1769 * but don't touch the SLAB redzone.
1770 */
1771 if (init) {
1772 int rsize;
1773
1774 if (!kasan_has_integrated_init())
1775 memset(kasan_reset_tag(x), 0, s->object_size);
1776 rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad : 0;
1777 memset((char *)kasan_reset_tag(x) + s->inuse, 0,
1778 s->size - s->inuse - rsize);
1779 }
1780 /* KASAN might put x into memory quarantine, delaying its reuse. */
1781 return kasan_slab_free(s, x, init);
1782}
1783
1784static inline bool slab_free_freelist_hook(struct kmem_cache *s,
1785 void **head, void **tail,
1786 int *cnt)
1787{
1788
1789 void *object;
1790 void *next = *head;
1791 void *old_tail = *tail ? *tail : *head;
1792
1793 if (is_kfence_address(next)) {
1794 slab_free_hook(s, next, false);
1795 return true;
1796 }
1797
1798 /* Head and tail of the reconstructed freelist */
1799 *head = NULL;
1800 *tail = NULL;
1801
1802 do {
1803 object = next;
1804 next = get_freepointer(s, object);
1805
1806 /* If object's reuse doesn't have to be delayed */
1807 if (!slab_free_hook(s, object, slab_want_init_on_free(s))) {
1808 /* Move object to the new freelist */
1809 set_freepointer(s, object, *head);
1810 *head = object;
1811 if (!*tail)
1812 *tail = object;
1813 } else {
1814 /*
1815 * Adjust the reconstructed freelist depth
1816 * accordingly if object's reuse is delayed.
1817 */
1818 --(*cnt);
1819 }
1820 } while (object != old_tail);
1821
1822 if (*head == *tail)
1823 *tail = NULL;
1824
1825 return *head != NULL;
1826}
1827
1828static void *setup_object(struct kmem_cache *s, void *object)
1829{
1830 setup_object_debug(s, object);
1831 object = kasan_init_slab_obj(s, object);
1832 if (unlikely(s->ctor)) {
1833 kasan_unpoison_object_data(s, object);
1834 s->ctor(object);
1835 kasan_poison_object_data(s, object);
1836 }
1837 return object;
1838}
1839
1840/*
1841 * Slab allocation and freeing
1842 */
1843static inline struct slab *alloc_slab_page(gfp_t flags, int node,
1844 struct kmem_cache_order_objects oo)
1845{
1846 struct folio *folio;
1847 struct slab *slab;
1848 unsigned int order = oo_order(oo);
1849
1850 if (node == NUMA_NO_NODE)
1851 folio = (struct folio *)alloc_pages(flags, order);
1852 else
1853 folio = (struct folio *)__alloc_pages_node(node, flags, order);
1854
1855 if (!folio)
1856 return NULL;
1857
1858 slab = folio_slab(folio);
1859 __folio_set_slab(folio);
1860 /* Make the flag visible before any changes to folio->mapping */
1861 smp_wmb();
1862 if (page_is_pfmemalloc(folio_page(folio, 0)))
1863 slab_set_pfmemalloc(slab);
1864
1865 return slab;
1866}
1867
1868#ifdef CONFIG_SLAB_FREELIST_RANDOM
1869/* Pre-initialize the random sequence cache */
1870static int init_cache_random_seq(struct kmem_cache *s)
1871{
1872 unsigned int count = oo_objects(s->oo);
1873 int err;
1874
1875 /* Bailout if already initialised */
1876 if (s->random_seq)
1877 return 0;
1878
1879 err = cache_random_seq_create(s, count, GFP_KERNEL);
1880 if (err) {
1881 pr_err("SLUB: Unable to initialize free list for %s\n",
1882 s->name);
1883 return err;
1884 }
1885
1886 /* Transform to an offset on the set of pages */
1887 if (s->random_seq) {
1888 unsigned int i;
1889
1890 for (i = 0; i < count; i++)
1891 s->random_seq[i] *= s->size;
1892 }
1893 return 0;
1894}
1895
1896/* Initialize each random sequence freelist per cache */
1897static void __init init_freelist_randomization(void)
1898{
1899 struct kmem_cache *s;
1900
1901 mutex_lock(&slab_mutex);
1902
1903 list_for_each_entry(s, &slab_caches, list)
1904 init_cache_random_seq(s);
1905
1906 mutex_unlock(&slab_mutex);
1907}
1908
1909/* Get the next entry on the pre-computed freelist randomized */
1910static void *next_freelist_entry(struct kmem_cache *s, struct slab *slab,
1911 unsigned long *pos, void *start,
1912 unsigned long page_limit,
1913 unsigned long freelist_count)
1914{
1915 unsigned int idx;
1916
1917 /*
1918 * If the target page allocation failed, the number of objects on the
1919 * page might be smaller than the usual size defined by the cache.
1920 */
1921 do {
1922 idx = s->random_seq[*pos];
1923 *pos += 1;
1924 if (*pos >= freelist_count)
1925 *pos = 0;
1926 } while (unlikely(idx >= page_limit));
1927
1928 return (char *)start + idx;
1929}
1930
1931/* Shuffle the single linked freelist based on a random pre-computed sequence */
1932static bool shuffle_freelist(struct kmem_cache *s, struct slab *slab)
1933{
1934 void *start;
1935 void *cur;
1936 void *next;
1937 unsigned long idx, pos, page_limit, freelist_count;
1938
1939 if (slab->objects < 2 || !s->random_seq)
1940 return false;
1941
1942 freelist_count = oo_objects(s->oo);
1943 pos = get_random_u32_below(freelist_count);
1944
1945 page_limit = slab->objects * s->size;
1946 start = fixup_red_left(s, slab_address(slab));
1947
1948 /* First entry is used as the base of the freelist */
1949 cur = next_freelist_entry(s, slab, &pos, start, page_limit,
1950 freelist_count);
1951 cur = setup_object(s, cur);
1952 slab->freelist = cur;
1953
1954 for (idx = 1; idx < slab->objects; idx++) {
1955 next = next_freelist_entry(s, slab, &pos, start, page_limit,
1956 freelist_count);
1957 next = setup_object(s, next);
1958 set_freepointer(s, cur, next);
1959 cur = next;
1960 }
1961 set_freepointer(s, cur, NULL);
1962
1963 return true;
1964}
1965#else
1966static inline int init_cache_random_seq(struct kmem_cache *s)
1967{
1968 return 0;
1969}
1970static inline void init_freelist_randomization(void) { }
1971static inline bool shuffle_freelist(struct kmem_cache *s, struct slab *slab)
1972{
1973 return false;
1974}
1975#endif /* CONFIG_SLAB_FREELIST_RANDOM */
1976
1977static struct slab *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1978{
1979 struct slab *slab;
1980 struct kmem_cache_order_objects oo = s->oo;
1981 gfp_t alloc_gfp;
1982 void *start, *p, *next;
1983 int idx;
1984 bool shuffle;
1985
1986 flags &= gfp_allowed_mask;
1987
1988 flags |= s->allocflags;
1989
1990 /*
1991 * Let the initial higher-order allocation fail under memory pressure
1992 * so we fall-back to the minimum order allocation.
1993 */
1994 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1995 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1996 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~__GFP_RECLAIM;
1997
1998 slab = alloc_slab_page(alloc_gfp, node, oo);
1999 if (unlikely(!slab)) {
2000 oo = s->min;
2001 alloc_gfp = flags;
2002 /*
2003 * Allocation may have failed due to fragmentation.
2004 * Try a lower order alloc if possible
2005 */
2006 slab = alloc_slab_page(alloc_gfp, node, oo);
2007 if (unlikely(!slab))
2008 return NULL;
2009 stat(s, ORDER_FALLBACK);
2010 }
2011
2012 slab->objects = oo_objects(oo);
2013 slab->inuse = 0;
2014 slab->frozen = 0;
2015
2016 account_slab(slab, oo_order(oo), s, flags);
2017
2018 slab->slab_cache = s;
2019
2020 kasan_poison_slab(slab);
2021
2022 start = slab_address(slab);
2023
2024 setup_slab_debug(s, slab, start);
2025
2026 shuffle = shuffle_freelist(s, slab);
2027
2028 if (!shuffle) {
2029 start = fixup_red_left(s, start);
2030 start = setup_object(s, start);
2031 slab->freelist = start;
2032 for (idx = 0, p = start; idx < slab->objects - 1; idx++) {
2033 next = p + s->size;
2034 next = setup_object(s, next);
2035 set_freepointer(s, p, next);
2036 p = next;
2037 }
2038 set_freepointer(s, p, NULL);
2039 }
2040
2041 return slab;
2042}
2043
2044static struct slab *new_slab(struct kmem_cache *s, gfp_t flags, int node)
2045{
2046 if (unlikely(flags & GFP_SLAB_BUG_MASK))
2047 flags = kmalloc_fix_flags(flags);
2048
2049 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
2050
2051 return allocate_slab(s,
2052 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
2053}
2054
2055static void __free_slab(struct kmem_cache *s, struct slab *slab)
2056{
2057 struct folio *folio = slab_folio(slab);
2058 int order = folio_order(folio);
2059 int pages = 1 << order;
2060
2061 __slab_clear_pfmemalloc(slab);
2062 folio->mapping = NULL;
2063 /* Make the mapping reset visible before clearing the flag */
2064 smp_wmb();
2065 __folio_clear_slab(folio);
2066 if (current->reclaim_state)
2067 current->reclaim_state->reclaimed_slab += pages;
2068 unaccount_slab(slab, order, s);
2069 __free_pages(folio_page(folio, 0), order);
2070}
2071
2072static void rcu_free_slab(struct rcu_head *h)
2073{
2074 struct slab *slab = container_of(h, struct slab, rcu_head);
2075
2076 __free_slab(slab->slab_cache, slab);
2077}
2078
2079static void free_slab(struct kmem_cache *s, struct slab *slab)
2080{
2081 if (kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS)) {
2082 void *p;
2083
2084 slab_pad_check(s, slab);
2085 for_each_object(p, s, slab_address(slab), slab->objects)
2086 check_object(s, slab, p, SLUB_RED_INACTIVE);
2087 }
2088
2089 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU))
2090 call_rcu(&slab->rcu_head, rcu_free_slab);
2091 else
2092 __free_slab(s, slab);
2093}
2094
2095static void discard_slab(struct kmem_cache *s, struct slab *slab)
2096{
2097 dec_slabs_node(s, slab_nid(slab), slab->objects);
2098 free_slab(s, slab);
2099}
2100
2101/*
2102 * Management of partially allocated slabs.
2103 */
2104static inline void
2105__add_partial(struct kmem_cache_node *n, struct slab *slab, int tail)
2106{
2107 n->nr_partial++;
2108 if (tail == DEACTIVATE_TO_TAIL)
2109 list_add_tail(&slab->slab_list, &n->partial);
2110 else
2111 list_add(&slab->slab_list, &n->partial);
2112}
2113
2114static inline void add_partial(struct kmem_cache_node *n,
2115 struct slab *slab, int tail)
2116{
2117 lockdep_assert_held(&n->list_lock);
2118 __add_partial(n, slab, tail);
2119}
2120
2121static inline void remove_partial(struct kmem_cache_node *n,
2122 struct slab *slab)
2123{
2124 lockdep_assert_held(&n->list_lock);
2125 list_del(&slab->slab_list);
2126 n->nr_partial--;
2127}
2128
2129/*
2130 * Called only for kmem_cache_debug() caches instead of acquire_slab(), with a
2131 * slab from the n->partial list. Remove only a single object from the slab, do
2132 * the alloc_debug_processing() checks and leave the slab on the list, or move
2133 * it to full list if it was the last free object.
2134 */
2135static void *alloc_single_from_partial(struct kmem_cache *s,
2136 struct kmem_cache_node *n, struct slab *slab, int orig_size)
2137{
2138 void *object;
2139
2140 lockdep_assert_held(&n->list_lock);
2141
2142 object = slab->freelist;
2143 slab->freelist = get_freepointer(s, object);
2144 slab->inuse++;
2145
2146 if (!alloc_debug_processing(s, slab, object, orig_size)) {
2147 remove_partial(n, slab);
2148 return NULL;
2149 }
2150
2151 if (slab->inuse == slab->objects) {
2152 remove_partial(n, slab);
2153 add_full(s, n, slab);
2154 }
2155
2156 return object;
2157}
2158
2159/*
2160 * Called only for kmem_cache_debug() caches to allocate from a freshly
2161 * allocated slab. Allocate a single object instead of whole freelist
2162 * and put the slab to the partial (or full) list.
2163 */
2164static void *alloc_single_from_new_slab(struct kmem_cache *s,
2165 struct slab *slab, int orig_size)
2166{
2167 int nid = slab_nid(slab);
2168 struct kmem_cache_node *n = get_node(s, nid);
2169 unsigned long flags;
2170 void *object;
2171
2172
2173 object = slab->freelist;
2174 slab->freelist = get_freepointer(s, object);
2175 slab->inuse = 1;
2176
2177 if (!alloc_debug_processing(s, slab, object, orig_size))
2178 /*
2179 * It's not really expected that this would fail on a
2180 * freshly allocated slab, but a concurrent memory
2181 * corruption in theory could cause that.
2182 */
2183 return NULL;
2184
2185 spin_lock_irqsave(&n->list_lock, flags);
2186
2187 if (slab->inuse == slab->objects)
2188 add_full(s, n, slab);
2189 else
2190 add_partial(n, slab, DEACTIVATE_TO_HEAD);
2191
2192 inc_slabs_node(s, nid, slab->objects);
2193 spin_unlock_irqrestore(&n->list_lock, flags);
2194
2195 return object;
2196}
2197
2198/*
2199 * Remove slab from the partial list, freeze it and
2200 * return the pointer to the freelist.
2201 *
2202 * Returns a list of objects or NULL if it fails.
2203 */
2204static inline void *acquire_slab(struct kmem_cache *s,
2205 struct kmem_cache_node *n, struct slab *slab,
2206 int mode)
2207{
2208 void *freelist;
2209 unsigned long counters;
2210 struct slab new;
2211
2212 lockdep_assert_held(&n->list_lock);
2213
2214 /*
2215 * Zap the freelist and set the frozen bit.
2216 * The old freelist is the list of objects for the
2217 * per cpu allocation list.
2218 */
2219 freelist = slab->freelist;
2220 counters = slab->counters;
2221 new.counters = counters;
2222 if (mode) {
2223 new.inuse = slab->objects;
2224 new.freelist = NULL;
2225 } else {
2226 new.freelist = freelist;
2227 }
2228
2229 VM_BUG_ON(new.frozen);
2230 new.frozen = 1;
2231
2232 if (!__cmpxchg_double_slab(s, slab,
2233 freelist, counters,
2234 new.freelist, new.counters,
2235 "acquire_slab"))
2236 return NULL;
2237
2238 remove_partial(n, slab);
2239 WARN_ON(!freelist);
2240 return freelist;
2241}
2242
2243#ifdef CONFIG_SLUB_CPU_PARTIAL
2244static void put_cpu_partial(struct kmem_cache *s, struct slab *slab, int drain);
2245#else
2246static inline void put_cpu_partial(struct kmem_cache *s, struct slab *slab,
2247 int drain) { }
2248#endif
2249static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags);
2250
2251/*
2252 * Try to allocate a partial slab from a specific node.
2253 */
2254static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
2255 struct partial_context *pc)
2256{
2257 struct slab *slab, *slab2;
2258 void *object = NULL;
2259 unsigned long flags;
2260 unsigned int partial_slabs = 0;
2261
2262 /*
2263 * Racy check. If we mistakenly see no partial slabs then we
2264 * just allocate an empty slab. If we mistakenly try to get a
2265 * partial slab and there is none available then get_partial()
2266 * will return NULL.
2267 */
2268 if (!n || !n->nr_partial)
2269 return NULL;
2270
2271 spin_lock_irqsave(&n->list_lock, flags);
2272 list_for_each_entry_safe(slab, slab2, &n->partial, slab_list) {
2273 void *t;
2274
2275 if (!pfmemalloc_match(slab, pc->flags))
2276 continue;
2277
2278 if (IS_ENABLED(CONFIG_SLUB_TINY) || kmem_cache_debug(s)) {
2279 object = alloc_single_from_partial(s, n, slab,
2280 pc->orig_size);
2281 if (object)
2282 break;
2283 continue;
2284 }
2285
2286 t = acquire_slab(s, n, slab, object == NULL);
2287 if (!t)
2288 break;
2289
2290 if (!object) {
2291 *pc->slab = slab;
2292 stat(s, ALLOC_FROM_PARTIAL);
2293 object = t;
2294 } else {
2295 put_cpu_partial(s, slab, 0);
2296 stat(s, CPU_PARTIAL_NODE);
2297 partial_slabs++;
2298 }
2299#ifdef CONFIG_SLUB_CPU_PARTIAL
2300 if (!kmem_cache_has_cpu_partial(s)
2301 || partial_slabs > s->cpu_partial_slabs / 2)
2302 break;
2303#else
2304 break;
2305#endif
2306
2307 }
2308 spin_unlock_irqrestore(&n->list_lock, flags);
2309 return object;
2310}
2311
2312/*
2313 * Get a slab from somewhere. Search in increasing NUMA distances.
2314 */
2315static void *get_any_partial(struct kmem_cache *s, struct partial_context *pc)
2316{
2317#ifdef CONFIG_NUMA
2318 struct zonelist *zonelist;
2319 struct zoneref *z;
2320 struct zone *zone;
2321 enum zone_type highest_zoneidx = gfp_zone(pc->flags);
2322 void *object;
2323 unsigned int cpuset_mems_cookie;
2324
2325 /*
2326 * The defrag ratio allows a configuration of the tradeoffs between
2327 * inter node defragmentation and node local allocations. A lower
2328 * defrag_ratio increases the tendency to do local allocations
2329 * instead of attempting to obtain partial slabs from other nodes.
2330 *
2331 * If the defrag_ratio is set to 0 then kmalloc() always
2332 * returns node local objects. If the ratio is higher then kmalloc()
2333 * may return off node objects because partial slabs are obtained
2334 * from other nodes and filled up.
2335 *
2336 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
2337 * (which makes defrag_ratio = 1000) then every (well almost)
2338 * allocation will first attempt to defrag slab caches on other nodes.
2339 * This means scanning over all nodes to look for partial slabs which
2340 * may be expensive if we do it every time we are trying to find a slab
2341 * with available objects.
2342 */
2343 if (!s->remote_node_defrag_ratio ||
2344 get_cycles() % 1024 > s->remote_node_defrag_ratio)
2345 return NULL;
2346
2347 do {
2348 cpuset_mems_cookie = read_mems_allowed_begin();
2349 zonelist = node_zonelist(mempolicy_slab_node(), pc->flags);
2350 for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) {
2351 struct kmem_cache_node *n;
2352
2353 n = get_node(s, zone_to_nid(zone));
2354
2355 if (n && cpuset_zone_allowed(zone, pc->flags) &&
2356 n->nr_partial > s->min_partial) {
2357 object = get_partial_node(s, n, pc);
2358 if (object) {
2359 /*
2360 * Don't check read_mems_allowed_retry()
2361 * here - if mems_allowed was updated in
2362 * parallel, that was a harmless race
2363 * between allocation and the cpuset
2364 * update
2365 */
2366 return object;
2367 }
2368 }
2369 }
2370 } while (read_mems_allowed_retry(cpuset_mems_cookie));
2371#endif /* CONFIG_NUMA */
2372 return NULL;
2373}
2374
2375/*
2376 * Get a partial slab, lock it and return it.
2377 */
2378static void *get_partial(struct kmem_cache *s, int node, struct partial_context *pc)
2379{
2380 void *object;
2381 int searchnode = node;
2382
2383 if (node == NUMA_NO_NODE)
2384 searchnode = numa_mem_id();
2385
2386 object = get_partial_node(s, get_node(s, searchnode), pc);
2387 if (object || node != NUMA_NO_NODE)
2388 return object;
2389
2390 return get_any_partial(s, pc);
2391}
2392
2393#ifndef CONFIG_SLUB_TINY
2394
2395#ifdef CONFIG_PREEMPTION
2396/*
2397 * Calculate the next globally unique transaction for disambiguation
2398 * during cmpxchg. The transactions start with the cpu number and are then
2399 * incremented by CONFIG_NR_CPUS.
2400 */
2401#define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
2402#else
2403/*
2404 * No preemption supported therefore also no need to check for
2405 * different cpus.
2406 */
2407#define TID_STEP 1
2408#endif /* CONFIG_PREEMPTION */
2409
2410static inline unsigned long next_tid(unsigned long tid)
2411{
2412 return tid + TID_STEP;
2413}
2414
2415#ifdef SLUB_DEBUG_CMPXCHG
2416static inline unsigned int tid_to_cpu(unsigned long tid)
2417{
2418 return tid % TID_STEP;
2419}
2420
2421static inline unsigned long tid_to_event(unsigned long tid)
2422{
2423 return tid / TID_STEP;
2424}
2425#endif
2426
2427static inline unsigned int init_tid(int cpu)
2428{
2429 return cpu;
2430}
2431
2432static inline void note_cmpxchg_failure(const char *n,
2433 const struct kmem_cache *s, unsigned long tid)
2434{
2435#ifdef SLUB_DEBUG_CMPXCHG
2436 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
2437
2438 pr_info("%s %s: cmpxchg redo ", n, s->name);
2439
2440#ifdef CONFIG_PREEMPTION
2441 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
2442 pr_warn("due to cpu change %d -> %d\n",
2443 tid_to_cpu(tid), tid_to_cpu(actual_tid));
2444 else
2445#endif
2446 if (tid_to_event(tid) != tid_to_event(actual_tid))
2447 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2448 tid_to_event(tid), tid_to_event(actual_tid));
2449 else
2450 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2451 actual_tid, tid, next_tid(tid));
2452#endif
2453 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2454}
2455
2456static void init_kmem_cache_cpus(struct kmem_cache *s)
2457{
2458 int cpu;
2459 struct kmem_cache_cpu *c;
2460
2461 for_each_possible_cpu(cpu) {
2462 c = per_cpu_ptr(s->cpu_slab, cpu);
2463 local_lock_init(&c->lock);
2464 c->tid = init_tid(cpu);
2465 }
2466}
2467
2468/*
2469 * Finishes removing the cpu slab. Merges cpu's freelist with slab's freelist,
2470 * unfreezes the slabs and puts it on the proper list.
2471 * Assumes the slab has been already safely taken away from kmem_cache_cpu
2472 * by the caller.
2473 */
2474static void deactivate_slab(struct kmem_cache *s, struct slab *slab,
2475 void *freelist)
2476{
2477 enum slab_modes { M_NONE, M_PARTIAL, M_FREE, M_FULL_NOLIST };
2478 struct kmem_cache_node *n = get_node(s, slab_nid(slab));
2479 int free_delta = 0;
2480 enum slab_modes mode = M_NONE;
2481 void *nextfree, *freelist_iter, *freelist_tail;
2482 int tail = DEACTIVATE_TO_HEAD;
2483 unsigned long flags = 0;
2484 struct slab new;
2485 struct slab old;
2486
2487 if (slab->freelist) {
2488 stat(s, DEACTIVATE_REMOTE_FREES);
2489 tail = DEACTIVATE_TO_TAIL;
2490 }
2491
2492 /*
2493 * Stage one: Count the objects on cpu's freelist as free_delta and
2494 * remember the last object in freelist_tail for later splicing.
2495 */
2496 freelist_tail = NULL;
2497 freelist_iter = freelist;
2498 while (freelist_iter) {
2499 nextfree = get_freepointer(s, freelist_iter);
2500
2501 /*
2502 * If 'nextfree' is invalid, it is possible that the object at
2503 * 'freelist_iter' is already corrupted. So isolate all objects
2504 * starting at 'freelist_iter' by skipping them.
2505 */
2506 if (freelist_corrupted(s, slab, &freelist_iter, nextfree))
2507 break;
2508
2509 freelist_tail = freelist_iter;
2510 free_delta++;
2511
2512 freelist_iter = nextfree;
2513 }
2514
2515 /*
2516 * Stage two: Unfreeze the slab while splicing the per-cpu
2517 * freelist to the head of slab's freelist.
2518 *
2519 * Ensure that the slab is unfrozen while the list presence
2520 * reflects the actual number of objects during unfreeze.
2521 *
2522 * We first perform cmpxchg holding lock and insert to list
2523 * when it succeed. If there is mismatch then the slab is not
2524 * unfrozen and number of objects in the slab may have changed.
2525 * Then release lock and retry cmpxchg again.
2526 */
2527redo:
2528
2529 old.freelist = READ_ONCE(slab->freelist);
2530 old.counters = READ_ONCE(slab->counters);
2531 VM_BUG_ON(!old.frozen);
2532
2533 /* Determine target state of the slab */
2534 new.counters = old.counters;
2535 if (freelist_tail) {
2536 new.inuse -= free_delta;
2537 set_freepointer(s, freelist_tail, old.freelist);
2538 new.freelist = freelist;
2539 } else
2540 new.freelist = old.freelist;
2541
2542 new.frozen = 0;
2543
2544 if (!new.inuse && n->nr_partial >= s->min_partial) {
2545 mode = M_FREE;
2546 } else if (new.freelist) {
2547 mode = M_PARTIAL;
2548 /*
2549 * Taking the spinlock removes the possibility that
2550 * acquire_slab() will see a slab that is frozen
2551 */
2552 spin_lock_irqsave(&n->list_lock, flags);
2553 } else {
2554 mode = M_FULL_NOLIST;
2555 }
2556
2557
2558 if (!cmpxchg_double_slab(s, slab,
2559 old.freelist, old.counters,
2560 new.freelist, new.counters,
2561 "unfreezing slab")) {
2562 if (mode == M_PARTIAL)
2563 spin_unlock_irqrestore(&n->list_lock, flags);
2564 goto redo;
2565 }
2566
2567
2568 if (mode == M_PARTIAL) {
2569 add_partial(n, slab, tail);
2570 spin_unlock_irqrestore(&n->list_lock, flags);
2571 stat(s, tail);
2572 } else if (mode == M_FREE) {
2573 stat(s, DEACTIVATE_EMPTY);
2574 discard_slab(s, slab);
2575 stat(s, FREE_SLAB);
2576 } else if (mode == M_FULL_NOLIST) {
2577 stat(s, DEACTIVATE_FULL);
2578 }
2579}
2580
2581#ifdef CONFIG_SLUB_CPU_PARTIAL
2582static void __unfreeze_partials(struct kmem_cache *s, struct slab *partial_slab)
2583{
2584 struct kmem_cache_node *n = NULL, *n2 = NULL;
2585 struct slab *slab, *slab_to_discard = NULL;
2586 unsigned long flags = 0;
2587
2588 while (partial_slab) {
2589 struct slab new;
2590 struct slab old;
2591
2592 slab = partial_slab;
2593 partial_slab = slab->next;
2594
2595 n2 = get_node(s, slab_nid(slab));
2596 if (n != n2) {
2597 if (n)
2598 spin_unlock_irqrestore(&n->list_lock, flags);
2599
2600 n = n2;
2601 spin_lock_irqsave(&n->list_lock, flags);
2602 }
2603
2604 do {
2605
2606 old.freelist = slab->freelist;
2607 old.counters = slab->counters;
2608 VM_BUG_ON(!old.frozen);
2609
2610 new.counters = old.counters;
2611 new.freelist = old.freelist;
2612
2613 new.frozen = 0;
2614
2615 } while (!__cmpxchg_double_slab(s, slab,
2616 old.freelist, old.counters,
2617 new.freelist, new.counters,
2618 "unfreezing slab"));
2619
2620 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2621 slab->next = slab_to_discard;
2622 slab_to_discard = slab;
2623 } else {
2624 add_partial(n, slab, DEACTIVATE_TO_TAIL);
2625 stat(s, FREE_ADD_PARTIAL);
2626 }
2627 }
2628
2629 if (n)
2630 spin_unlock_irqrestore(&n->list_lock, flags);
2631
2632 while (slab_to_discard) {
2633 slab = slab_to_discard;
2634 slab_to_discard = slab_to_discard->next;
2635
2636 stat(s, DEACTIVATE_EMPTY);
2637 discard_slab(s, slab);
2638 stat(s, FREE_SLAB);
2639 }
2640}
2641
2642/*
2643 * Unfreeze all the cpu partial slabs.
2644 */
2645static void unfreeze_partials(struct kmem_cache *s)
2646{
2647 struct slab *partial_slab;
2648 unsigned long flags;
2649
2650 local_lock_irqsave(&s->cpu_slab->lock, flags);
2651 partial_slab = this_cpu_read(s->cpu_slab->partial);
2652 this_cpu_write(s->cpu_slab->partial, NULL);
2653 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2654
2655 if (partial_slab)
2656 __unfreeze_partials(s, partial_slab);
2657}
2658
2659static void unfreeze_partials_cpu(struct kmem_cache *s,
2660 struct kmem_cache_cpu *c)
2661{
2662 struct slab *partial_slab;
2663
2664 partial_slab = slub_percpu_partial(c);
2665 c->partial = NULL;
2666
2667 if (partial_slab)
2668 __unfreeze_partials(s, partial_slab);
2669}
2670
2671/*
2672 * Put a slab that was just frozen (in __slab_free|get_partial_node) into a
2673 * partial slab slot if available.
2674 *
2675 * If we did not find a slot then simply move all the partials to the
2676 * per node partial list.
2677 */
2678static void put_cpu_partial(struct kmem_cache *s, struct slab *slab, int drain)
2679{
2680 struct slab *oldslab;
2681 struct slab *slab_to_unfreeze = NULL;
2682 unsigned long flags;
2683 int slabs = 0;
2684
2685 local_lock_irqsave(&s->cpu_slab->lock, flags);
2686
2687 oldslab = this_cpu_read(s->cpu_slab->partial);
2688
2689 if (oldslab) {
2690 if (drain && oldslab->slabs >= s->cpu_partial_slabs) {
2691 /*
2692 * Partial array is full. Move the existing set to the
2693 * per node partial list. Postpone the actual unfreezing
2694 * outside of the critical section.
2695 */
2696 slab_to_unfreeze = oldslab;
2697 oldslab = NULL;
2698 } else {
2699 slabs = oldslab->slabs;
2700 }
2701 }
2702
2703 slabs++;
2704
2705 slab->slabs = slabs;
2706 slab->next = oldslab;
2707
2708 this_cpu_write(s->cpu_slab->partial, slab);
2709
2710 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2711
2712 if (slab_to_unfreeze) {
2713 __unfreeze_partials(s, slab_to_unfreeze);
2714 stat(s, CPU_PARTIAL_DRAIN);
2715 }
2716}
2717
2718#else /* CONFIG_SLUB_CPU_PARTIAL */
2719
2720static inline void unfreeze_partials(struct kmem_cache *s) { }
2721static inline void unfreeze_partials_cpu(struct kmem_cache *s,
2722 struct kmem_cache_cpu *c) { }
2723
2724#endif /* CONFIG_SLUB_CPU_PARTIAL */
2725
2726static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2727{
2728 unsigned long flags;
2729 struct slab *slab;
2730 void *freelist;
2731
2732 local_lock_irqsave(&s->cpu_slab->lock, flags);
2733
2734 slab = c->slab;
2735 freelist = c->freelist;
2736
2737 c->slab = NULL;
2738 c->freelist = NULL;
2739 c->tid = next_tid(c->tid);
2740
2741 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2742
2743 if (slab) {
2744 deactivate_slab(s, slab, freelist);
2745 stat(s, CPUSLAB_FLUSH);
2746 }
2747}
2748
2749static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2750{
2751 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2752 void *freelist = c->freelist;
2753 struct slab *slab = c->slab;
2754
2755 c->slab = NULL;
2756 c->freelist = NULL;
2757 c->tid = next_tid(c->tid);
2758
2759 if (slab) {
2760 deactivate_slab(s, slab, freelist);
2761 stat(s, CPUSLAB_FLUSH);
2762 }
2763
2764 unfreeze_partials_cpu(s, c);
2765}
2766
2767struct slub_flush_work {
2768 struct work_struct work;
2769 struct kmem_cache *s;
2770 bool skip;
2771};
2772
2773/*
2774 * Flush cpu slab.
2775 *
2776 * Called from CPU work handler with migration disabled.
2777 */
2778static void flush_cpu_slab(struct work_struct *w)
2779{
2780 struct kmem_cache *s;
2781 struct kmem_cache_cpu *c;
2782 struct slub_flush_work *sfw;
2783
2784 sfw = container_of(w, struct slub_flush_work, work);
2785
2786 s = sfw->s;
2787 c = this_cpu_ptr(s->cpu_slab);
2788
2789 if (c->slab)
2790 flush_slab(s, c);
2791
2792 unfreeze_partials(s);
2793}
2794
2795static bool has_cpu_slab(int cpu, struct kmem_cache *s)
2796{
2797 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2798
2799 return c->slab || slub_percpu_partial(c);
2800}
2801
2802static DEFINE_MUTEX(flush_lock);
2803static DEFINE_PER_CPU(struct slub_flush_work, slub_flush);
2804
2805static void flush_all_cpus_locked(struct kmem_cache *s)
2806{
2807 struct slub_flush_work *sfw;
2808 unsigned int cpu;
2809
2810 lockdep_assert_cpus_held();
2811 mutex_lock(&flush_lock);
2812
2813 for_each_online_cpu(cpu) {
2814 sfw = &per_cpu(slub_flush, cpu);
2815 if (!has_cpu_slab(cpu, s)) {
2816 sfw->skip = true;
2817 continue;
2818 }
2819 INIT_WORK(&sfw->work, flush_cpu_slab);
2820 sfw->skip = false;
2821 sfw->s = s;
2822 queue_work_on(cpu, flushwq, &sfw->work);
2823 }
2824
2825 for_each_online_cpu(cpu) {
2826 sfw = &per_cpu(slub_flush, cpu);
2827 if (sfw->skip)
2828 continue;
2829 flush_work(&sfw->work);
2830 }
2831
2832 mutex_unlock(&flush_lock);
2833}
2834
2835static void flush_all(struct kmem_cache *s)
2836{
2837 cpus_read_lock();
2838 flush_all_cpus_locked(s);
2839 cpus_read_unlock();
2840}
2841
2842/*
2843 * Use the cpu notifier to insure that the cpu slabs are flushed when
2844 * necessary.
2845 */
2846static int slub_cpu_dead(unsigned int cpu)
2847{
2848 struct kmem_cache *s;
2849
2850 mutex_lock(&slab_mutex);
2851 list_for_each_entry(s, &slab_caches, list)
2852 __flush_cpu_slab(s, cpu);
2853 mutex_unlock(&slab_mutex);
2854 return 0;
2855}
2856
2857#else /* CONFIG_SLUB_TINY */
2858static inline void flush_all_cpus_locked(struct kmem_cache *s) { }
2859static inline void flush_all(struct kmem_cache *s) { }
2860static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) { }
2861static inline int slub_cpu_dead(unsigned int cpu) { return 0; }
2862#endif /* CONFIG_SLUB_TINY */
2863
2864/*
2865 * Check if the objects in a per cpu structure fit numa
2866 * locality expectations.
2867 */
2868static inline int node_match(struct slab *slab, int node)
2869{
2870#ifdef CONFIG_NUMA
2871 if (node != NUMA_NO_NODE && slab_nid(slab) != node)
2872 return 0;
2873#endif
2874 return 1;
2875}
2876
2877#ifdef CONFIG_SLUB_DEBUG
2878static int count_free(struct slab *slab)
2879{
2880 return slab->objects - slab->inuse;
2881}
2882
2883static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2884{
2885 return atomic_long_read(&n->total_objects);
2886}
2887
2888/* Supports checking bulk free of a constructed freelist */
2889static inline bool free_debug_processing(struct kmem_cache *s,
2890 struct slab *slab, void *head, void *tail, int *bulk_cnt,
2891 unsigned long addr, depot_stack_handle_t handle)
2892{
2893 bool checks_ok = false;
2894 void *object = head;
2895 int cnt = 0;
2896
2897 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
2898 if (!check_slab(s, slab))
2899 goto out;
2900 }
2901
2902 if (slab->inuse < *bulk_cnt) {
2903 slab_err(s, slab, "Slab has %d allocated objects but %d are to be freed\n",
2904 slab->inuse, *bulk_cnt);
2905 goto out;
2906 }
2907
2908next_object:
2909
2910 if (++cnt > *bulk_cnt)
2911 goto out_cnt;
2912
2913 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
2914 if (!free_consistency_checks(s, slab, object, addr))
2915 goto out;
2916 }
2917
2918 if (s->flags & SLAB_STORE_USER)
2919 set_track_update(s, object, TRACK_FREE, addr, handle);
2920 trace(s, slab, object, 0);
2921 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
2922 init_object(s, object, SLUB_RED_INACTIVE);
2923
2924 /* Reached end of constructed freelist yet? */
2925 if (object != tail) {
2926 object = get_freepointer(s, object);
2927 goto next_object;
2928 }
2929 checks_ok = true;
2930
2931out_cnt:
2932 if (cnt != *bulk_cnt) {
2933 slab_err(s, slab, "Bulk free expected %d objects but found %d\n",
2934 *bulk_cnt, cnt);
2935 *bulk_cnt = cnt;
2936 }
2937
2938out:
2939
2940 if (!checks_ok)
2941 slab_fix(s, "Object at 0x%p not freed", object);
2942
2943 return checks_ok;
2944}
2945#endif /* CONFIG_SLUB_DEBUG */
2946
2947#if defined(CONFIG_SLUB_DEBUG) || defined(SLAB_SUPPORTS_SYSFS)
2948static unsigned long count_partial(struct kmem_cache_node *n,
2949 int (*get_count)(struct slab *))
2950{
2951 unsigned long flags;
2952 unsigned long x = 0;
2953 struct slab *slab;
2954
2955 spin_lock_irqsave(&n->list_lock, flags);
2956 list_for_each_entry(slab, &n->partial, slab_list)
2957 x += get_count(slab);
2958 spin_unlock_irqrestore(&n->list_lock, flags);
2959 return x;
2960}
2961#endif /* CONFIG_SLUB_DEBUG || SLAB_SUPPORTS_SYSFS */
2962
2963#ifdef CONFIG_SLUB_DEBUG
2964static noinline void
2965slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2966{
2967 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2968 DEFAULT_RATELIMIT_BURST);
2969 int node;
2970 struct kmem_cache_node *n;
2971
2972 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2973 return;
2974
2975 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2976 nid, gfpflags, &gfpflags);
2977 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2978 s->name, s->object_size, s->size, oo_order(s->oo),
2979 oo_order(s->min));
2980
2981 if (oo_order(s->min) > get_order(s->object_size))
2982 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2983 s->name);
2984
2985 for_each_kmem_cache_node(s, node, n) {
2986 unsigned long nr_slabs;
2987 unsigned long nr_objs;
2988 unsigned long nr_free;
2989
2990 nr_free = count_partial(n, count_free);
2991 nr_slabs = node_nr_slabs(n);
2992 nr_objs = node_nr_objs(n);
2993
2994 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2995 node, nr_slabs, nr_objs, nr_free);
2996 }
2997}
2998#else /* CONFIG_SLUB_DEBUG */
2999static inline void
3000slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid) { }
3001#endif
3002
3003static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags)
3004{
3005 if (unlikely(slab_test_pfmemalloc(slab)))
3006 return gfp_pfmemalloc_allowed(gfpflags);
3007
3008 return true;
3009}
3010
3011#ifndef CONFIG_SLUB_TINY
3012/*
3013 * Check the slab->freelist and either transfer the freelist to the
3014 * per cpu freelist or deactivate the slab.
3015 *
3016 * The slab is still frozen if the return value is not NULL.
3017 *
3018 * If this function returns NULL then the slab has been unfrozen.
3019 */
3020static inline void *get_freelist(struct kmem_cache *s, struct slab *slab)
3021{
3022 struct slab new;
3023 unsigned long counters;
3024 void *freelist;
3025
3026 lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock));
3027
3028 do {
3029 freelist = slab->freelist;
3030 counters = slab->counters;
3031
3032 new.counters = counters;
3033 VM_BUG_ON(!new.frozen);
3034
3035 new.inuse = slab->objects;
3036 new.frozen = freelist != NULL;
3037
3038 } while (!__cmpxchg_double_slab(s, slab,
3039 freelist, counters,
3040 NULL, new.counters,
3041 "get_freelist"));
3042
3043 return freelist;
3044}
3045
3046/*
3047 * Slow path. The lockless freelist is empty or we need to perform
3048 * debugging duties.
3049 *
3050 * Processing is still very fast if new objects have been freed to the
3051 * regular freelist. In that case we simply take over the regular freelist
3052 * as the lockless freelist and zap the regular freelist.
3053 *
3054 * If that is not working then we fall back to the partial lists. We take the
3055 * first element of the freelist as the object to allocate now and move the
3056 * rest of the freelist to the lockless freelist.
3057 *
3058 * And if we were unable to get a new slab from the partial slab lists then
3059 * we need to allocate a new slab. This is the slowest path since it involves
3060 * a call to the page allocator and the setup of a new slab.
3061 *
3062 * Version of __slab_alloc to use when we know that preemption is
3063 * already disabled (which is the case for bulk allocation).
3064 */
3065static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
3066 unsigned long addr, struct kmem_cache_cpu *c, unsigned int orig_size)
3067{
3068 void *freelist;
3069 struct slab *slab;
3070 unsigned long flags;
3071 struct partial_context pc;
3072
3073 stat(s, ALLOC_SLOWPATH);
3074
3075reread_slab:
3076
3077 slab = READ_ONCE(c->slab);
3078 if (!slab) {
3079 /*
3080 * if the node is not online or has no normal memory, just
3081 * ignore the node constraint
3082 */
3083 if (unlikely(node != NUMA_NO_NODE &&
3084 !node_isset(node, slab_nodes)))
3085 node = NUMA_NO_NODE;
3086 goto new_slab;
3087 }
3088redo:
3089
3090 if (unlikely(!node_match(slab, node))) {
3091 /*
3092 * same as above but node_match() being false already
3093 * implies node != NUMA_NO_NODE
3094 */
3095 if (!node_isset(node, slab_nodes)) {
3096 node = NUMA_NO_NODE;
3097 } else {
3098 stat(s, ALLOC_NODE_MISMATCH);
3099 goto deactivate_slab;
3100 }
3101 }
3102
3103 /*
3104 * By rights, we should be searching for a slab page that was
3105 * PFMEMALLOC but right now, we are losing the pfmemalloc
3106 * information when the page leaves the per-cpu allocator
3107 */
3108 if (unlikely(!pfmemalloc_match(slab, gfpflags)))
3109 goto deactivate_slab;
3110
3111 /* must check again c->slab in case we got preempted and it changed */
3112 local_lock_irqsave(&s->cpu_slab->lock, flags);
3113 if (unlikely(slab != c->slab)) {
3114 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3115 goto reread_slab;
3116 }
3117 freelist = c->freelist;
3118 if (freelist)
3119 goto load_freelist;
3120
3121 freelist = get_freelist(s, slab);
3122
3123 if (!freelist) {
3124 c->slab = NULL;
3125 c->tid = next_tid(c->tid);
3126 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3127 stat(s, DEACTIVATE_BYPASS);
3128 goto new_slab;
3129 }
3130
3131 stat(s, ALLOC_REFILL);
3132
3133load_freelist:
3134
3135 lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock));
3136
3137 /*
3138 * freelist is pointing to the list of objects to be used.
3139 * slab is pointing to the slab from which the objects are obtained.
3140 * That slab must be frozen for per cpu allocations to work.
3141 */
3142 VM_BUG_ON(!c->slab->frozen);
3143 c->freelist = get_freepointer(s, freelist);
3144 c->tid = next_tid(c->tid);
3145 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3146 return freelist;
3147
3148deactivate_slab:
3149
3150 local_lock_irqsave(&s->cpu_slab->lock, flags);
3151 if (slab != c->slab) {
3152 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3153 goto reread_slab;
3154 }
3155 freelist = c->freelist;
3156 c->slab = NULL;
3157 c->freelist = NULL;
3158 c->tid = next_tid(c->tid);
3159 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3160 deactivate_slab(s, slab, freelist);
3161
3162new_slab:
3163
3164 if (slub_percpu_partial(c)) {
3165 local_lock_irqsave(&s->cpu_slab->lock, flags);
3166 if (unlikely(c->slab)) {
3167 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3168 goto reread_slab;
3169 }
3170 if (unlikely(!slub_percpu_partial(c))) {
3171 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3172 /* we were preempted and partial list got empty */
3173 goto new_objects;
3174 }
3175
3176 slab = c->slab = slub_percpu_partial(c);
3177 slub_set_percpu_partial(c, slab);
3178 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3179 stat(s, CPU_PARTIAL_ALLOC);
3180 goto redo;
3181 }
3182
3183new_objects:
3184
3185 pc.flags = gfpflags;
3186 pc.slab = &slab;
3187 pc.orig_size = orig_size;
3188 freelist = get_partial(s, node, &pc);
3189 if (freelist)
3190 goto check_new_slab;
3191
3192 slub_put_cpu_ptr(s->cpu_slab);
3193 slab = new_slab(s, gfpflags, node);
3194 c = slub_get_cpu_ptr(s->cpu_slab);
3195
3196 if (unlikely(!slab)) {
3197 slab_out_of_memory(s, gfpflags, node);
3198 return NULL;
3199 }
3200
3201 stat(s, ALLOC_SLAB);
3202
3203 if (kmem_cache_debug(s)) {
3204 freelist = alloc_single_from_new_slab(s, slab, orig_size);
3205
3206 if (unlikely(!freelist))
3207 goto new_objects;
3208
3209 if (s->flags & SLAB_STORE_USER)
3210 set_track(s, freelist, TRACK_ALLOC, addr);
3211
3212 return freelist;
3213 }
3214
3215 /*
3216 * No other reference to the slab yet so we can
3217 * muck around with it freely without cmpxchg
3218 */
3219 freelist = slab->freelist;
3220 slab->freelist = NULL;
3221 slab->inuse = slab->objects;
3222 slab->frozen = 1;
3223
3224 inc_slabs_node(s, slab_nid(slab), slab->objects);
3225
3226check_new_slab:
3227
3228 if (kmem_cache_debug(s)) {
3229 /*
3230 * For debug caches here we had to go through
3231 * alloc_single_from_partial() so just store the tracking info
3232 * and return the object
3233 */
3234 if (s->flags & SLAB_STORE_USER)
3235 set_track(s, freelist, TRACK_ALLOC, addr);
3236
3237 return freelist;
3238 }
3239
3240 if (unlikely(!pfmemalloc_match(slab, gfpflags))) {
3241 /*
3242 * For !pfmemalloc_match() case we don't load freelist so that
3243 * we don't make further mismatched allocations easier.
3244 */
3245 deactivate_slab(s, slab, get_freepointer(s, freelist));
3246 return freelist;
3247 }
3248
3249retry_load_slab:
3250
3251 local_lock_irqsave(&s->cpu_slab->lock, flags);
3252 if (unlikely(c->slab)) {
3253 void *flush_freelist = c->freelist;
3254 struct slab *flush_slab = c->slab;
3255
3256 c->slab = NULL;
3257 c->freelist = NULL;
3258 c->tid = next_tid(c->tid);
3259
3260 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3261
3262 deactivate_slab(s, flush_slab, flush_freelist);
3263
3264 stat(s, CPUSLAB_FLUSH);
3265
3266 goto retry_load_slab;
3267 }
3268 c->slab = slab;
3269
3270 goto load_freelist;
3271}
3272
3273/*
3274 * A wrapper for ___slab_alloc() for contexts where preemption is not yet
3275 * disabled. Compensates for possible cpu changes by refetching the per cpu area
3276 * pointer.
3277 */
3278static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
3279 unsigned long addr, struct kmem_cache_cpu *c, unsigned int orig_size)
3280{
3281 void *p;
3282
3283#ifdef CONFIG_PREEMPT_COUNT
3284 /*
3285 * We may have been preempted and rescheduled on a different
3286 * cpu before disabling preemption. Need to reload cpu area
3287 * pointer.
3288 */
3289 c = slub_get_cpu_ptr(s->cpu_slab);
3290#endif
3291
3292 p = ___slab_alloc(s, gfpflags, node, addr, c, orig_size);
3293#ifdef CONFIG_PREEMPT_COUNT
3294 slub_put_cpu_ptr(s->cpu_slab);
3295#endif
3296 return p;
3297}
3298
3299static __always_inline void *__slab_alloc_node(struct kmem_cache *s,
3300 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
3301{
3302 struct kmem_cache_cpu *c;
3303 struct slab *slab;
3304 unsigned long tid;
3305 void *object;
3306
3307redo:
3308 /*
3309 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
3310 * enabled. We may switch back and forth between cpus while
3311 * reading from one cpu area. That does not matter as long
3312 * as we end up on the original cpu again when doing the cmpxchg.
3313 *
3314 * We must guarantee that tid and kmem_cache_cpu are retrieved on the
3315 * same cpu. We read first the kmem_cache_cpu pointer and use it to read
3316 * the tid. If we are preempted and switched to another cpu between the
3317 * two reads, it's OK as the two are still associated with the same cpu
3318 * and cmpxchg later will validate the cpu.
3319 */
3320 c = raw_cpu_ptr(s->cpu_slab);
3321 tid = READ_ONCE(c->tid);
3322
3323 /*
3324 * Irqless object alloc/free algorithm used here depends on sequence
3325 * of fetching cpu_slab's data. tid should be fetched before anything
3326 * on c to guarantee that object and slab associated with previous tid
3327 * won't be used with current tid. If we fetch tid first, object and
3328 * slab could be one associated with next tid and our alloc/free
3329 * request will be failed. In this case, we will retry. So, no problem.
3330 */
3331 barrier();
3332
3333 /*
3334 * The transaction ids are globally unique per cpu and per operation on
3335 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
3336 * occurs on the right processor and that there was no operation on the
3337 * linked list in between.
3338 */
3339
3340 object = c->freelist;
3341 slab = c->slab;
3342
3343 if (!USE_LOCKLESS_FAST_PATH() ||
3344 unlikely(!object || !slab || !node_match(slab, node))) {
3345 object = __slab_alloc(s, gfpflags, node, addr, c, orig_size);
3346 } else {
3347 void *next_object = get_freepointer_safe(s, object);
3348
3349 /*
3350 * The cmpxchg will only match if there was no additional
3351 * operation and if we are on the right processor.
3352 *
3353 * The cmpxchg does the following atomically (without lock
3354 * semantics!)
3355 * 1. Relocate first pointer to the current per cpu area.
3356 * 2. Verify that tid and freelist have not been changed
3357 * 3. If they were not changed replace tid and freelist
3358 *
3359 * Since this is without lock semantics the protection is only
3360 * against code executing on this cpu *not* from access by
3361 * other cpus.
3362 */
3363 if (unlikely(!this_cpu_cmpxchg_double(
3364 s->cpu_slab->freelist, s->cpu_slab->tid,
3365 object, tid,
3366 next_object, next_tid(tid)))) {
3367
3368 note_cmpxchg_failure("slab_alloc", s, tid);
3369 goto redo;
3370 }
3371 prefetch_freepointer(s, next_object);
3372 stat(s, ALLOC_FASTPATH);
3373 }
3374
3375 return object;
3376}
3377#else /* CONFIG_SLUB_TINY */
3378static void *__slab_alloc_node(struct kmem_cache *s,
3379 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
3380{
3381 struct partial_context pc;
3382 struct slab *slab;
3383 void *object;
3384
3385 pc.flags = gfpflags;
3386 pc.slab = &slab;
3387 pc.orig_size = orig_size;
3388 object = get_partial(s, node, &pc);
3389
3390 if (object)
3391 return object;
3392
3393 slab = new_slab(s, gfpflags, node);
3394 if (unlikely(!slab)) {
3395 slab_out_of_memory(s, gfpflags, node);
3396 return NULL;
3397 }
3398
3399 object = alloc_single_from_new_slab(s, slab, orig_size);
3400
3401 return object;
3402}
3403#endif /* CONFIG_SLUB_TINY */
3404
3405/*
3406 * If the object has been wiped upon free, make sure it's fully initialized by
3407 * zeroing out freelist pointer.
3408 */
3409static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s,
3410 void *obj)
3411{
3412 if (unlikely(slab_want_init_on_free(s)) && obj)
3413 memset((void *)((char *)kasan_reset_tag(obj) + s->offset),
3414 0, sizeof(void *));
3415}
3416
3417/*
3418 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
3419 * have the fastpath folded into their functions. So no function call
3420 * overhead for requests that can be satisfied on the fastpath.
3421 *
3422 * The fastpath works by first checking if the lockless freelist can be used.
3423 * If not then __slab_alloc is called for slow processing.
3424 *
3425 * Otherwise we can simply pick the next object from the lockless free list.
3426 */
3427static __fastpath_inline void *slab_alloc_node(struct kmem_cache *s, struct list_lru *lru,
3428 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
3429{
3430 void *object;
3431 struct obj_cgroup *objcg = NULL;
3432 bool init = false;
3433
3434 s = slab_pre_alloc_hook(s, lru, &objcg, 1, gfpflags);
3435 if (!s)
3436 return NULL;
3437
3438 object = kfence_alloc(s, orig_size, gfpflags);
3439 if (unlikely(object))
3440 goto out;
3441
3442 object = __slab_alloc_node(s, gfpflags, node, addr, orig_size);
3443
3444 maybe_wipe_obj_freeptr(s, object);
3445 init = slab_want_init_on_alloc(gfpflags, s);
3446
3447out:
3448 /*
3449 * When init equals 'true', like for kzalloc() family, only
3450 * @orig_size bytes might be zeroed instead of s->object_size
3451 */
3452 slab_post_alloc_hook(s, objcg, gfpflags, 1, &object, init, orig_size);
3453
3454 return object;
3455}
3456
3457static __fastpath_inline void *slab_alloc(struct kmem_cache *s, struct list_lru *lru,
3458 gfp_t gfpflags, unsigned long addr, size_t orig_size)
3459{
3460 return slab_alloc_node(s, lru, gfpflags, NUMA_NO_NODE, addr, orig_size);
3461}
3462
3463static __fastpath_inline
3464void *__kmem_cache_alloc_lru(struct kmem_cache *s, struct list_lru *lru,
3465 gfp_t gfpflags)
3466{
3467 void *ret = slab_alloc(s, lru, gfpflags, _RET_IP_, s->object_size);
3468
3469 trace_kmem_cache_alloc(_RET_IP_, ret, s, gfpflags, NUMA_NO_NODE);
3470
3471 return ret;
3472}
3473
3474void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
3475{
3476 return __kmem_cache_alloc_lru(s, NULL, gfpflags);
3477}
3478EXPORT_SYMBOL(kmem_cache_alloc);
3479
3480void *kmem_cache_alloc_lru(struct kmem_cache *s, struct list_lru *lru,
3481 gfp_t gfpflags)
3482{
3483 return __kmem_cache_alloc_lru(s, lru, gfpflags);
3484}
3485EXPORT_SYMBOL(kmem_cache_alloc_lru);
3486
3487void *__kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags,
3488 int node, size_t orig_size,
3489 unsigned long caller)
3490{
3491 return slab_alloc_node(s, NULL, gfpflags, node,
3492 caller, orig_size);
3493}
3494
3495void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
3496{
3497 void *ret = slab_alloc_node(s, NULL, gfpflags, node, _RET_IP_, s->object_size);
3498
3499 trace_kmem_cache_alloc(_RET_IP_, ret, s, gfpflags, node);
3500
3501 return ret;
3502}
3503EXPORT_SYMBOL(kmem_cache_alloc_node);
3504
3505static noinline void free_to_partial_list(
3506 struct kmem_cache *s, struct slab *slab,
3507 void *head, void *tail, int bulk_cnt,
3508 unsigned long addr)
3509{
3510 struct kmem_cache_node *n = get_node(s, slab_nid(slab));
3511 struct slab *slab_free = NULL;
3512 int cnt = bulk_cnt;
3513 unsigned long flags;
3514 depot_stack_handle_t handle = 0;
3515
3516 if (s->flags & SLAB_STORE_USER)
3517 handle = set_track_prepare();
3518
3519 spin_lock_irqsave(&n->list_lock, flags);
3520
3521 if (free_debug_processing(s, slab, head, tail, &cnt, addr, handle)) {
3522 void *prior = slab->freelist;
3523
3524 /* Perform the actual freeing while we still hold the locks */
3525 slab->inuse -= cnt;
3526 set_freepointer(s, tail, prior);
3527 slab->freelist = head;
3528
3529 /*
3530 * If the slab is empty, and node's partial list is full,
3531 * it should be discarded anyway no matter it's on full or
3532 * partial list.
3533 */
3534 if (slab->inuse == 0 && n->nr_partial >= s->min_partial)
3535 slab_free = slab;
3536
3537 if (!prior) {
3538 /* was on full list */
3539 remove_full(s, n, slab);
3540 if (!slab_free) {
3541 add_partial(n, slab, DEACTIVATE_TO_TAIL);
3542 stat(s, FREE_ADD_PARTIAL);
3543 }
3544 } else if (slab_free) {
3545 remove_partial(n, slab);
3546 stat(s, FREE_REMOVE_PARTIAL);
3547 }
3548 }
3549
3550 if (slab_free) {
3551 /*
3552 * Update the counters while still holding n->list_lock to
3553 * prevent spurious validation warnings
3554 */
3555 dec_slabs_node(s, slab_nid(slab_free), slab_free->objects);
3556 }
3557
3558 spin_unlock_irqrestore(&n->list_lock, flags);
3559
3560 if (slab_free) {
3561 stat(s, FREE_SLAB);
3562 free_slab(s, slab_free);
3563 }
3564}
3565
3566/*
3567 * Slow path handling. This may still be called frequently since objects
3568 * have a longer lifetime than the cpu slabs in most processing loads.
3569 *
3570 * So we still attempt to reduce cache line usage. Just take the slab
3571 * lock and free the item. If there is no additional partial slab
3572 * handling required then we can return immediately.
3573 */
3574static void __slab_free(struct kmem_cache *s, struct slab *slab,
3575 void *head, void *tail, int cnt,
3576 unsigned long addr)
3577
3578{
3579 void *prior;
3580 int was_frozen;
3581 struct slab new;
3582 unsigned long counters;
3583 struct kmem_cache_node *n = NULL;
3584 unsigned long flags;
3585
3586 stat(s, FREE_SLOWPATH);
3587
3588 if (kfence_free(head))
3589 return;
3590
3591 if (IS_ENABLED(CONFIG_SLUB_TINY) || kmem_cache_debug(s)) {
3592 free_to_partial_list(s, slab, head, tail, cnt, addr);
3593 return;
3594 }
3595
3596 do {
3597 if (unlikely(n)) {
3598 spin_unlock_irqrestore(&n->list_lock, flags);
3599 n = NULL;
3600 }
3601 prior = slab->freelist;
3602 counters = slab->counters;
3603 set_freepointer(s, tail, prior);
3604 new.counters = counters;
3605 was_frozen = new.frozen;
3606 new.inuse -= cnt;
3607 if ((!new.inuse || !prior) && !was_frozen) {
3608
3609 if (kmem_cache_has_cpu_partial(s) && !prior) {
3610
3611 /*
3612 * Slab was on no list before and will be
3613 * partially empty
3614 * We can defer the list move and instead
3615 * freeze it.
3616 */
3617 new.frozen = 1;
3618
3619 } else { /* Needs to be taken off a list */
3620
3621 n = get_node(s, slab_nid(slab));
3622 /*
3623 * Speculatively acquire the list_lock.
3624 * If the cmpxchg does not succeed then we may
3625 * drop the list_lock without any processing.
3626 *
3627 * Otherwise the list_lock will synchronize with
3628 * other processors updating the list of slabs.
3629 */
3630 spin_lock_irqsave(&n->list_lock, flags);
3631
3632 }
3633 }
3634
3635 } while (!cmpxchg_double_slab(s, slab,
3636 prior, counters,
3637 head, new.counters,
3638 "__slab_free"));
3639
3640 if (likely(!n)) {
3641
3642 if (likely(was_frozen)) {
3643 /*
3644 * The list lock was not taken therefore no list
3645 * activity can be necessary.
3646 */
3647 stat(s, FREE_FROZEN);
3648 } else if (new.frozen) {
3649 /*
3650 * If we just froze the slab then put it onto the
3651 * per cpu partial list.
3652 */
3653 put_cpu_partial(s, slab, 1);
3654 stat(s, CPU_PARTIAL_FREE);
3655 }
3656
3657 return;
3658 }
3659
3660 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
3661 goto slab_empty;
3662
3663 /*
3664 * Objects left in the slab. If it was not on the partial list before
3665 * then add it.
3666 */
3667 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
3668 remove_full(s, n, slab);
3669 add_partial(n, slab, DEACTIVATE_TO_TAIL);
3670 stat(s, FREE_ADD_PARTIAL);
3671 }
3672 spin_unlock_irqrestore(&n->list_lock, flags);
3673 return;
3674
3675slab_empty:
3676 if (prior) {
3677 /*
3678 * Slab on the partial list.
3679 */
3680 remove_partial(n, slab);
3681 stat(s, FREE_REMOVE_PARTIAL);
3682 } else {
3683 /* Slab must be on the full list */
3684 remove_full(s, n, slab);
3685 }
3686
3687 spin_unlock_irqrestore(&n->list_lock, flags);
3688 stat(s, FREE_SLAB);
3689 discard_slab(s, slab);
3690}
3691
3692#ifndef CONFIG_SLUB_TINY
3693/*
3694 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
3695 * can perform fastpath freeing without additional function calls.
3696 *
3697 * The fastpath is only possible if we are freeing to the current cpu slab
3698 * of this processor. This typically the case if we have just allocated
3699 * the item before.
3700 *
3701 * If fastpath is not possible then fall back to __slab_free where we deal
3702 * with all sorts of special processing.
3703 *
3704 * Bulk free of a freelist with several objects (all pointing to the
3705 * same slab) possible by specifying head and tail ptr, plus objects
3706 * count (cnt). Bulk free indicated by tail pointer being set.
3707 */
3708static __always_inline void do_slab_free(struct kmem_cache *s,
3709 struct slab *slab, void *head, void *tail,
3710 int cnt, unsigned long addr)
3711{
3712 void *tail_obj = tail ? : head;
3713 struct kmem_cache_cpu *c;
3714 unsigned long tid;
3715 void **freelist;
3716
3717redo:
3718 /*
3719 * Determine the currently cpus per cpu slab.
3720 * The cpu may change afterward. However that does not matter since
3721 * data is retrieved via this pointer. If we are on the same cpu
3722 * during the cmpxchg then the free will succeed.
3723 */
3724 c = raw_cpu_ptr(s->cpu_slab);
3725 tid = READ_ONCE(c->tid);
3726
3727 /* Same with comment on barrier() in slab_alloc_node() */
3728 barrier();
3729
3730 if (unlikely(slab != c->slab)) {
3731 __slab_free(s, slab, head, tail_obj, cnt, addr);
3732 return;
3733 }
3734
3735 if (USE_LOCKLESS_FAST_PATH()) {
3736 freelist = READ_ONCE(c->freelist);
3737
3738 set_freepointer(s, tail_obj, freelist);
3739
3740 if (unlikely(!this_cpu_cmpxchg_double(
3741 s->cpu_slab->freelist, s->cpu_slab->tid,
3742 freelist, tid,
3743 head, next_tid(tid)))) {
3744
3745 note_cmpxchg_failure("slab_free", s, tid);
3746 goto redo;
3747 }
3748 } else {
3749 /* Update the free list under the local lock */
3750 local_lock(&s->cpu_slab->lock);
3751 c = this_cpu_ptr(s->cpu_slab);
3752 if (unlikely(slab != c->slab)) {
3753 local_unlock(&s->cpu_slab->lock);
3754 goto redo;
3755 }
3756 tid = c->tid;
3757 freelist = c->freelist;
3758
3759 set_freepointer(s, tail_obj, freelist);
3760 c->freelist = head;
3761 c->tid = next_tid(tid);
3762
3763 local_unlock(&s->cpu_slab->lock);
3764 }
3765 stat(s, FREE_FASTPATH);
3766}
3767#else /* CONFIG_SLUB_TINY */
3768static void do_slab_free(struct kmem_cache *s,
3769 struct slab *slab, void *head, void *tail,
3770 int cnt, unsigned long addr)
3771{
3772 void *tail_obj = tail ? : head;
3773
3774 __slab_free(s, slab, head, tail_obj, cnt, addr);
3775}
3776#endif /* CONFIG_SLUB_TINY */
3777
3778static __fastpath_inline void slab_free(struct kmem_cache *s, struct slab *slab,
3779 void *head, void *tail, void **p, int cnt,
3780 unsigned long addr)
3781{
3782 memcg_slab_free_hook(s, slab, p, cnt);
3783 /*
3784 * With KASAN enabled slab_free_freelist_hook modifies the freelist
3785 * to remove objects, whose reuse must be delayed.
3786 */
3787 if (slab_free_freelist_hook(s, &head, &tail, &cnt))
3788 do_slab_free(s, slab, head, tail, cnt, addr);
3789}
3790
3791#ifdef CONFIG_KASAN_GENERIC
3792void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
3793{
3794 do_slab_free(cache, virt_to_slab(x), x, NULL, 1, addr);
3795}
3796#endif
3797
3798void __kmem_cache_free(struct kmem_cache *s, void *x, unsigned long caller)
3799{
3800 slab_free(s, virt_to_slab(x), x, NULL, &x, 1, caller);
3801}
3802
3803void kmem_cache_free(struct kmem_cache *s, void *x)
3804{
3805 s = cache_from_obj(s, x);
3806 if (!s)
3807 return;
3808 trace_kmem_cache_free(_RET_IP_, x, s);
3809 slab_free(s, virt_to_slab(x), x, NULL, &x, 1, _RET_IP_);
3810}
3811EXPORT_SYMBOL(kmem_cache_free);
3812
3813struct detached_freelist {
3814 struct slab *slab;
3815 void *tail;
3816 void *freelist;
3817 int cnt;
3818 struct kmem_cache *s;
3819};
3820
3821/*
3822 * This function progressively scans the array with free objects (with
3823 * a limited look ahead) and extract objects belonging to the same
3824 * slab. It builds a detached freelist directly within the given
3825 * slab/objects. This can happen without any need for
3826 * synchronization, because the objects are owned by running process.
3827 * The freelist is build up as a single linked list in the objects.
3828 * The idea is, that this detached freelist can then be bulk
3829 * transferred to the real freelist(s), but only requiring a single
3830 * synchronization primitive. Look ahead in the array is limited due
3831 * to performance reasons.
3832 */
3833static inline
3834int build_detached_freelist(struct kmem_cache *s, size_t size,
3835 void **p, struct detached_freelist *df)
3836{
3837 int lookahead = 3;
3838 void *object;
3839 struct folio *folio;
3840 size_t same;
3841
3842 object = p[--size];
3843 folio = virt_to_folio(object);
3844 if (!s) {
3845 /* Handle kalloc'ed objects */
3846 if (unlikely(!folio_test_slab(folio))) {
3847 free_large_kmalloc(folio, object);
3848 df->slab = NULL;
3849 return size;
3850 }
3851 /* Derive kmem_cache from object */
3852 df->slab = folio_slab(folio);
3853 df->s = df->slab->slab_cache;
3854 } else {
3855 df->slab = folio_slab(folio);
3856 df->s = cache_from_obj(s, object); /* Support for memcg */
3857 }
3858
3859 /* Start new detached freelist */
3860 df->tail = object;
3861 df->freelist = object;
3862 df->cnt = 1;
3863
3864 if (is_kfence_address(object))
3865 return size;
3866
3867 set_freepointer(df->s, object, NULL);
3868
3869 same = size;
3870 while (size) {
3871 object = p[--size];
3872 /* df->slab is always set at this point */
3873 if (df->slab == virt_to_slab(object)) {
3874 /* Opportunity build freelist */
3875 set_freepointer(df->s, object, df->freelist);
3876 df->freelist = object;
3877 df->cnt++;
3878 same--;
3879 if (size != same)
3880 swap(p[size], p[same]);
3881 continue;
3882 }
3883
3884 /* Limit look ahead search */
3885 if (!--lookahead)
3886 break;
3887 }
3888
3889 return same;
3890}
3891
3892/* Note that interrupts must be enabled when calling this function. */
3893void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3894{
3895 if (!size)
3896 return;
3897
3898 do {
3899 struct detached_freelist df;
3900
3901 size = build_detached_freelist(s, size, p, &df);
3902 if (!df.slab)
3903 continue;
3904
3905 slab_free(df.s, df.slab, df.freelist, df.tail, &p[size], df.cnt,
3906 _RET_IP_);
3907 } while (likely(size));
3908}
3909EXPORT_SYMBOL(kmem_cache_free_bulk);
3910
3911#ifndef CONFIG_SLUB_TINY
3912static inline int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags,
3913 size_t size, void **p, struct obj_cgroup *objcg)
3914{
3915 struct kmem_cache_cpu *c;
3916 int i;
3917
3918 /*
3919 * Drain objects in the per cpu slab, while disabling local
3920 * IRQs, which protects against PREEMPT and interrupts
3921 * handlers invoking normal fastpath.
3922 */
3923 c = slub_get_cpu_ptr(s->cpu_slab);
3924 local_lock_irq(&s->cpu_slab->lock);
3925
3926 for (i = 0; i < size; i++) {
3927 void *object = kfence_alloc(s, s->object_size, flags);
3928
3929 if (unlikely(object)) {
3930 p[i] = object;
3931 continue;
3932 }
3933
3934 object = c->freelist;
3935 if (unlikely(!object)) {
3936 /*
3937 * We may have removed an object from c->freelist using
3938 * the fastpath in the previous iteration; in that case,
3939 * c->tid has not been bumped yet.
3940 * Since ___slab_alloc() may reenable interrupts while
3941 * allocating memory, we should bump c->tid now.
3942 */
3943 c->tid = next_tid(c->tid);
3944
3945 local_unlock_irq(&s->cpu_slab->lock);
3946
3947 /*
3948 * Invoking slow path likely have side-effect
3949 * of re-populating per CPU c->freelist
3950 */
3951 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3952 _RET_IP_, c, s->object_size);
3953 if (unlikely(!p[i]))
3954 goto error;
3955
3956 c = this_cpu_ptr(s->cpu_slab);
3957 maybe_wipe_obj_freeptr(s, p[i]);
3958
3959 local_lock_irq(&s->cpu_slab->lock);
3960
3961 continue; /* goto for-loop */
3962 }
3963 c->freelist = get_freepointer(s, object);
3964 p[i] = object;
3965 maybe_wipe_obj_freeptr(s, p[i]);
3966 }
3967 c->tid = next_tid(c->tid);
3968 local_unlock_irq(&s->cpu_slab->lock);
3969 slub_put_cpu_ptr(s->cpu_slab);
3970
3971 return i;
3972
3973error:
3974 slub_put_cpu_ptr(s->cpu_slab);
3975 slab_post_alloc_hook(s, objcg, flags, i, p, false, s->object_size);
3976 kmem_cache_free_bulk(s, i, p);
3977 return 0;
3978
3979}
3980#else /* CONFIG_SLUB_TINY */
3981static int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags,
3982 size_t size, void **p, struct obj_cgroup *objcg)
3983{
3984 int i;
3985
3986 for (i = 0; i < size; i++) {
3987 void *object = kfence_alloc(s, s->object_size, flags);
3988
3989 if (unlikely(object)) {
3990 p[i] = object;
3991 continue;
3992 }
3993
3994 p[i] = __slab_alloc_node(s, flags, NUMA_NO_NODE,
3995 _RET_IP_, s->object_size);
3996 if (unlikely(!p[i]))
3997 goto error;
3998
3999 maybe_wipe_obj_freeptr(s, p[i]);
4000 }
4001
4002 return i;
4003
4004error:
4005 slab_post_alloc_hook(s, objcg, flags, i, p, false, s->object_size);
4006 kmem_cache_free_bulk(s, i, p);
4007 return 0;
4008}
4009#endif /* CONFIG_SLUB_TINY */
4010
4011/* Note that interrupts must be enabled when calling this function. */
4012int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
4013 void **p)
4014{
4015 int i;
4016 struct obj_cgroup *objcg = NULL;
4017
4018 if (!size)
4019 return 0;
4020
4021 /* memcg and kmem_cache debug support */
4022 s = slab_pre_alloc_hook(s, NULL, &objcg, size, flags);
4023 if (unlikely(!s))
4024 return 0;
4025
4026 i = __kmem_cache_alloc_bulk(s, flags, size, p, objcg);
4027
4028 /*
4029 * memcg and kmem_cache debug support and memory initialization.
4030 * Done outside of the IRQ disabled fastpath loop.
4031 */
4032 if (i != 0)
4033 slab_post_alloc_hook(s, objcg, flags, size, p,
4034 slab_want_init_on_alloc(flags, s), s->object_size);
4035 return i;
4036}
4037EXPORT_SYMBOL(kmem_cache_alloc_bulk);
4038
4039
4040/*
4041 * Object placement in a slab is made very easy because we always start at
4042 * offset 0. If we tune the size of the object to the alignment then we can
4043 * get the required alignment by putting one properly sized object after
4044 * another.
4045 *
4046 * Notice that the allocation order determines the sizes of the per cpu
4047 * caches. Each processor has always one slab available for allocations.
4048 * Increasing the allocation order reduces the number of times that slabs
4049 * must be moved on and off the partial lists and is therefore a factor in
4050 * locking overhead.
4051 */
4052
4053/*
4054 * Minimum / Maximum order of slab pages. This influences locking overhead
4055 * and slab fragmentation. A higher order reduces the number of partial slabs
4056 * and increases the number of allocations possible without having to
4057 * take the list_lock.
4058 */
4059static unsigned int slub_min_order;
4060static unsigned int slub_max_order =
4061 IS_ENABLED(CONFIG_SLUB_TINY) ? 1 : PAGE_ALLOC_COSTLY_ORDER;
4062static unsigned int slub_min_objects;
4063
4064/*
4065 * Calculate the order of allocation given an slab object size.
4066 *
4067 * The order of allocation has significant impact on performance and other
4068 * system components. Generally order 0 allocations should be preferred since
4069 * order 0 does not cause fragmentation in the page allocator. Larger objects
4070 * be problematic to put into order 0 slabs because there may be too much
4071 * unused space left. We go to a higher order if more than 1/16th of the slab
4072 * would be wasted.
4073 *
4074 * In order to reach satisfactory performance we must ensure that a minimum
4075 * number of objects is in one slab. Otherwise we may generate too much
4076 * activity on the partial lists which requires taking the list_lock. This is
4077 * less a concern for large slabs though which are rarely used.
4078 *
4079 * slub_max_order specifies the order where we begin to stop considering the
4080 * number of objects in a slab as critical. If we reach slub_max_order then
4081 * we try to keep the page order as low as possible. So we accept more waste
4082 * of space in favor of a small page order.
4083 *
4084 * Higher order allocations also allow the placement of more objects in a
4085 * slab and thereby reduce object handling overhead. If the user has
4086 * requested a higher minimum order then we start with that one instead of
4087 * the smallest order which will fit the object.
4088 */
4089static inline unsigned int calc_slab_order(unsigned int size,
4090 unsigned int min_objects, unsigned int max_order,
4091 unsigned int fract_leftover)
4092{
4093 unsigned int min_order = slub_min_order;
4094 unsigned int order;
4095
4096 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
4097 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
4098
4099 for (order = max(min_order, (unsigned int)get_order(min_objects * size));
4100 order <= max_order; order++) {
4101
4102 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
4103 unsigned int rem;
4104
4105 rem = slab_size % size;
4106
4107 if (rem <= slab_size / fract_leftover)
4108 break;
4109 }
4110
4111 return order;
4112}
4113
4114static inline int calculate_order(unsigned int size)
4115{
4116 unsigned int order;
4117 unsigned int min_objects;
4118 unsigned int max_objects;
4119 unsigned int nr_cpus;
4120
4121 /*
4122 * Attempt to find best configuration for a slab. This
4123 * works by first attempting to generate a layout with
4124 * the best configuration and backing off gradually.
4125 *
4126 * First we increase the acceptable waste in a slab. Then
4127 * we reduce the minimum objects required in a slab.
4128 */
4129 min_objects = slub_min_objects;
4130 if (!min_objects) {
4131 /*
4132 * Some architectures will only update present cpus when
4133 * onlining them, so don't trust the number if it's just 1. But
4134 * we also don't want to use nr_cpu_ids always, as on some other
4135 * architectures, there can be many possible cpus, but never
4136 * onlined. Here we compromise between trying to avoid too high
4137 * order on systems that appear larger than they are, and too
4138 * low order on systems that appear smaller than they are.
4139 */
4140 nr_cpus = num_present_cpus();
4141 if (nr_cpus <= 1)
4142 nr_cpus = nr_cpu_ids;
4143 min_objects = 4 * (fls(nr_cpus) + 1);
4144 }
4145 max_objects = order_objects(slub_max_order, size);
4146 min_objects = min(min_objects, max_objects);
4147
4148 while (min_objects > 1) {
4149 unsigned int fraction;
4150
4151 fraction = 16;
4152 while (fraction >= 4) {
4153 order = calc_slab_order(size, min_objects,
4154 slub_max_order, fraction);
4155 if (order <= slub_max_order)
4156 return order;
4157 fraction /= 2;
4158 }
4159 min_objects--;
4160 }
4161
4162 /*
4163 * We were unable to place multiple objects in a slab. Now
4164 * lets see if we can place a single object there.
4165 */
4166 order = calc_slab_order(size, 1, slub_max_order, 1);
4167 if (order <= slub_max_order)
4168 return order;
4169
4170 /*
4171 * Doh this slab cannot be placed using slub_max_order.
4172 */
4173 order = calc_slab_order(size, 1, MAX_ORDER, 1);
4174 if (order < MAX_ORDER)
4175 return order;
4176 return -ENOSYS;
4177}
4178
4179static void
4180init_kmem_cache_node(struct kmem_cache_node *n)
4181{
4182 n->nr_partial = 0;
4183 spin_lock_init(&n->list_lock);
4184 INIT_LIST_HEAD(&n->partial);
4185#ifdef CONFIG_SLUB_DEBUG
4186 atomic_long_set(&n->nr_slabs, 0);
4187 atomic_long_set(&n->total_objects, 0);
4188 INIT_LIST_HEAD(&n->full);
4189#endif
4190}
4191
4192#ifndef CONFIG_SLUB_TINY
4193static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
4194{
4195 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
4196 NR_KMALLOC_TYPES * KMALLOC_SHIFT_HIGH *
4197 sizeof(struct kmem_cache_cpu));
4198
4199 /*
4200 * Must align to double word boundary for the double cmpxchg
4201 * instructions to work; see __pcpu_double_call_return_bool().
4202 */
4203 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
4204 2 * sizeof(void *));
4205
4206 if (!s->cpu_slab)
4207 return 0;
4208
4209 init_kmem_cache_cpus(s);
4210
4211 return 1;
4212}
4213#else
4214static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
4215{
4216 return 1;
4217}
4218#endif /* CONFIG_SLUB_TINY */
4219
4220static struct kmem_cache *kmem_cache_node;
4221
4222/*
4223 * No kmalloc_node yet so do it by hand. We know that this is the first
4224 * slab on the node for this slabcache. There are no concurrent accesses
4225 * possible.
4226 *
4227 * Note that this function only works on the kmem_cache_node
4228 * when allocating for the kmem_cache_node. This is used for bootstrapping
4229 * memory on a fresh node that has no slab structures yet.
4230 */
4231static void early_kmem_cache_node_alloc(int node)
4232{
4233 struct slab *slab;
4234 struct kmem_cache_node *n;
4235
4236 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
4237
4238 slab = new_slab(kmem_cache_node, GFP_NOWAIT, node);
4239
4240 BUG_ON(!slab);
4241 inc_slabs_node(kmem_cache_node, slab_nid(slab), slab->objects);
4242 if (slab_nid(slab) != node) {
4243 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
4244 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
4245 }
4246
4247 n = slab->freelist;
4248 BUG_ON(!n);
4249#ifdef CONFIG_SLUB_DEBUG
4250 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
4251 init_tracking(kmem_cache_node, n);
4252#endif
4253 n = kasan_slab_alloc(kmem_cache_node, n, GFP_KERNEL, false);
4254 slab->freelist = get_freepointer(kmem_cache_node, n);
4255 slab->inuse = 1;
4256 kmem_cache_node->node[node] = n;
4257 init_kmem_cache_node(n);
4258 inc_slabs_node(kmem_cache_node, node, slab->objects);
4259
4260 /*
4261 * No locks need to be taken here as it has just been
4262 * initialized and there is no concurrent access.
4263 */
4264 __add_partial(n, slab, DEACTIVATE_TO_HEAD);
4265}
4266
4267static void free_kmem_cache_nodes(struct kmem_cache *s)
4268{
4269 int node;
4270 struct kmem_cache_node *n;
4271
4272 for_each_kmem_cache_node(s, node, n) {
4273 s->node[node] = NULL;
4274 kmem_cache_free(kmem_cache_node, n);
4275 }
4276}
4277
4278void __kmem_cache_release(struct kmem_cache *s)
4279{
4280 cache_random_seq_destroy(s);
4281#ifndef CONFIG_SLUB_TINY
4282 free_percpu(s->cpu_slab);
4283#endif
4284 free_kmem_cache_nodes(s);
4285}
4286
4287static int init_kmem_cache_nodes(struct kmem_cache *s)
4288{
4289 int node;
4290
4291 for_each_node_mask(node, slab_nodes) {
4292 struct kmem_cache_node *n;
4293
4294 if (slab_state == DOWN) {
4295 early_kmem_cache_node_alloc(node);
4296 continue;
4297 }
4298 n = kmem_cache_alloc_node(kmem_cache_node,
4299 GFP_KERNEL, node);
4300
4301 if (!n) {
4302 free_kmem_cache_nodes(s);
4303 return 0;
4304 }
4305
4306 init_kmem_cache_node(n);
4307 s->node[node] = n;
4308 }
4309 return 1;
4310}
4311
4312static void set_cpu_partial(struct kmem_cache *s)
4313{
4314#ifdef CONFIG_SLUB_CPU_PARTIAL
4315 unsigned int nr_objects;
4316
4317 /*
4318 * cpu_partial determined the maximum number of objects kept in the
4319 * per cpu partial lists of a processor.
4320 *
4321 * Per cpu partial lists mainly contain slabs that just have one
4322 * object freed. If they are used for allocation then they can be
4323 * filled up again with minimal effort. The slab will never hit the
4324 * per node partial lists and therefore no locking will be required.
4325 *
4326 * For backwards compatibility reasons, this is determined as number
4327 * of objects, even though we now limit maximum number of pages, see
4328 * slub_set_cpu_partial()
4329 */
4330 if (!kmem_cache_has_cpu_partial(s))
4331 nr_objects = 0;
4332 else if (s->size >= PAGE_SIZE)
4333 nr_objects = 6;
4334 else if (s->size >= 1024)
4335 nr_objects = 24;
4336 else if (s->size >= 256)
4337 nr_objects = 52;
4338 else
4339 nr_objects = 120;
4340
4341 slub_set_cpu_partial(s, nr_objects);
4342#endif
4343}
4344
4345/*
4346 * calculate_sizes() determines the order and the distribution of data within
4347 * a slab object.
4348 */
4349static int calculate_sizes(struct kmem_cache *s)
4350{
4351 slab_flags_t flags = s->flags;
4352 unsigned int size = s->object_size;
4353 unsigned int order;
4354
4355 /*
4356 * Round up object size to the next word boundary. We can only
4357 * place the free pointer at word boundaries and this determines
4358 * the possible location of the free pointer.
4359 */
4360 size = ALIGN(size, sizeof(void *));
4361
4362#ifdef CONFIG_SLUB_DEBUG
4363 /*
4364 * Determine if we can poison the object itself. If the user of
4365 * the slab may touch the object after free or before allocation
4366 * then we should never poison the object itself.
4367 */
4368 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
4369 !s->ctor)
4370 s->flags |= __OBJECT_POISON;
4371 else
4372 s->flags &= ~__OBJECT_POISON;
4373
4374
4375 /*
4376 * If we are Redzoning then check if there is some space between the
4377 * end of the object and the free pointer. If not then add an
4378 * additional word to have some bytes to store Redzone information.
4379 */
4380 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
4381 size += sizeof(void *);
4382#endif
4383
4384 /*
4385 * With that we have determined the number of bytes in actual use
4386 * by the object and redzoning.
4387 */
4388 s->inuse = size;
4389
4390 if (slub_debug_orig_size(s) ||
4391 (flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
4392 ((flags & SLAB_RED_ZONE) && s->object_size < sizeof(void *)) ||
4393 s->ctor) {
4394 /*
4395 * Relocate free pointer after the object if it is not
4396 * permitted to overwrite the first word of the object on
4397 * kmem_cache_free.
4398 *
4399 * This is the case if we do RCU, have a constructor or
4400 * destructor, are poisoning the objects, or are
4401 * redzoning an object smaller than sizeof(void *).
4402 *
4403 * The assumption that s->offset >= s->inuse means free
4404 * pointer is outside of the object is used in the
4405 * freeptr_outside_object() function. If that is no
4406 * longer true, the function needs to be modified.
4407 */
4408 s->offset = size;
4409 size += sizeof(void *);
4410 } else {
4411 /*
4412 * Store freelist pointer near middle of object to keep
4413 * it away from the edges of the object to avoid small
4414 * sized over/underflows from neighboring allocations.
4415 */
4416 s->offset = ALIGN_DOWN(s->object_size / 2, sizeof(void *));
4417 }
4418
4419#ifdef CONFIG_SLUB_DEBUG
4420 if (flags & SLAB_STORE_USER) {
4421 /*
4422 * Need to store information about allocs and frees after
4423 * the object.
4424 */
4425 size += 2 * sizeof(struct track);
4426
4427 /* Save the original kmalloc request size */
4428 if (flags & SLAB_KMALLOC)
4429 size += sizeof(unsigned int);
4430 }
4431#endif
4432
4433 kasan_cache_create(s, &size, &s->flags);
4434#ifdef CONFIG_SLUB_DEBUG
4435 if (flags & SLAB_RED_ZONE) {
4436 /*
4437 * Add some empty padding so that we can catch
4438 * overwrites from earlier objects rather than let
4439 * tracking information or the free pointer be
4440 * corrupted if a user writes before the start
4441 * of the object.
4442 */
4443 size += sizeof(void *);
4444
4445 s->red_left_pad = sizeof(void *);
4446 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
4447 size += s->red_left_pad;
4448 }
4449#endif
4450
4451 /*
4452 * SLUB stores one object immediately after another beginning from
4453 * offset 0. In order to align the objects we have to simply size
4454 * each object to conform to the alignment.
4455 */
4456 size = ALIGN(size, s->align);
4457 s->size = size;
4458 s->reciprocal_size = reciprocal_value(size);
4459 order = calculate_order(size);
4460
4461 if ((int)order < 0)
4462 return 0;
4463
4464 s->allocflags = 0;
4465 if (order)
4466 s->allocflags |= __GFP_COMP;
4467
4468 if (s->flags & SLAB_CACHE_DMA)
4469 s->allocflags |= GFP_DMA;
4470
4471 if (s->flags & SLAB_CACHE_DMA32)
4472 s->allocflags |= GFP_DMA32;
4473
4474 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4475 s->allocflags |= __GFP_RECLAIMABLE;
4476
4477 /*
4478 * Determine the number of objects per slab
4479 */
4480 s->oo = oo_make(order, size);
4481 s->min = oo_make(get_order(size), size);
4482
4483 return !!oo_objects(s->oo);
4484}
4485
4486static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
4487{
4488 s->flags = kmem_cache_flags(s->size, flags, s->name);
4489#ifdef CONFIG_SLAB_FREELIST_HARDENED
4490 s->random = get_random_long();
4491#endif
4492
4493 if (!calculate_sizes(s))
4494 goto error;
4495 if (disable_higher_order_debug) {
4496 /*
4497 * Disable debugging flags that store metadata if the min slab
4498 * order increased.
4499 */
4500 if (get_order(s->size) > get_order(s->object_size)) {
4501 s->flags &= ~DEBUG_METADATA_FLAGS;
4502 s->offset = 0;
4503 if (!calculate_sizes(s))
4504 goto error;
4505 }
4506 }
4507
4508#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
4509 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
4510 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
4511 /* Enable fast mode */
4512 s->flags |= __CMPXCHG_DOUBLE;
4513#endif
4514
4515 /*
4516 * The larger the object size is, the more slabs we want on the partial
4517 * list to avoid pounding the page allocator excessively.
4518 */
4519 s->min_partial = min_t(unsigned long, MAX_PARTIAL, ilog2(s->size) / 2);
4520 s->min_partial = max_t(unsigned long, MIN_PARTIAL, s->min_partial);
4521
4522 set_cpu_partial(s);
4523
4524#ifdef CONFIG_NUMA
4525 s->remote_node_defrag_ratio = 1000;
4526#endif
4527
4528 /* Initialize the pre-computed randomized freelist if slab is up */
4529 if (slab_state >= UP) {
4530 if (init_cache_random_seq(s))
4531 goto error;
4532 }
4533
4534 if (!init_kmem_cache_nodes(s))
4535 goto error;
4536
4537 if (alloc_kmem_cache_cpus(s))
4538 return 0;
4539
4540error:
4541 __kmem_cache_release(s);
4542 return -EINVAL;
4543}
4544
4545static void list_slab_objects(struct kmem_cache *s, struct slab *slab,
4546 const char *text)
4547{
4548#ifdef CONFIG_SLUB_DEBUG
4549 void *addr = slab_address(slab);
4550 void *p;
4551
4552 slab_err(s, slab, text, s->name);
4553
4554 spin_lock(&object_map_lock);
4555 __fill_map(object_map, s, slab);
4556
4557 for_each_object(p, s, addr, slab->objects) {
4558
4559 if (!test_bit(__obj_to_index(s, addr, p), object_map)) {
4560 pr_err("Object 0x%p @offset=%tu\n", p, p - addr);
4561 print_tracking(s, p);
4562 }
4563 }
4564 spin_unlock(&object_map_lock);
4565#endif
4566}
4567
4568/*
4569 * Attempt to free all partial slabs on a node.
4570 * This is called from __kmem_cache_shutdown(). We must take list_lock
4571 * because sysfs file might still access partial list after the shutdowning.
4572 */
4573static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
4574{
4575 LIST_HEAD(discard);
4576 struct slab *slab, *h;
4577
4578 BUG_ON(irqs_disabled());
4579 spin_lock_irq(&n->list_lock);
4580 list_for_each_entry_safe(slab, h, &n->partial, slab_list) {
4581 if (!slab->inuse) {
4582 remove_partial(n, slab);
4583 list_add(&slab->slab_list, &discard);
4584 } else {
4585 list_slab_objects(s, slab,
4586 "Objects remaining in %s on __kmem_cache_shutdown()");
4587 }
4588 }
4589 spin_unlock_irq(&n->list_lock);
4590
4591 list_for_each_entry_safe(slab, h, &discard, slab_list)
4592 discard_slab(s, slab);
4593}
4594
4595bool __kmem_cache_empty(struct kmem_cache *s)
4596{
4597 int node;
4598 struct kmem_cache_node *n;
4599
4600 for_each_kmem_cache_node(s, node, n)
4601 if (n->nr_partial || slabs_node(s, node))
4602 return false;
4603 return true;
4604}
4605
4606/*
4607 * Release all resources used by a slab cache.
4608 */
4609int __kmem_cache_shutdown(struct kmem_cache *s)
4610{
4611 int node;
4612 struct kmem_cache_node *n;
4613
4614 flush_all_cpus_locked(s);
4615 /* Attempt to free all objects */
4616 for_each_kmem_cache_node(s, node, n) {
4617 free_partial(s, n);
4618 if (n->nr_partial || slabs_node(s, node))
4619 return 1;
4620 }
4621 return 0;
4622}
4623
4624#ifdef CONFIG_PRINTK
4625void __kmem_obj_info(struct kmem_obj_info *kpp, void *object, struct slab *slab)
4626{
4627 void *base;
4628 int __maybe_unused i;
4629 unsigned int objnr;
4630 void *objp;
4631 void *objp0;
4632 struct kmem_cache *s = slab->slab_cache;
4633 struct track __maybe_unused *trackp;
4634
4635 kpp->kp_ptr = object;
4636 kpp->kp_slab = slab;
4637 kpp->kp_slab_cache = s;
4638 base = slab_address(slab);
4639 objp0 = kasan_reset_tag(object);
4640#ifdef CONFIG_SLUB_DEBUG
4641 objp = restore_red_left(s, objp0);
4642#else
4643 objp = objp0;
4644#endif
4645 objnr = obj_to_index(s, slab, objp);
4646 kpp->kp_data_offset = (unsigned long)((char *)objp0 - (char *)objp);
4647 objp = base + s->size * objnr;
4648 kpp->kp_objp = objp;
4649 if (WARN_ON_ONCE(objp < base || objp >= base + slab->objects * s->size
4650 || (objp - base) % s->size) ||
4651 !(s->flags & SLAB_STORE_USER))
4652 return;
4653#ifdef CONFIG_SLUB_DEBUG
4654 objp = fixup_red_left(s, objp);
4655 trackp = get_track(s, objp, TRACK_ALLOC);
4656 kpp->kp_ret = (void *)trackp->addr;
4657#ifdef CONFIG_STACKDEPOT
4658 {
4659 depot_stack_handle_t handle;
4660 unsigned long *entries;
4661 unsigned int nr_entries;
4662
4663 handle = READ_ONCE(trackp->handle);
4664 if (handle) {
4665 nr_entries = stack_depot_fetch(handle, &entries);
4666 for (i = 0; i < KS_ADDRS_COUNT && i < nr_entries; i++)
4667 kpp->kp_stack[i] = (void *)entries[i];
4668 }
4669
4670 trackp = get_track(s, objp, TRACK_FREE);
4671 handle = READ_ONCE(trackp->handle);
4672 if (handle) {
4673 nr_entries = stack_depot_fetch(handle, &entries);
4674 for (i = 0; i < KS_ADDRS_COUNT && i < nr_entries; i++)
4675 kpp->kp_free_stack[i] = (void *)entries[i];
4676 }
4677 }
4678#endif
4679#endif
4680}
4681#endif
4682
4683/********************************************************************
4684 * Kmalloc subsystem
4685 *******************************************************************/
4686
4687static int __init setup_slub_min_order(char *str)
4688{
4689 get_option(&str, (int *)&slub_min_order);
4690
4691 return 1;
4692}
4693
4694__setup("slub_min_order=", setup_slub_min_order);
4695
4696static int __init setup_slub_max_order(char *str)
4697{
4698 get_option(&str, (int *)&slub_max_order);
4699 slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1);
4700
4701 return 1;
4702}
4703
4704__setup("slub_max_order=", setup_slub_max_order);
4705
4706static int __init setup_slub_min_objects(char *str)
4707{
4708 get_option(&str, (int *)&slub_min_objects);
4709
4710 return 1;
4711}
4712
4713__setup("slub_min_objects=", setup_slub_min_objects);
4714
4715#ifdef CONFIG_HARDENED_USERCOPY
4716/*
4717 * Rejects incorrectly sized objects and objects that are to be copied
4718 * to/from userspace but do not fall entirely within the containing slab
4719 * cache's usercopy region.
4720 *
4721 * Returns NULL if check passes, otherwise const char * to name of cache
4722 * to indicate an error.
4723 */
4724void __check_heap_object(const void *ptr, unsigned long n,
4725 const struct slab *slab, bool to_user)
4726{
4727 struct kmem_cache *s;
4728 unsigned int offset;
4729 bool is_kfence = is_kfence_address(ptr);
4730
4731 ptr = kasan_reset_tag(ptr);
4732
4733 /* Find object and usable object size. */
4734 s = slab->slab_cache;
4735
4736 /* Reject impossible pointers. */
4737 if (ptr < slab_address(slab))
4738 usercopy_abort("SLUB object not in SLUB page?!", NULL,
4739 to_user, 0, n);
4740
4741 /* Find offset within object. */
4742 if (is_kfence)
4743 offset = ptr - kfence_object_start(ptr);
4744 else
4745 offset = (ptr - slab_address(slab)) % s->size;
4746
4747 /* Adjust for redzone and reject if within the redzone. */
4748 if (!is_kfence && kmem_cache_debug_flags(s, SLAB_RED_ZONE)) {
4749 if (offset < s->red_left_pad)
4750 usercopy_abort("SLUB object in left red zone",
4751 s->name, to_user, offset, n);
4752 offset -= s->red_left_pad;
4753 }
4754
4755 /* Allow address range falling entirely within usercopy region. */
4756 if (offset >= s->useroffset &&
4757 offset - s->useroffset <= s->usersize &&
4758 n <= s->useroffset - offset + s->usersize)
4759 return;
4760
4761 usercopy_abort("SLUB object", s->name, to_user, offset, n);
4762}
4763#endif /* CONFIG_HARDENED_USERCOPY */
4764
4765#define SHRINK_PROMOTE_MAX 32
4766
4767/*
4768 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
4769 * up most to the head of the partial lists. New allocations will then
4770 * fill those up and thus they can be removed from the partial lists.
4771 *
4772 * The slabs with the least items are placed last. This results in them
4773 * being allocated from last increasing the chance that the last objects
4774 * are freed in them.
4775 */
4776static int __kmem_cache_do_shrink(struct kmem_cache *s)
4777{
4778 int node;
4779 int i;
4780 struct kmem_cache_node *n;
4781 struct slab *slab;
4782 struct slab *t;
4783 struct list_head discard;
4784 struct list_head promote[SHRINK_PROMOTE_MAX];
4785 unsigned long flags;
4786 int ret = 0;
4787
4788 for_each_kmem_cache_node(s, node, n) {
4789 INIT_LIST_HEAD(&discard);
4790 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
4791 INIT_LIST_HEAD(promote + i);
4792
4793 spin_lock_irqsave(&n->list_lock, flags);
4794
4795 /*
4796 * Build lists of slabs to discard or promote.
4797 *
4798 * Note that concurrent frees may occur while we hold the
4799 * list_lock. slab->inuse here is the upper limit.
4800 */
4801 list_for_each_entry_safe(slab, t, &n->partial, slab_list) {
4802 int free = slab->objects - slab->inuse;
4803
4804 /* Do not reread slab->inuse */
4805 barrier();
4806
4807 /* We do not keep full slabs on the list */
4808 BUG_ON(free <= 0);
4809
4810 if (free == slab->objects) {
4811 list_move(&slab->slab_list, &discard);
4812 n->nr_partial--;
4813 dec_slabs_node(s, node, slab->objects);
4814 } else if (free <= SHRINK_PROMOTE_MAX)
4815 list_move(&slab->slab_list, promote + free - 1);
4816 }
4817
4818 /*
4819 * Promote the slabs filled up most to the head of the
4820 * partial list.
4821 */
4822 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
4823 list_splice(promote + i, &n->partial);
4824
4825 spin_unlock_irqrestore(&n->list_lock, flags);
4826
4827 /* Release empty slabs */
4828 list_for_each_entry_safe(slab, t, &discard, slab_list)
4829 free_slab(s, slab);
4830
4831 if (slabs_node(s, node))
4832 ret = 1;
4833 }
4834
4835 return ret;
4836}
4837
4838int __kmem_cache_shrink(struct kmem_cache *s)
4839{
4840 flush_all(s);
4841 return __kmem_cache_do_shrink(s);
4842}
4843
4844static int slab_mem_going_offline_callback(void *arg)
4845{
4846 struct kmem_cache *s;
4847
4848 mutex_lock(&slab_mutex);
4849 list_for_each_entry(s, &slab_caches, list) {
4850 flush_all_cpus_locked(s);
4851 __kmem_cache_do_shrink(s);
4852 }
4853 mutex_unlock(&slab_mutex);
4854
4855 return 0;
4856}
4857
4858static void slab_mem_offline_callback(void *arg)
4859{
4860 struct memory_notify *marg = arg;
4861 int offline_node;
4862
4863 offline_node = marg->status_change_nid_normal;
4864
4865 /*
4866 * If the node still has available memory. we need kmem_cache_node
4867 * for it yet.
4868 */
4869 if (offline_node < 0)
4870 return;
4871
4872 mutex_lock(&slab_mutex);
4873 node_clear(offline_node, slab_nodes);
4874 /*
4875 * We no longer free kmem_cache_node structures here, as it would be
4876 * racy with all get_node() users, and infeasible to protect them with
4877 * slab_mutex.
4878 */
4879 mutex_unlock(&slab_mutex);
4880}
4881
4882static int slab_mem_going_online_callback(void *arg)
4883{
4884 struct kmem_cache_node *n;
4885 struct kmem_cache *s;
4886 struct memory_notify *marg = arg;
4887 int nid = marg->status_change_nid_normal;
4888 int ret = 0;
4889
4890 /*
4891 * If the node's memory is already available, then kmem_cache_node is
4892 * already created. Nothing to do.
4893 */
4894 if (nid < 0)
4895 return 0;
4896
4897 /*
4898 * We are bringing a node online. No memory is available yet. We must
4899 * allocate a kmem_cache_node structure in order to bring the node
4900 * online.
4901 */
4902 mutex_lock(&slab_mutex);
4903 list_for_each_entry(s, &slab_caches, list) {
4904 /*
4905 * The structure may already exist if the node was previously
4906 * onlined and offlined.
4907 */
4908 if (get_node(s, nid))
4909 continue;
4910 /*
4911 * XXX: kmem_cache_alloc_node will fallback to other nodes
4912 * since memory is not yet available from the node that
4913 * is brought up.
4914 */
4915 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4916 if (!n) {
4917 ret = -ENOMEM;
4918 goto out;
4919 }
4920 init_kmem_cache_node(n);
4921 s->node[nid] = n;
4922 }
4923 /*
4924 * Any cache created after this point will also have kmem_cache_node
4925 * initialized for the new node.
4926 */
4927 node_set(nid, slab_nodes);
4928out:
4929 mutex_unlock(&slab_mutex);
4930 return ret;
4931}
4932
4933static int slab_memory_callback(struct notifier_block *self,
4934 unsigned long action, void *arg)
4935{
4936 int ret = 0;
4937
4938 switch (action) {
4939 case MEM_GOING_ONLINE:
4940 ret = slab_mem_going_online_callback(arg);
4941 break;
4942 case MEM_GOING_OFFLINE:
4943 ret = slab_mem_going_offline_callback(arg);
4944 break;
4945 case MEM_OFFLINE:
4946 case MEM_CANCEL_ONLINE:
4947 slab_mem_offline_callback(arg);
4948 break;
4949 case MEM_ONLINE:
4950 case MEM_CANCEL_OFFLINE:
4951 break;
4952 }
4953 if (ret)
4954 ret = notifier_from_errno(ret);
4955 else
4956 ret = NOTIFY_OK;
4957 return ret;
4958}
4959
4960/********************************************************************
4961 * Basic setup of slabs
4962 *******************************************************************/
4963
4964/*
4965 * Used for early kmem_cache structures that were allocated using
4966 * the page allocator. Allocate them properly then fix up the pointers
4967 * that may be pointing to the wrong kmem_cache structure.
4968 */
4969
4970static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4971{
4972 int node;
4973 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4974 struct kmem_cache_node *n;
4975
4976 memcpy(s, static_cache, kmem_cache->object_size);
4977
4978 /*
4979 * This runs very early, and only the boot processor is supposed to be
4980 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4981 * IPIs around.
4982 */
4983 __flush_cpu_slab(s, smp_processor_id());
4984 for_each_kmem_cache_node(s, node, n) {
4985 struct slab *p;
4986
4987 list_for_each_entry(p, &n->partial, slab_list)
4988 p->slab_cache = s;
4989
4990#ifdef CONFIG_SLUB_DEBUG
4991 list_for_each_entry(p, &n->full, slab_list)
4992 p->slab_cache = s;
4993#endif
4994 }
4995 list_add(&s->list, &slab_caches);
4996 return s;
4997}
4998
4999void __init kmem_cache_init(void)
5000{
5001 static __initdata struct kmem_cache boot_kmem_cache,
5002 boot_kmem_cache_node;
5003 int node;
5004
5005 if (debug_guardpage_minorder())
5006 slub_max_order = 0;
5007
5008 /* Print slub debugging pointers without hashing */
5009 if (__slub_debug_enabled())
5010 no_hash_pointers_enable(NULL);
5011
5012 kmem_cache_node = &boot_kmem_cache_node;
5013 kmem_cache = &boot_kmem_cache;
5014
5015 /*
5016 * Initialize the nodemask for which we will allocate per node
5017 * structures. Here we don't need taking slab_mutex yet.
5018 */
5019 for_each_node_state(node, N_NORMAL_MEMORY)
5020 node_set(node, slab_nodes);
5021
5022 create_boot_cache(kmem_cache_node, "kmem_cache_node",
5023 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
5024
5025 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
5026
5027 /* Able to allocate the per node structures */
5028 slab_state = PARTIAL;
5029
5030 create_boot_cache(kmem_cache, "kmem_cache",
5031 offsetof(struct kmem_cache, node) +
5032 nr_node_ids * sizeof(struct kmem_cache_node *),
5033 SLAB_HWCACHE_ALIGN, 0, 0);
5034
5035 kmem_cache = bootstrap(&boot_kmem_cache);
5036 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
5037
5038 /* Now we can use the kmem_cache to allocate kmalloc slabs */
5039 setup_kmalloc_cache_index_table();
5040 create_kmalloc_caches(0);
5041
5042 /* Setup random freelists for each cache */
5043 init_freelist_randomization();
5044
5045 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
5046 slub_cpu_dead);
5047
5048 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
5049 cache_line_size(),
5050 slub_min_order, slub_max_order, slub_min_objects,
5051 nr_cpu_ids, nr_node_ids);
5052}
5053
5054void __init kmem_cache_init_late(void)
5055{
5056#ifndef CONFIG_SLUB_TINY
5057 flushwq = alloc_workqueue("slub_flushwq", WQ_MEM_RECLAIM, 0);
5058 WARN_ON(!flushwq);
5059#endif
5060}
5061
5062struct kmem_cache *
5063__kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
5064 slab_flags_t flags, void (*ctor)(void *))
5065{
5066 struct kmem_cache *s;
5067
5068 s = find_mergeable(size, align, flags, name, ctor);
5069 if (s) {
5070 if (sysfs_slab_alias(s, name))
5071 return NULL;
5072
5073 s->refcount++;
5074
5075 /*
5076 * Adjust the object sizes so that we clear
5077 * the complete object on kzalloc.
5078 */
5079 s->object_size = max(s->object_size, size);
5080 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
5081 }
5082
5083 return s;
5084}
5085
5086int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
5087{
5088 int err;
5089
5090 err = kmem_cache_open(s, flags);
5091 if (err)
5092 return err;
5093
5094 /* Mutex is not taken during early boot */
5095 if (slab_state <= UP)
5096 return 0;
5097
5098 err = sysfs_slab_add(s);
5099 if (err) {
5100 __kmem_cache_release(s);
5101 return err;
5102 }
5103
5104 if (s->flags & SLAB_STORE_USER)
5105 debugfs_slab_add(s);
5106
5107 return 0;
5108}
5109
5110#ifdef SLAB_SUPPORTS_SYSFS
5111static int count_inuse(struct slab *slab)
5112{
5113 return slab->inuse;
5114}
5115
5116static int count_total(struct slab *slab)
5117{
5118 return slab->objects;
5119}
5120#endif
5121
5122#ifdef CONFIG_SLUB_DEBUG
5123static void validate_slab(struct kmem_cache *s, struct slab *slab,
5124 unsigned long *obj_map)
5125{
5126 void *p;
5127 void *addr = slab_address(slab);
5128
5129 if (!check_slab(s, slab) || !on_freelist(s, slab, NULL))
5130 return;
5131
5132 /* Now we know that a valid freelist exists */
5133 __fill_map(obj_map, s, slab);
5134 for_each_object(p, s, addr, slab->objects) {
5135 u8 val = test_bit(__obj_to_index(s, addr, p), obj_map) ?
5136 SLUB_RED_INACTIVE : SLUB_RED_ACTIVE;
5137
5138 if (!check_object(s, slab, p, val))
5139 break;
5140 }
5141}
5142
5143static int validate_slab_node(struct kmem_cache *s,
5144 struct kmem_cache_node *n, unsigned long *obj_map)
5145{
5146 unsigned long count = 0;
5147 struct slab *slab;
5148 unsigned long flags;
5149
5150 spin_lock_irqsave(&n->list_lock, flags);
5151
5152 list_for_each_entry(slab, &n->partial, slab_list) {
5153 validate_slab(s, slab, obj_map);
5154 count++;
5155 }
5156 if (count != n->nr_partial) {
5157 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
5158 s->name, count, n->nr_partial);
5159 slab_add_kunit_errors();
5160 }
5161
5162 if (!(s->flags & SLAB_STORE_USER))
5163 goto out;
5164
5165 list_for_each_entry(slab, &n->full, slab_list) {
5166 validate_slab(s, slab, obj_map);
5167 count++;
5168 }
5169 if (count != atomic_long_read(&n->nr_slabs)) {
5170 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
5171 s->name, count, atomic_long_read(&n->nr_slabs));
5172 slab_add_kunit_errors();
5173 }
5174
5175out:
5176 spin_unlock_irqrestore(&n->list_lock, flags);
5177 return count;
5178}
5179
5180long validate_slab_cache(struct kmem_cache *s)
5181{
5182 int node;
5183 unsigned long count = 0;
5184 struct kmem_cache_node *n;
5185 unsigned long *obj_map;
5186
5187 obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL);
5188 if (!obj_map)
5189 return -ENOMEM;
5190
5191 flush_all(s);
5192 for_each_kmem_cache_node(s, node, n)
5193 count += validate_slab_node(s, n, obj_map);
5194
5195 bitmap_free(obj_map);
5196
5197 return count;
5198}
5199EXPORT_SYMBOL(validate_slab_cache);
5200
5201#ifdef CONFIG_DEBUG_FS
5202/*
5203 * Generate lists of code addresses where slabcache objects are allocated
5204 * and freed.
5205 */
5206
5207struct location {
5208 depot_stack_handle_t handle;
5209 unsigned long count;
5210 unsigned long addr;
5211 unsigned long waste;
5212 long long sum_time;
5213 long min_time;
5214 long max_time;
5215 long min_pid;
5216 long max_pid;
5217 DECLARE_BITMAP(cpus, NR_CPUS);
5218 nodemask_t nodes;
5219};
5220
5221struct loc_track {
5222 unsigned long max;
5223 unsigned long count;
5224 struct location *loc;
5225 loff_t idx;
5226};
5227
5228static struct dentry *slab_debugfs_root;
5229
5230static void free_loc_track(struct loc_track *t)
5231{
5232 if (t->max)
5233 free_pages((unsigned long)t->loc,
5234 get_order(sizeof(struct location) * t->max));
5235}
5236
5237static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
5238{
5239 struct location *l;
5240 int order;
5241
5242 order = get_order(sizeof(struct location) * max);
5243
5244 l = (void *)__get_free_pages(flags, order);
5245 if (!l)
5246 return 0;
5247
5248 if (t->count) {
5249 memcpy(l, t->loc, sizeof(struct location) * t->count);
5250 free_loc_track(t);
5251 }
5252 t->max = max;
5253 t->loc = l;
5254 return 1;
5255}
5256
5257static int add_location(struct loc_track *t, struct kmem_cache *s,
5258 const struct track *track,
5259 unsigned int orig_size)
5260{
5261 long start, end, pos;
5262 struct location *l;
5263 unsigned long caddr, chandle, cwaste;
5264 unsigned long age = jiffies - track->when;
5265 depot_stack_handle_t handle = 0;
5266 unsigned int waste = s->object_size - orig_size;
5267
5268#ifdef CONFIG_STACKDEPOT
5269 handle = READ_ONCE(track->handle);
5270#endif
5271 start = -1;
5272 end = t->count;
5273
5274 for ( ; ; ) {
5275 pos = start + (end - start + 1) / 2;
5276
5277 /*
5278 * There is nothing at "end". If we end up there
5279 * we need to add something to before end.
5280 */
5281 if (pos == end)
5282 break;
5283
5284 l = &t->loc[pos];
5285 caddr = l->addr;
5286 chandle = l->handle;
5287 cwaste = l->waste;
5288 if ((track->addr == caddr) && (handle == chandle) &&
5289 (waste == cwaste)) {
5290
5291 l->count++;
5292 if (track->when) {
5293 l->sum_time += age;
5294 if (age < l->min_time)
5295 l->min_time = age;
5296 if (age > l->max_time)
5297 l->max_time = age;
5298
5299 if (track->pid < l->min_pid)
5300 l->min_pid = track->pid;
5301 if (track->pid > l->max_pid)
5302 l->max_pid = track->pid;
5303
5304 cpumask_set_cpu(track->cpu,
5305 to_cpumask(l->cpus));
5306 }
5307 node_set(page_to_nid(virt_to_page(track)), l->nodes);
5308 return 1;
5309 }
5310
5311 if (track->addr < caddr)
5312 end = pos;
5313 else if (track->addr == caddr && handle < chandle)
5314 end = pos;
5315 else if (track->addr == caddr && handle == chandle &&
5316 waste < cwaste)
5317 end = pos;
5318 else
5319 start = pos;
5320 }
5321
5322 /*
5323 * Not found. Insert new tracking element.
5324 */
5325 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
5326 return 0;
5327
5328 l = t->loc + pos;
5329 if (pos < t->count)
5330 memmove(l + 1, l,
5331 (t->count - pos) * sizeof(struct location));
5332 t->count++;
5333 l->count = 1;
5334 l->addr = track->addr;
5335 l->sum_time = age;
5336 l->min_time = age;
5337 l->max_time = age;
5338 l->min_pid = track->pid;
5339 l->max_pid = track->pid;
5340 l->handle = handle;
5341 l->waste = waste;
5342 cpumask_clear(to_cpumask(l->cpus));
5343 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
5344 nodes_clear(l->nodes);
5345 node_set(page_to_nid(virt_to_page(track)), l->nodes);
5346 return 1;
5347}
5348
5349static void process_slab(struct loc_track *t, struct kmem_cache *s,
5350 struct slab *slab, enum track_item alloc,
5351 unsigned long *obj_map)
5352{
5353 void *addr = slab_address(slab);
5354 bool is_alloc = (alloc == TRACK_ALLOC);
5355 void *p;
5356
5357 __fill_map(obj_map, s, slab);
5358
5359 for_each_object(p, s, addr, slab->objects)
5360 if (!test_bit(__obj_to_index(s, addr, p), obj_map))
5361 add_location(t, s, get_track(s, p, alloc),
5362 is_alloc ? get_orig_size(s, p) :
5363 s->object_size);
5364}
5365#endif /* CONFIG_DEBUG_FS */
5366#endif /* CONFIG_SLUB_DEBUG */
5367
5368#ifdef SLAB_SUPPORTS_SYSFS
5369enum slab_stat_type {
5370 SL_ALL, /* All slabs */
5371 SL_PARTIAL, /* Only partially allocated slabs */
5372 SL_CPU, /* Only slabs used for cpu caches */
5373 SL_OBJECTS, /* Determine allocated objects not slabs */
5374 SL_TOTAL /* Determine object capacity not slabs */
5375};
5376
5377#define SO_ALL (1 << SL_ALL)
5378#define SO_PARTIAL (1 << SL_PARTIAL)
5379#define SO_CPU (1 << SL_CPU)
5380#define SO_OBJECTS (1 << SL_OBJECTS)
5381#define SO_TOTAL (1 << SL_TOTAL)
5382
5383static ssize_t show_slab_objects(struct kmem_cache *s,
5384 char *buf, unsigned long flags)
5385{
5386 unsigned long total = 0;
5387 int node;
5388 int x;
5389 unsigned long *nodes;
5390 int len = 0;
5391
5392 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
5393 if (!nodes)
5394 return -ENOMEM;
5395
5396 if (flags & SO_CPU) {
5397 int cpu;
5398
5399 for_each_possible_cpu(cpu) {
5400 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
5401 cpu);
5402 int node;
5403 struct slab *slab;
5404
5405 slab = READ_ONCE(c->slab);
5406 if (!slab)
5407 continue;
5408
5409 node = slab_nid(slab);
5410 if (flags & SO_TOTAL)
5411 x = slab->objects;
5412 else if (flags & SO_OBJECTS)
5413 x = slab->inuse;
5414 else
5415 x = 1;
5416
5417 total += x;
5418 nodes[node] += x;
5419
5420#ifdef CONFIG_SLUB_CPU_PARTIAL
5421 slab = slub_percpu_partial_read_once(c);
5422 if (slab) {
5423 node = slab_nid(slab);
5424 if (flags & SO_TOTAL)
5425 WARN_ON_ONCE(1);
5426 else if (flags & SO_OBJECTS)
5427 WARN_ON_ONCE(1);
5428 else
5429 x = slab->slabs;
5430 total += x;
5431 nodes[node] += x;
5432 }
5433#endif
5434 }
5435 }
5436
5437 /*
5438 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
5439 * already held which will conflict with an existing lock order:
5440 *
5441 * mem_hotplug_lock->slab_mutex->kernfs_mutex
5442 *
5443 * We don't really need mem_hotplug_lock (to hold off
5444 * slab_mem_going_offline_callback) here because slab's memory hot
5445 * unplug code doesn't destroy the kmem_cache->node[] data.
5446 */
5447
5448#ifdef CONFIG_SLUB_DEBUG
5449 if (flags & SO_ALL) {
5450 struct kmem_cache_node *n;
5451
5452 for_each_kmem_cache_node(s, node, n) {
5453
5454 if (flags & SO_TOTAL)
5455 x = atomic_long_read(&n->total_objects);
5456 else if (flags & SO_OBJECTS)
5457 x = atomic_long_read(&n->total_objects) -
5458 count_partial(n, count_free);
5459 else
5460 x = atomic_long_read(&n->nr_slabs);
5461 total += x;
5462 nodes[node] += x;
5463 }
5464
5465 } else
5466#endif
5467 if (flags & SO_PARTIAL) {
5468 struct kmem_cache_node *n;
5469
5470 for_each_kmem_cache_node(s, node, n) {
5471 if (flags & SO_TOTAL)
5472 x = count_partial(n, count_total);
5473 else if (flags & SO_OBJECTS)
5474 x = count_partial(n, count_inuse);
5475 else
5476 x = n->nr_partial;
5477 total += x;
5478 nodes[node] += x;
5479 }
5480 }
5481
5482 len += sysfs_emit_at(buf, len, "%lu", total);
5483#ifdef CONFIG_NUMA
5484 for (node = 0; node < nr_node_ids; node++) {
5485 if (nodes[node])
5486 len += sysfs_emit_at(buf, len, " N%d=%lu",
5487 node, nodes[node]);
5488 }
5489#endif
5490 len += sysfs_emit_at(buf, len, "\n");
5491 kfree(nodes);
5492
5493 return len;
5494}
5495
5496#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
5497#define to_slab(n) container_of(n, struct kmem_cache, kobj)
5498
5499struct slab_attribute {
5500 struct attribute attr;
5501 ssize_t (*show)(struct kmem_cache *s, char *buf);
5502 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
5503};
5504
5505#define SLAB_ATTR_RO(_name) \
5506 static struct slab_attribute _name##_attr = __ATTR_RO_MODE(_name, 0400)
5507
5508#define SLAB_ATTR(_name) \
5509 static struct slab_attribute _name##_attr = __ATTR_RW_MODE(_name, 0600)
5510
5511static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
5512{
5513 return sysfs_emit(buf, "%u\n", s->size);
5514}
5515SLAB_ATTR_RO(slab_size);
5516
5517static ssize_t align_show(struct kmem_cache *s, char *buf)
5518{
5519 return sysfs_emit(buf, "%u\n", s->align);
5520}
5521SLAB_ATTR_RO(align);
5522
5523static ssize_t object_size_show(struct kmem_cache *s, char *buf)
5524{
5525 return sysfs_emit(buf, "%u\n", s->object_size);
5526}
5527SLAB_ATTR_RO(object_size);
5528
5529static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
5530{
5531 return sysfs_emit(buf, "%u\n", oo_objects(s->oo));
5532}
5533SLAB_ATTR_RO(objs_per_slab);
5534
5535static ssize_t order_show(struct kmem_cache *s, char *buf)
5536{
5537 return sysfs_emit(buf, "%u\n", oo_order(s->oo));
5538}
5539SLAB_ATTR_RO(order);
5540
5541static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
5542{
5543 return sysfs_emit(buf, "%lu\n", s->min_partial);
5544}
5545
5546static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
5547 size_t length)
5548{
5549 unsigned long min;
5550 int err;
5551
5552 err = kstrtoul(buf, 10, &min);
5553 if (err)
5554 return err;
5555
5556 s->min_partial = min;
5557 return length;
5558}
5559SLAB_ATTR(min_partial);
5560
5561static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
5562{
5563 unsigned int nr_partial = 0;
5564#ifdef CONFIG_SLUB_CPU_PARTIAL
5565 nr_partial = s->cpu_partial;
5566#endif
5567
5568 return sysfs_emit(buf, "%u\n", nr_partial);
5569}
5570
5571static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
5572 size_t length)
5573{
5574 unsigned int objects;
5575 int err;
5576
5577 err = kstrtouint(buf, 10, &objects);
5578 if (err)
5579 return err;
5580 if (objects && !kmem_cache_has_cpu_partial(s))
5581 return -EINVAL;
5582
5583 slub_set_cpu_partial(s, objects);
5584 flush_all(s);
5585 return length;
5586}
5587SLAB_ATTR(cpu_partial);
5588
5589static ssize_t ctor_show(struct kmem_cache *s, char *buf)
5590{
5591 if (!s->ctor)
5592 return 0;
5593 return sysfs_emit(buf, "%pS\n", s->ctor);
5594}
5595SLAB_ATTR_RO(ctor);
5596
5597static ssize_t aliases_show(struct kmem_cache *s, char *buf)
5598{
5599 return sysfs_emit(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
5600}
5601SLAB_ATTR_RO(aliases);
5602
5603static ssize_t partial_show(struct kmem_cache *s, char *buf)
5604{
5605 return show_slab_objects(s, buf, SO_PARTIAL);
5606}
5607SLAB_ATTR_RO(partial);
5608
5609static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
5610{
5611 return show_slab_objects(s, buf, SO_CPU);
5612}
5613SLAB_ATTR_RO(cpu_slabs);
5614
5615static ssize_t objects_show(struct kmem_cache *s, char *buf)
5616{
5617 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5618}
5619SLAB_ATTR_RO(objects);
5620
5621static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5622{
5623 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5624}
5625SLAB_ATTR_RO(objects_partial);
5626
5627static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5628{
5629 int objects = 0;
5630 int slabs = 0;
5631 int cpu __maybe_unused;
5632 int len = 0;
5633
5634#ifdef CONFIG_SLUB_CPU_PARTIAL
5635 for_each_online_cpu(cpu) {
5636 struct slab *slab;
5637
5638 slab = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5639
5640 if (slab)
5641 slabs += slab->slabs;
5642 }
5643#endif
5644
5645 /* Approximate half-full slabs, see slub_set_cpu_partial() */
5646 objects = (slabs * oo_objects(s->oo)) / 2;
5647 len += sysfs_emit_at(buf, len, "%d(%d)", objects, slabs);
5648
5649#if defined(CONFIG_SLUB_CPU_PARTIAL) && defined(CONFIG_SMP)
5650 for_each_online_cpu(cpu) {
5651 struct slab *slab;
5652
5653 slab = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5654 if (slab) {
5655 slabs = READ_ONCE(slab->slabs);
5656 objects = (slabs * oo_objects(s->oo)) / 2;
5657 len += sysfs_emit_at(buf, len, " C%d=%d(%d)",
5658 cpu, objects, slabs);
5659 }
5660 }
5661#endif
5662 len += sysfs_emit_at(buf, len, "\n");
5663
5664 return len;
5665}
5666SLAB_ATTR_RO(slabs_cpu_partial);
5667
5668static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5669{
5670 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5671}
5672SLAB_ATTR_RO(reclaim_account);
5673
5674static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5675{
5676 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5677}
5678SLAB_ATTR_RO(hwcache_align);
5679
5680#ifdef CONFIG_ZONE_DMA
5681static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5682{
5683 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5684}
5685SLAB_ATTR_RO(cache_dma);
5686#endif
5687
5688#ifdef CONFIG_HARDENED_USERCOPY
5689static ssize_t usersize_show(struct kmem_cache *s, char *buf)
5690{
5691 return sysfs_emit(buf, "%u\n", s->usersize);
5692}
5693SLAB_ATTR_RO(usersize);
5694#endif
5695
5696static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5697{
5698 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
5699}
5700SLAB_ATTR_RO(destroy_by_rcu);
5701
5702#ifdef CONFIG_SLUB_DEBUG
5703static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5704{
5705 return show_slab_objects(s, buf, SO_ALL);
5706}
5707SLAB_ATTR_RO(slabs);
5708
5709static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5710{
5711 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5712}
5713SLAB_ATTR_RO(total_objects);
5714
5715static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5716{
5717 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5718}
5719SLAB_ATTR_RO(sanity_checks);
5720
5721static ssize_t trace_show(struct kmem_cache *s, char *buf)
5722{
5723 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5724}
5725SLAB_ATTR_RO(trace);
5726
5727static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5728{
5729 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5730}
5731
5732SLAB_ATTR_RO(red_zone);
5733
5734static ssize_t poison_show(struct kmem_cache *s, char *buf)
5735{
5736 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_POISON));
5737}
5738
5739SLAB_ATTR_RO(poison);
5740
5741static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5742{
5743 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5744}
5745
5746SLAB_ATTR_RO(store_user);
5747
5748static ssize_t validate_show(struct kmem_cache *s, char *buf)
5749{
5750 return 0;
5751}
5752
5753static ssize_t validate_store(struct kmem_cache *s,
5754 const char *buf, size_t length)
5755{
5756 int ret = -EINVAL;
5757
5758 if (buf[0] == '1' && kmem_cache_debug(s)) {
5759 ret = validate_slab_cache(s);
5760 if (ret >= 0)
5761 ret = length;
5762 }
5763 return ret;
5764}
5765SLAB_ATTR(validate);
5766
5767#endif /* CONFIG_SLUB_DEBUG */
5768
5769#ifdef CONFIG_FAILSLAB
5770static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5771{
5772 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5773}
5774
5775static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
5776 size_t length)
5777{
5778 if (s->refcount > 1)
5779 return -EINVAL;
5780
5781 if (buf[0] == '1')
5782 WRITE_ONCE(s->flags, s->flags | SLAB_FAILSLAB);
5783 else
5784 WRITE_ONCE(s->flags, s->flags & ~SLAB_FAILSLAB);
5785
5786 return length;
5787}
5788SLAB_ATTR(failslab);
5789#endif
5790
5791static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5792{
5793 return 0;
5794}
5795
5796static ssize_t shrink_store(struct kmem_cache *s,
5797 const char *buf, size_t length)
5798{
5799 if (buf[0] == '1')
5800 kmem_cache_shrink(s);
5801 else
5802 return -EINVAL;
5803 return length;
5804}
5805SLAB_ATTR(shrink);
5806
5807#ifdef CONFIG_NUMA
5808static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5809{
5810 return sysfs_emit(buf, "%u\n", s->remote_node_defrag_ratio / 10);
5811}
5812
5813static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5814 const char *buf, size_t length)
5815{
5816 unsigned int ratio;
5817 int err;
5818
5819 err = kstrtouint(buf, 10, &ratio);
5820 if (err)
5821 return err;
5822 if (ratio > 100)
5823 return -ERANGE;
5824
5825 s->remote_node_defrag_ratio = ratio * 10;
5826
5827 return length;
5828}
5829SLAB_ATTR(remote_node_defrag_ratio);
5830#endif
5831
5832#ifdef CONFIG_SLUB_STATS
5833static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5834{
5835 unsigned long sum = 0;
5836 int cpu;
5837 int len = 0;
5838 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
5839
5840 if (!data)
5841 return -ENOMEM;
5842
5843 for_each_online_cpu(cpu) {
5844 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5845
5846 data[cpu] = x;
5847 sum += x;
5848 }
5849
5850 len += sysfs_emit_at(buf, len, "%lu", sum);
5851
5852#ifdef CONFIG_SMP
5853 for_each_online_cpu(cpu) {
5854 if (data[cpu])
5855 len += sysfs_emit_at(buf, len, " C%d=%u",
5856 cpu, data[cpu]);
5857 }
5858#endif
5859 kfree(data);
5860 len += sysfs_emit_at(buf, len, "\n");
5861
5862 return len;
5863}
5864
5865static void clear_stat(struct kmem_cache *s, enum stat_item si)
5866{
5867 int cpu;
5868
5869 for_each_online_cpu(cpu)
5870 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5871}
5872
5873#define STAT_ATTR(si, text) \
5874static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5875{ \
5876 return show_stat(s, buf, si); \
5877} \
5878static ssize_t text##_store(struct kmem_cache *s, \
5879 const char *buf, size_t length) \
5880{ \
5881 if (buf[0] != '0') \
5882 return -EINVAL; \
5883 clear_stat(s, si); \
5884 return length; \
5885} \
5886SLAB_ATTR(text); \
5887
5888STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5889STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5890STAT_ATTR(FREE_FASTPATH, free_fastpath);
5891STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5892STAT_ATTR(FREE_FROZEN, free_frozen);
5893STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5894STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5895STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5896STAT_ATTR(ALLOC_SLAB, alloc_slab);
5897STAT_ATTR(ALLOC_REFILL, alloc_refill);
5898STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5899STAT_ATTR(FREE_SLAB, free_slab);
5900STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5901STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5902STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5903STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5904STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5905STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5906STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5907STAT_ATTR(ORDER_FALLBACK, order_fallback);
5908STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5909STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5910STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5911STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5912STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5913STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5914#endif /* CONFIG_SLUB_STATS */
5915
5916#ifdef CONFIG_KFENCE
5917static ssize_t skip_kfence_show(struct kmem_cache *s, char *buf)
5918{
5919 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_SKIP_KFENCE));
5920}
5921
5922static ssize_t skip_kfence_store(struct kmem_cache *s,
5923 const char *buf, size_t length)
5924{
5925 int ret = length;
5926
5927 if (buf[0] == '0')
5928 s->flags &= ~SLAB_SKIP_KFENCE;
5929 else if (buf[0] == '1')
5930 s->flags |= SLAB_SKIP_KFENCE;
5931 else
5932 ret = -EINVAL;
5933
5934 return ret;
5935}
5936SLAB_ATTR(skip_kfence);
5937#endif
5938
5939static struct attribute *slab_attrs[] = {
5940 &slab_size_attr.attr,
5941 &object_size_attr.attr,
5942 &objs_per_slab_attr.attr,
5943 &order_attr.attr,
5944 &min_partial_attr.attr,
5945 &cpu_partial_attr.attr,
5946 &objects_attr.attr,
5947 &objects_partial_attr.attr,
5948 &partial_attr.attr,
5949 &cpu_slabs_attr.attr,
5950 &ctor_attr.attr,
5951 &aliases_attr.attr,
5952 &align_attr.attr,
5953 &hwcache_align_attr.attr,
5954 &reclaim_account_attr.attr,
5955 &destroy_by_rcu_attr.attr,
5956 &shrink_attr.attr,
5957 &slabs_cpu_partial_attr.attr,
5958#ifdef CONFIG_SLUB_DEBUG
5959 &total_objects_attr.attr,
5960 &slabs_attr.attr,
5961 &sanity_checks_attr.attr,
5962 &trace_attr.attr,
5963 &red_zone_attr.attr,
5964 &poison_attr.attr,
5965 &store_user_attr.attr,
5966 &validate_attr.attr,
5967#endif
5968#ifdef CONFIG_ZONE_DMA
5969 &cache_dma_attr.attr,
5970#endif
5971#ifdef CONFIG_NUMA
5972 &remote_node_defrag_ratio_attr.attr,
5973#endif
5974#ifdef CONFIG_SLUB_STATS
5975 &alloc_fastpath_attr.attr,
5976 &alloc_slowpath_attr.attr,
5977 &free_fastpath_attr.attr,
5978 &free_slowpath_attr.attr,
5979 &free_frozen_attr.attr,
5980 &free_add_partial_attr.attr,
5981 &free_remove_partial_attr.attr,
5982 &alloc_from_partial_attr.attr,
5983 &alloc_slab_attr.attr,
5984 &alloc_refill_attr.attr,
5985 &alloc_node_mismatch_attr.attr,
5986 &free_slab_attr.attr,
5987 &cpuslab_flush_attr.attr,
5988 &deactivate_full_attr.attr,
5989 &deactivate_empty_attr.attr,
5990 &deactivate_to_head_attr.attr,
5991 &deactivate_to_tail_attr.attr,
5992 &deactivate_remote_frees_attr.attr,
5993 &deactivate_bypass_attr.attr,
5994 &order_fallback_attr.attr,
5995 &cmpxchg_double_fail_attr.attr,
5996 &cmpxchg_double_cpu_fail_attr.attr,
5997 &cpu_partial_alloc_attr.attr,
5998 &cpu_partial_free_attr.attr,
5999 &cpu_partial_node_attr.attr,
6000 &cpu_partial_drain_attr.attr,
6001#endif
6002#ifdef CONFIG_FAILSLAB
6003 &failslab_attr.attr,
6004#endif
6005#ifdef CONFIG_HARDENED_USERCOPY
6006 &usersize_attr.attr,
6007#endif
6008#ifdef CONFIG_KFENCE
6009 &skip_kfence_attr.attr,
6010#endif
6011
6012 NULL
6013};
6014
6015static const struct attribute_group slab_attr_group = {
6016 .attrs = slab_attrs,
6017};
6018
6019static ssize_t slab_attr_show(struct kobject *kobj,
6020 struct attribute *attr,
6021 char *buf)
6022{
6023 struct slab_attribute *attribute;
6024 struct kmem_cache *s;
6025
6026 attribute = to_slab_attr(attr);
6027 s = to_slab(kobj);
6028
6029 if (!attribute->show)
6030 return -EIO;
6031
6032 return attribute->show(s, buf);
6033}
6034
6035static ssize_t slab_attr_store(struct kobject *kobj,
6036 struct attribute *attr,
6037 const char *buf, size_t len)
6038{
6039 struct slab_attribute *attribute;
6040 struct kmem_cache *s;
6041
6042 attribute = to_slab_attr(attr);
6043 s = to_slab(kobj);
6044
6045 if (!attribute->store)
6046 return -EIO;
6047
6048 return attribute->store(s, buf, len);
6049}
6050
6051static void kmem_cache_release(struct kobject *k)
6052{
6053 slab_kmem_cache_release(to_slab(k));
6054}
6055
6056static const struct sysfs_ops slab_sysfs_ops = {
6057 .show = slab_attr_show,
6058 .store = slab_attr_store,
6059};
6060
6061static struct kobj_type slab_ktype = {
6062 .sysfs_ops = &slab_sysfs_ops,
6063 .release = kmem_cache_release,
6064};
6065
6066static struct kset *slab_kset;
6067
6068static inline struct kset *cache_kset(struct kmem_cache *s)
6069{
6070 return slab_kset;
6071}
6072
6073#define ID_STR_LENGTH 32
6074
6075/* Create a unique string id for a slab cache:
6076 *
6077 * Format :[flags-]size
6078 */
6079static char *create_unique_id(struct kmem_cache *s)
6080{
6081 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
6082 char *p = name;
6083
6084 if (!name)
6085 return ERR_PTR(-ENOMEM);
6086
6087 *p++ = ':';
6088 /*
6089 * First flags affecting slabcache operations. We will only
6090 * get here for aliasable slabs so we do not need to support
6091 * too many flags. The flags here must cover all flags that
6092 * are matched during merging to guarantee that the id is
6093 * unique.
6094 */
6095 if (s->flags & SLAB_CACHE_DMA)
6096 *p++ = 'd';
6097 if (s->flags & SLAB_CACHE_DMA32)
6098 *p++ = 'D';
6099 if (s->flags & SLAB_RECLAIM_ACCOUNT)
6100 *p++ = 'a';
6101 if (s->flags & SLAB_CONSISTENCY_CHECKS)
6102 *p++ = 'F';
6103 if (s->flags & SLAB_ACCOUNT)
6104 *p++ = 'A';
6105 if (p != name + 1)
6106 *p++ = '-';
6107 p += snprintf(p, ID_STR_LENGTH - (p - name), "%07u", s->size);
6108
6109 if (WARN_ON(p > name + ID_STR_LENGTH - 1)) {
6110 kfree(name);
6111 return ERR_PTR(-EINVAL);
6112 }
6113 kmsan_unpoison_memory(name, p - name);
6114 return name;
6115}
6116
6117static int sysfs_slab_add(struct kmem_cache *s)
6118{
6119 int err;
6120 const char *name;
6121 struct kset *kset = cache_kset(s);
6122 int unmergeable = slab_unmergeable(s);
6123
6124 if (!unmergeable && disable_higher_order_debug &&
6125 (slub_debug & DEBUG_METADATA_FLAGS))
6126 unmergeable = 1;
6127
6128 if (unmergeable) {
6129 /*
6130 * Slabcache can never be merged so we can use the name proper.
6131 * This is typically the case for debug situations. In that
6132 * case we can catch duplicate names easily.
6133 */
6134 sysfs_remove_link(&slab_kset->kobj, s->name);
6135 name = s->name;
6136 } else {
6137 /*
6138 * Create a unique name for the slab as a target
6139 * for the symlinks.
6140 */
6141 name = create_unique_id(s);
6142 if (IS_ERR(name))
6143 return PTR_ERR(name);
6144 }
6145
6146 s->kobj.kset = kset;
6147 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
6148 if (err)
6149 goto out;
6150
6151 err = sysfs_create_group(&s->kobj, &slab_attr_group);
6152 if (err)
6153 goto out_del_kobj;
6154
6155 if (!unmergeable) {
6156 /* Setup first alias */
6157 sysfs_slab_alias(s, s->name);
6158 }
6159out:
6160 if (!unmergeable)
6161 kfree(name);
6162 return err;
6163out_del_kobj:
6164 kobject_del(&s->kobj);
6165 goto out;
6166}
6167
6168void sysfs_slab_unlink(struct kmem_cache *s)
6169{
6170 if (slab_state >= FULL)
6171 kobject_del(&s->kobj);
6172}
6173
6174void sysfs_slab_release(struct kmem_cache *s)
6175{
6176 if (slab_state >= FULL)
6177 kobject_put(&s->kobj);
6178}
6179
6180/*
6181 * Need to buffer aliases during bootup until sysfs becomes
6182 * available lest we lose that information.
6183 */
6184struct saved_alias {
6185 struct kmem_cache *s;
6186 const char *name;
6187 struct saved_alias *next;
6188};
6189
6190static struct saved_alias *alias_list;
6191
6192static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
6193{
6194 struct saved_alias *al;
6195
6196 if (slab_state == FULL) {
6197 /*
6198 * If we have a leftover link then remove it.
6199 */
6200 sysfs_remove_link(&slab_kset->kobj, name);
6201 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
6202 }
6203
6204 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
6205 if (!al)
6206 return -ENOMEM;
6207
6208 al->s = s;
6209 al->name = name;
6210 al->next = alias_list;
6211 alias_list = al;
6212 kmsan_unpoison_memory(al, sizeof(*al));
6213 return 0;
6214}
6215
6216static int __init slab_sysfs_init(void)
6217{
6218 struct kmem_cache *s;
6219 int err;
6220
6221 mutex_lock(&slab_mutex);
6222
6223 slab_kset = kset_create_and_add("slab", NULL, kernel_kobj);
6224 if (!slab_kset) {
6225 mutex_unlock(&slab_mutex);
6226 pr_err("Cannot register slab subsystem.\n");
6227 return -ENOSYS;
6228 }
6229
6230 slab_state = FULL;
6231
6232 list_for_each_entry(s, &slab_caches, list) {
6233 err = sysfs_slab_add(s);
6234 if (err)
6235 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
6236 s->name);
6237 }
6238
6239 while (alias_list) {
6240 struct saved_alias *al = alias_list;
6241
6242 alias_list = alias_list->next;
6243 err = sysfs_slab_alias(al->s, al->name);
6244 if (err)
6245 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
6246 al->name);
6247 kfree(al);
6248 }
6249
6250 mutex_unlock(&slab_mutex);
6251 return 0;
6252}
6253late_initcall(slab_sysfs_init);
6254#endif /* SLAB_SUPPORTS_SYSFS */
6255
6256#if defined(CONFIG_SLUB_DEBUG) && defined(CONFIG_DEBUG_FS)
6257static int slab_debugfs_show(struct seq_file *seq, void *v)
6258{
6259 struct loc_track *t = seq->private;
6260 struct location *l;
6261 unsigned long idx;
6262
6263 idx = (unsigned long) t->idx;
6264 if (idx < t->count) {
6265 l = &t->loc[idx];
6266
6267 seq_printf(seq, "%7ld ", l->count);
6268
6269 if (l->addr)
6270 seq_printf(seq, "%pS", (void *)l->addr);
6271 else
6272 seq_puts(seq, "<not-available>");
6273
6274 if (l->waste)
6275 seq_printf(seq, " waste=%lu/%lu",
6276 l->count * l->waste, l->waste);
6277
6278 if (l->sum_time != l->min_time) {
6279 seq_printf(seq, " age=%ld/%llu/%ld",
6280 l->min_time, div_u64(l->sum_time, l->count),
6281 l->max_time);
6282 } else
6283 seq_printf(seq, " age=%ld", l->min_time);
6284
6285 if (l->min_pid != l->max_pid)
6286 seq_printf(seq, " pid=%ld-%ld", l->min_pid, l->max_pid);
6287 else
6288 seq_printf(seq, " pid=%ld",
6289 l->min_pid);
6290
6291 if (num_online_cpus() > 1 && !cpumask_empty(to_cpumask(l->cpus)))
6292 seq_printf(seq, " cpus=%*pbl",
6293 cpumask_pr_args(to_cpumask(l->cpus)));
6294
6295 if (nr_online_nodes > 1 && !nodes_empty(l->nodes))
6296 seq_printf(seq, " nodes=%*pbl",
6297 nodemask_pr_args(&l->nodes));
6298
6299#ifdef CONFIG_STACKDEPOT
6300 {
6301 depot_stack_handle_t handle;
6302 unsigned long *entries;
6303 unsigned int nr_entries, j;
6304
6305 handle = READ_ONCE(l->handle);
6306 if (handle) {
6307 nr_entries = stack_depot_fetch(handle, &entries);
6308 seq_puts(seq, "\n");
6309 for (j = 0; j < nr_entries; j++)
6310 seq_printf(seq, " %pS\n", (void *)entries[j]);
6311 }
6312 }
6313#endif
6314 seq_puts(seq, "\n");
6315 }
6316
6317 if (!idx && !t->count)
6318 seq_puts(seq, "No data\n");
6319
6320 return 0;
6321}
6322
6323static void slab_debugfs_stop(struct seq_file *seq, void *v)
6324{
6325}
6326
6327static void *slab_debugfs_next(struct seq_file *seq, void *v, loff_t *ppos)
6328{
6329 struct loc_track *t = seq->private;
6330
6331 t->idx = ++(*ppos);
6332 if (*ppos <= t->count)
6333 return ppos;
6334
6335 return NULL;
6336}
6337
6338static int cmp_loc_by_count(const void *a, const void *b, const void *data)
6339{
6340 struct location *loc1 = (struct location *)a;
6341 struct location *loc2 = (struct location *)b;
6342
6343 if (loc1->count > loc2->count)
6344 return -1;
6345 else
6346 return 1;
6347}
6348
6349static void *slab_debugfs_start(struct seq_file *seq, loff_t *ppos)
6350{
6351 struct loc_track *t = seq->private;
6352
6353 t->idx = *ppos;
6354 return ppos;
6355}
6356
6357static const struct seq_operations slab_debugfs_sops = {
6358 .start = slab_debugfs_start,
6359 .next = slab_debugfs_next,
6360 .stop = slab_debugfs_stop,
6361 .show = slab_debugfs_show,
6362};
6363
6364static int slab_debug_trace_open(struct inode *inode, struct file *filep)
6365{
6366
6367 struct kmem_cache_node *n;
6368 enum track_item alloc;
6369 int node;
6370 struct loc_track *t = __seq_open_private(filep, &slab_debugfs_sops,
6371 sizeof(struct loc_track));
6372 struct kmem_cache *s = file_inode(filep)->i_private;
6373 unsigned long *obj_map;
6374
6375 if (!t)
6376 return -ENOMEM;
6377
6378 obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL);
6379 if (!obj_map) {
6380 seq_release_private(inode, filep);
6381 return -ENOMEM;
6382 }
6383
6384 if (strcmp(filep->f_path.dentry->d_name.name, "alloc_traces") == 0)
6385 alloc = TRACK_ALLOC;
6386 else
6387 alloc = TRACK_FREE;
6388
6389 if (!alloc_loc_track(t, PAGE_SIZE / sizeof(struct location), GFP_KERNEL)) {
6390 bitmap_free(obj_map);
6391 seq_release_private(inode, filep);
6392 return -ENOMEM;
6393 }
6394
6395 for_each_kmem_cache_node(s, node, n) {
6396 unsigned long flags;
6397 struct slab *slab;
6398
6399 if (!atomic_long_read(&n->nr_slabs))
6400 continue;
6401
6402 spin_lock_irqsave(&n->list_lock, flags);
6403 list_for_each_entry(slab, &n->partial, slab_list)
6404 process_slab(t, s, slab, alloc, obj_map);
6405 list_for_each_entry(slab, &n->full, slab_list)
6406 process_slab(t, s, slab, alloc, obj_map);
6407 spin_unlock_irqrestore(&n->list_lock, flags);
6408 }
6409
6410 /* Sort locations by count */
6411 sort_r(t->loc, t->count, sizeof(struct location),
6412 cmp_loc_by_count, NULL, NULL);
6413
6414 bitmap_free(obj_map);
6415 return 0;
6416}
6417
6418static int slab_debug_trace_release(struct inode *inode, struct file *file)
6419{
6420 struct seq_file *seq = file->private_data;
6421 struct loc_track *t = seq->private;
6422
6423 free_loc_track(t);
6424 return seq_release_private(inode, file);
6425}
6426
6427static const struct file_operations slab_debugfs_fops = {
6428 .open = slab_debug_trace_open,
6429 .read = seq_read,
6430 .llseek = seq_lseek,
6431 .release = slab_debug_trace_release,
6432};
6433
6434static void debugfs_slab_add(struct kmem_cache *s)
6435{
6436 struct dentry *slab_cache_dir;
6437
6438 if (unlikely(!slab_debugfs_root))
6439 return;
6440
6441 slab_cache_dir = debugfs_create_dir(s->name, slab_debugfs_root);
6442
6443 debugfs_create_file("alloc_traces", 0400,
6444 slab_cache_dir, s, &slab_debugfs_fops);
6445
6446 debugfs_create_file("free_traces", 0400,
6447 slab_cache_dir, s, &slab_debugfs_fops);
6448}
6449
6450void debugfs_slab_release(struct kmem_cache *s)
6451{
6452 debugfs_remove_recursive(debugfs_lookup(s->name, slab_debugfs_root));
6453}
6454
6455static int __init slab_debugfs_init(void)
6456{
6457 struct kmem_cache *s;
6458
6459 slab_debugfs_root = debugfs_create_dir("slab", NULL);
6460
6461 list_for_each_entry(s, &slab_caches, list)
6462 if (s->flags & SLAB_STORE_USER)
6463 debugfs_slab_add(s);
6464
6465 return 0;
6466
6467}
6468__initcall(slab_debugfs_init);
6469#endif
6470/*
6471 * The /proc/slabinfo ABI
6472 */
6473#ifdef CONFIG_SLUB_DEBUG
6474void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
6475{
6476 unsigned long nr_slabs = 0;
6477 unsigned long nr_objs = 0;
6478 unsigned long nr_free = 0;
6479 int node;
6480 struct kmem_cache_node *n;
6481
6482 for_each_kmem_cache_node(s, node, n) {
6483 nr_slabs += node_nr_slabs(n);
6484 nr_objs += node_nr_objs(n);
6485 nr_free += count_partial(n, count_free);
6486 }
6487
6488 sinfo->active_objs = nr_objs - nr_free;
6489 sinfo->num_objs = nr_objs;
6490 sinfo->active_slabs = nr_slabs;
6491 sinfo->num_slabs = nr_slabs;
6492 sinfo->objects_per_slab = oo_objects(s->oo);
6493 sinfo->cache_order = oo_order(s->oo);
6494}
6495
6496void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
6497{
6498}
6499
6500ssize_t slabinfo_write(struct file *file, const char __user *buffer,
6501 size_t count, loff_t *ppos)
6502{
6503 return -EIO;
6504}
6505#endif /* CONFIG_SLUB_DEBUG */
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/kmemcheck.h>
24#include <linux/cpu.h>
25#include <linux/cpuset.h>
26#include <linux/mempolicy.h>
27#include <linux/ctype.h>
28#include <linux/debugobjects.h>
29#include <linux/kallsyms.h>
30#include <linux/memory.h>
31#include <linux/math64.h>
32#include <linux/fault-inject.h>
33#include <linux/stacktrace.h>
34#include <linux/prefetch.h>
35#include <linux/memcontrol.h>
36
37#include <trace/events/kmem.h>
38
39#include "internal.h"
40
41/*
42 * Lock order:
43 * 1. slab_mutex (Global Mutex)
44 * 2. node->list_lock
45 * 3. slab_lock(page) (Only on some arches and for debugging)
46 *
47 * slab_mutex
48 *
49 * The role of the slab_mutex is to protect the list of all the slabs
50 * and to synchronize major metadata changes to slab cache structures.
51 *
52 * The slab_lock is only used for debugging and on arches that do not
53 * have the ability to do a cmpxchg_double. It only protects the second
54 * double word in the page struct. Meaning
55 * A. page->freelist -> List of object free in a page
56 * B. page->counters -> Counters of objects
57 * C. page->frozen -> frozen state
58 *
59 * If a slab is frozen then it is exempt from list management. It is not
60 * on any list. The processor that froze the slab is the one who can
61 * perform list operations on the page. Other processors may put objects
62 * onto the freelist but the processor that froze the slab is the only
63 * one that can retrieve the objects from the page's freelist.
64 *
65 * The list_lock protects the partial and full list on each node and
66 * the partial slab counter. If taken then no new slabs may be added or
67 * removed from the lists nor make the number of partial slabs be modified.
68 * (Note that the total number of slabs is an atomic value that may be
69 * modified without taking the list lock).
70 *
71 * The list_lock is a centralized lock and thus we avoid taking it as
72 * much as possible. As long as SLUB does not have to handle partial
73 * slabs, operations can continue without any centralized lock. F.e.
74 * allocating a long series of objects that fill up slabs does not require
75 * the list lock.
76 * Interrupts are disabled during allocation and deallocation in order to
77 * make the slab allocator safe to use in the context of an irq. In addition
78 * interrupts are disabled to ensure that the processor does not change
79 * while handling per_cpu slabs, due to kernel preemption.
80 *
81 * SLUB assigns one slab for allocation to each processor.
82 * Allocations only occur from these slabs called cpu slabs.
83 *
84 * Slabs with free elements are kept on a partial list and during regular
85 * operations no list for full slabs is used. If an object in a full slab is
86 * freed then the slab will show up again on the partial lists.
87 * We track full slabs for debugging purposes though because otherwise we
88 * cannot scan all objects.
89 *
90 * Slabs are freed when they become empty. Teardown and setup is
91 * minimal so we rely on the page allocators per cpu caches for
92 * fast frees and allocs.
93 *
94 * Overloading of page flags that are otherwise used for LRU management.
95 *
96 * PageActive The slab is frozen and exempt from list processing.
97 * This means that the slab is dedicated to a purpose
98 * such as satisfying allocations for a specific
99 * processor. Objects may be freed in the slab while
100 * it is frozen but slab_free will then skip the usual
101 * list operations. It is up to the processor holding
102 * the slab to integrate the slab into the slab lists
103 * when the slab is no longer needed.
104 *
105 * One use of this flag is to mark slabs that are
106 * used for allocations. Then such a slab becomes a cpu
107 * slab. The cpu slab may be equipped with an additional
108 * freelist that allows lockless access to
109 * free objects in addition to the regular freelist
110 * that requires the slab lock.
111 *
112 * PageError Slab requires special handling due to debug
113 * options set. This moves slab handling out of
114 * the fast path and disables lockless freelists.
115 */
116
117static inline int kmem_cache_debug(struct kmem_cache *s)
118{
119#ifdef CONFIG_SLUB_DEBUG
120 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
121#else
122 return 0;
123#endif
124}
125
126static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
127{
128#ifdef CONFIG_SLUB_CPU_PARTIAL
129 return !kmem_cache_debug(s);
130#else
131 return false;
132#endif
133}
134
135/*
136 * Issues still to be resolved:
137 *
138 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
139 *
140 * - Variable sizing of the per node arrays
141 */
142
143/* Enable to test recovery from slab corruption on boot */
144#undef SLUB_RESILIENCY_TEST
145
146/* Enable to log cmpxchg failures */
147#undef SLUB_DEBUG_CMPXCHG
148
149/*
150 * Mininum number of partial slabs. These will be left on the partial
151 * lists even if they are empty. kmem_cache_shrink may reclaim them.
152 */
153#define MIN_PARTIAL 5
154
155/*
156 * Maximum number of desirable partial slabs.
157 * The existence of more partial slabs makes kmem_cache_shrink
158 * sort the partial list by the number of objects in use.
159 */
160#define MAX_PARTIAL 10
161
162#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
163 SLAB_POISON | SLAB_STORE_USER)
164
165/*
166 * Debugging flags that require metadata to be stored in the slab. These get
167 * disabled when slub_debug=O is used and a cache's min order increases with
168 * metadata.
169 */
170#define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
171
172/*
173 * Set of flags that will prevent slab merging
174 */
175#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
176 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
177 SLAB_FAILSLAB)
178
179#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
180 SLAB_CACHE_DMA | SLAB_NOTRACK)
181
182#define OO_SHIFT 16
183#define OO_MASK ((1 << OO_SHIFT) - 1)
184#define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
185
186/* Internal SLUB flags */
187#define __OBJECT_POISON 0x80000000UL /* Poison object */
188#define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
189
190#ifdef CONFIG_SMP
191static struct notifier_block slab_notifier;
192#endif
193
194/*
195 * Tracking user of a slab.
196 */
197#define TRACK_ADDRS_COUNT 16
198struct track {
199 unsigned long addr; /* Called from address */
200#ifdef CONFIG_STACKTRACE
201 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
202#endif
203 int cpu; /* Was running on cpu */
204 int pid; /* Pid context */
205 unsigned long when; /* When did the operation occur */
206};
207
208enum track_item { TRACK_ALLOC, TRACK_FREE };
209
210#ifdef CONFIG_SYSFS
211static int sysfs_slab_add(struct kmem_cache *);
212static int sysfs_slab_alias(struct kmem_cache *, const char *);
213static void memcg_propagate_slab_attrs(struct kmem_cache *s);
214#else
215static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
216static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
217 { return 0; }
218static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
219#endif
220
221static inline void stat(const struct kmem_cache *s, enum stat_item si)
222{
223#ifdef CONFIG_SLUB_STATS
224 /*
225 * The rmw is racy on a preemptible kernel but this is acceptable, so
226 * avoid this_cpu_add()'s irq-disable overhead.
227 */
228 raw_cpu_inc(s->cpu_slab->stat[si]);
229#endif
230}
231
232/********************************************************************
233 * Core slab cache functions
234 *******************************************************************/
235
236static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
237{
238 return s->node[node];
239}
240
241/* Verify that a pointer has an address that is valid within a slab page */
242static inline int check_valid_pointer(struct kmem_cache *s,
243 struct page *page, const void *object)
244{
245 void *base;
246
247 if (!object)
248 return 1;
249
250 base = page_address(page);
251 if (object < base || object >= base + page->objects * s->size ||
252 (object - base) % s->size) {
253 return 0;
254 }
255
256 return 1;
257}
258
259static inline void *get_freepointer(struct kmem_cache *s, void *object)
260{
261 return *(void **)(object + s->offset);
262}
263
264static void prefetch_freepointer(const struct kmem_cache *s, void *object)
265{
266 prefetch(object + s->offset);
267}
268
269static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
270{
271 void *p;
272
273#ifdef CONFIG_DEBUG_PAGEALLOC
274 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
275#else
276 p = get_freepointer(s, object);
277#endif
278 return p;
279}
280
281static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
282{
283 *(void **)(object + s->offset) = fp;
284}
285
286/* Loop over all objects in a slab */
287#define for_each_object(__p, __s, __addr, __objects) \
288 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
289 __p += (__s)->size)
290
291/* Determine object index from a given position */
292static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
293{
294 return (p - addr) / s->size;
295}
296
297static inline size_t slab_ksize(const struct kmem_cache *s)
298{
299#ifdef CONFIG_SLUB_DEBUG
300 /*
301 * Debugging requires use of the padding between object
302 * and whatever may come after it.
303 */
304 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
305 return s->object_size;
306
307#endif
308 /*
309 * If we have the need to store the freelist pointer
310 * back there or track user information then we can
311 * only use the space before that information.
312 */
313 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
314 return s->inuse;
315 /*
316 * Else we can use all the padding etc for the allocation
317 */
318 return s->size;
319}
320
321static inline int order_objects(int order, unsigned long size, int reserved)
322{
323 return ((PAGE_SIZE << order) - reserved) / size;
324}
325
326static inline struct kmem_cache_order_objects oo_make(int order,
327 unsigned long size, int reserved)
328{
329 struct kmem_cache_order_objects x = {
330 (order << OO_SHIFT) + order_objects(order, size, reserved)
331 };
332
333 return x;
334}
335
336static inline int oo_order(struct kmem_cache_order_objects x)
337{
338 return x.x >> OO_SHIFT;
339}
340
341static inline int oo_objects(struct kmem_cache_order_objects x)
342{
343 return x.x & OO_MASK;
344}
345
346/*
347 * Per slab locking using the pagelock
348 */
349static __always_inline void slab_lock(struct page *page)
350{
351 bit_spin_lock(PG_locked, &page->flags);
352}
353
354static __always_inline void slab_unlock(struct page *page)
355{
356 __bit_spin_unlock(PG_locked, &page->flags);
357}
358
359static inline void set_page_slub_counters(struct page *page, unsigned long counters_new)
360{
361 struct page tmp;
362 tmp.counters = counters_new;
363 /*
364 * page->counters can cover frozen/inuse/objects as well
365 * as page->_count. If we assign to ->counters directly
366 * we run the risk of losing updates to page->_count, so
367 * be careful and only assign to the fields we need.
368 */
369 page->frozen = tmp.frozen;
370 page->inuse = tmp.inuse;
371 page->objects = tmp.objects;
372}
373
374/* Interrupts must be disabled (for the fallback code to work right) */
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 VM_BUG_ON(!irqs_disabled());
381#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
382 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
383 if (s->flags & __CMPXCHG_DOUBLE) {
384 if (cmpxchg_double(&page->freelist, &page->counters,
385 freelist_old, counters_old,
386 freelist_new, counters_new))
387 return 1;
388 } else
389#endif
390 {
391 slab_lock(page);
392 if (page->freelist == freelist_old &&
393 page->counters == counters_old) {
394 page->freelist = freelist_new;
395 set_page_slub_counters(page, counters_new);
396 slab_unlock(page);
397 return 1;
398 }
399 slab_unlock(page);
400 }
401
402 cpu_relax();
403 stat(s, CMPXCHG_DOUBLE_FAIL);
404
405#ifdef SLUB_DEBUG_CMPXCHG
406 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
407#endif
408
409 return 0;
410}
411
412static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
413 void *freelist_old, unsigned long counters_old,
414 void *freelist_new, unsigned long counters_new,
415 const char *n)
416{
417#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
418 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
419 if (s->flags & __CMPXCHG_DOUBLE) {
420 if (cmpxchg_double(&page->freelist, &page->counters,
421 freelist_old, counters_old,
422 freelist_new, counters_new))
423 return 1;
424 } else
425#endif
426 {
427 unsigned long flags;
428
429 local_irq_save(flags);
430 slab_lock(page);
431 if (page->freelist == freelist_old &&
432 page->counters == counters_old) {
433 page->freelist = freelist_new;
434 set_page_slub_counters(page, counters_new);
435 slab_unlock(page);
436 local_irq_restore(flags);
437 return 1;
438 }
439 slab_unlock(page);
440 local_irq_restore(flags);
441 }
442
443 cpu_relax();
444 stat(s, CMPXCHG_DOUBLE_FAIL);
445
446#ifdef SLUB_DEBUG_CMPXCHG
447 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
448#endif
449
450 return 0;
451}
452
453#ifdef CONFIG_SLUB_DEBUG
454/*
455 * Determine a map of object in use on a page.
456 *
457 * Node listlock must be held to guarantee that the page does
458 * not vanish from under us.
459 */
460static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
461{
462 void *p;
463 void *addr = page_address(page);
464
465 for (p = page->freelist; p; p = get_freepointer(s, p))
466 set_bit(slab_index(p, s, addr), map);
467}
468
469/*
470 * Debug settings:
471 */
472#ifdef CONFIG_SLUB_DEBUG_ON
473static int slub_debug = DEBUG_DEFAULT_FLAGS;
474#else
475static int slub_debug;
476#endif
477
478static char *slub_debug_slabs;
479static int disable_higher_order_debug;
480
481/*
482 * Object debugging
483 */
484static void print_section(char *text, u8 *addr, unsigned int length)
485{
486 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
487 length, 1);
488}
489
490static struct track *get_track(struct kmem_cache *s, void *object,
491 enum track_item alloc)
492{
493 struct track *p;
494
495 if (s->offset)
496 p = object + s->offset + sizeof(void *);
497 else
498 p = object + s->inuse;
499
500 return p + alloc;
501}
502
503static void set_track(struct kmem_cache *s, void *object,
504 enum track_item alloc, unsigned long addr)
505{
506 struct track *p = get_track(s, object, alloc);
507
508 if (addr) {
509#ifdef CONFIG_STACKTRACE
510 struct stack_trace trace;
511 int i;
512
513 trace.nr_entries = 0;
514 trace.max_entries = TRACK_ADDRS_COUNT;
515 trace.entries = p->addrs;
516 trace.skip = 3;
517 save_stack_trace(&trace);
518
519 /* See rant in lockdep.c */
520 if (trace.nr_entries != 0 &&
521 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
522 trace.nr_entries--;
523
524 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
525 p->addrs[i] = 0;
526#endif
527 p->addr = addr;
528 p->cpu = smp_processor_id();
529 p->pid = current->pid;
530 p->when = jiffies;
531 } else
532 memset(p, 0, sizeof(struct track));
533}
534
535static void init_tracking(struct kmem_cache *s, void *object)
536{
537 if (!(s->flags & SLAB_STORE_USER))
538 return;
539
540 set_track(s, object, TRACK_FREE, 0UL);
541 set_track(s, object, TRACK_ALLOC, 0UL);
542}
543
544static void print_track(const char *s, struct track *t)
545{
546 if (!t->addr)
547 return;
548
549 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
550 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
551#ifdef CONFIG_STACKTRACE
552 {
553 int i;
554 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
555 if (t->addrs[i])
556 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
557 else
558 break;
559 }
560#endif
561}
562
563static void print_tracking(struct kmem_cache *s, void *object)
564{
565 if (!(s->flags & SLAB_STORE_USER))
566 return;
567
568 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
569 print_track("Freed", get_track(s, object, TRACK_FREE));
570}
571
572static void print_page_info(struct page *page)
573{
574 printk(KERN_ERR
575 "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
576 page, page->objects, page->inuse, page->freelist, page->flags);
577
578}
579
580static void slab_bug(struct kmem_cache *s, char *fmt, ...)
581{
582 va_list args;
583 char buf[100];
584
585 va_start(args, fmt);
586 vsnprintf(buf, sizeof(buf), fmt, args);
587 va_end(args);
588 printk(KERN_ERR "========================================"
589 "=====================================\n");
590 printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf);
591 printk(KERN_ERR "----------------------------------------"
592 "-------------------------------------\n\n");
593
594 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
595}
596
597static void slab_fix(struct kmem_cache *s, char *fmt, ...)
598{
599 va_list args;
600 char buf[100];
601
602 va_start(args, fmt);
603 vsnprintf(buf, sizeof(buf), fmt, args);
604 va_end(args);
605 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
606}
607
608static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
609{
610 unsigned int off; /* Offset of last byte */
611 u8 *addr = page_address(page);
612
613 print_tracking(s, p);
614
615 print_page_info(page);
616
617 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
618 p, p - addr, get_freepointer(s, p));
619
620 if (p > addr + 16)
621 print_section("Bytes b4 ", p - 16, 16);
622
623 print_section("Object ", p, min_t(unsigned long, s->object_size,
624 PAGE_SIZE));
625 if (s->flags & SLAB_RED_ZONE)
626 print_section("Redzone ", p + s->object_size,
627 s->inuse - s->object_size);
628
629 if (s->offset)
630 off = s->offset + sizeof(void *);
631 else
632 off = s->inuse;
633
634 if (s->flags & SLAB_STORE_USER)
635 off += 2 * sizeof(struct track);
636
637 if (off != s->size)
638 /* Beginning of the filler is the free pointer */
639 print_section("Padding ", p + off, s->size - off);
640
641 dump_stack();
642}
643
644static void object_err(struct kmem_cache *s, struct page *page,
645 u8 *object, char *reason)
646{
647 slab_bug(s, "%s", reason);
648 print_trailer(s, page, object);
649}
650
651static void slab_err(struct kmem_cache *s, struct page *page,
652 const char *fmt, ...)
653{
654 va_list args;
655 char buf[100];
656
657 va_start(args, fmt);
658 vsnprintf(buf, sizeof(buf), fmt, args);
659 va_end(args);
660 slab_bug(s, "%s", buf);
661 print_page_info(page);
662 dump_stack();
663}
664
665static void init_object(struct kmem_cache *s, void *object, u8 val)
666{
667 u8 *p = object;
668
669 if (s->flags & __OBJECT_POISON) {
670 memset(p, POISON_FREE, s->object_size - 1);
671 p[s->object_size - 1] = POISON_END;
672 }
673
674 if (s->flags & SLAB_RED_ZONE)
675 memset(p + s->object_size, val, s->inuse - s->object_size);
676}
677
678static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
679 void *from, void *to)
680{
681 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
682 memset(from, data, to - from);
683}
684
685static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
686 u8 *object, char *what,
687 u8 *start, unsigned int value, unsigned int bytes)
688{
689 u8 *fault;
690 u8 *end;
691
692 fault = memchr_inv(start, value, bytes);
693 if (!fault)
694 return 1;
695
696 end = start + bytes;
697 while (end > fault && end[-1] == value)
698 end--;
699
700 slab_bug(s, "%s overwritten", what);
701 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
702 fault, end - 1, fault[0], value);
703 print_trailer(s, page, object);
704
705 restore_bytes(s, what, value, fault, end);
706 return 0;
707}
708
709/*
710 * Object layout:
711 *
712 * object address
713 * Bytes of the object to be managed.
714 * If the freepointer may overlay the object then the free
715 * pointer is the first word of the object.
716 *
717 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
718 * 0xa5 (POISON_END)
719 *
720 * object + s->object_size
721 * Padding to reach word boundary. This is also used for Redzoning.
722 * Padding is extended by another word if Redzoning is enabled and
723 * object_size == inuse.
724 *
725 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
726 * 0xcc (RED_ACTIVE) for objects in use.
727 *
728 * object + s->inuse
729 * Meta data starts here.
730 *
731 * A. Free pointer (if we cannot overwrite object on free)
732 * B. Tracking data for SLAB_STORE_USER
733 * C. Padding to reach required alignment boundary or at mininum
734 * one word if debugging is on to be able to detect writes
735 * before the word boundary.
736 *
737 * Padding is done using 0x5a (POISON_INUSE)
738 *
739 * object + s->size
740 * Nothing is used beyond s->size.
741 *
742 * If slabcaches are merged then the object_size and inuse boundaries are mostly
743 * ignored. And therefore no slab options that rely on these boundaries
744 * may be used with merged slabcaches.
745 */
746
747static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
748{
749 unsigned long off = s->inuse; /* The end of info */
750
751 if (s->offset)
752 /* Freepointer is placed after the object. */
753 off += sizeof(void *);
754
755 if (s->flags & SLAB_STORE_USER)
756 /* We also have user information there */
757 off += 2 * sizeof(struct track);
758
759 if (s->size == off)
760 return 1;
761
762 return check_bytes_and_report(s, page, p, "Object padding",
763 p + off, POISON_INUSE, s->size - off);
764}
765
766/* Check the pad bytes at the end of a slab page */
767static int slab_pad_check(struct kmem_cache *s, struct page *page)
768{
769 u8 *start;
770 u8 *fault;
771 u8 *end;
772 int length;
773 int remainder;
774
775 if (!(s->flags & SLAB_POISON))
776 return 1;
777
778 start = page_address(page);
779 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
780 end = start + length;
781 remainder = length % s->size;
782 if (!remainder)
783 return 1;
784
785 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
786 if (!fault)
787 return 1;
788 while (end > fault && end[-1] == POISON_INUSE)
789 end--;
790
791 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
792 print_section("Padding ", end - remainder, remainder);
793
794 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
795 return 0;
796}
797
798static int check_object(struct kmem_cache *s, struct page *page,
799 void *object, u8 val)
800{
801 u8 *p = object;
802 u8 *endobject = object + s->object_size;
803
804 if (s->flags & SLAB_RED_ZONE) {
805 if (!check_bytes_and_report(s, page, object, "Redzone",
806 endobject, val, s->inuse - s->object_size))
807 return 0;
808 } else {
809 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
810 check_bytes_and_report(s, page, p, "Alignment padding",
811 endobject, POISON_INUSE,
812 s->inuse - s->object_size);
813 }
814 }
815
816 if (s->flags & SLAB_POISON) {
817 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
818 (!check_bytes_and_report(s, page, p, "Poison", p,
819 POISON_FREE, s->object_size - 1) ||
820 !check_bytes_and_report(s, page, p, "Poison",
821 p + s->object_size - 1, POISON_END, 1)))
822 return 0;
823 /*
824 * check_pad_bytes cleans up on its own.
825 */
826 check_pad_bytes(s, page, p);
827 }
828
829 if (!s->offset && val == SLUB_RED_ACTIVE)
830 /*
831 * Object and freepointer overlap. Cannot check
832 * freepointer while object is allocated.
833 */
834 return 1;
835
836 /* Check free pointer validity */
837 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
838 object_err(s, page, p, "Freepointer corrupt");
839 /*
840 * No choice but to zap it and thus lose the remainder
841 * of the free objects in this slab. May cause
842 * another error because the object count is now wrong.
843 */
844 set_freepointer(s, p, NULL);
845 return 0;
846 }
847 return 1;
848}
849
850static int check_slab(struct kmem_cache *s, struct page *page)
851{
852 int maxobj;
853
854 VM_BUG_ON(!irqs_disabled());
855
856 if (!PageSlab(page)) {
857 slab_err(s, page, "Not a valid slab page");
858 return 0;
859 }
860
861 maxobj = order_objects(compound_order(page), s->size, s->reserved);
862 if (page->objects > maxobj) {
863 slab_err(s, page, "objects %u > max %u",
864 s->name, page->objects, maxobj);
865 return 0;
866 }
867 if (page->inuse > page->objects) {
868 slab_err(s, page, "inuse %u > max %u",
869 s->name, page->inuse, page->objects);
870 return 0;
871 }
872 /* Slab_pad_check fixes things up after itself */
873 slab_pad_check(s, page);
874 return 1;
875}
876
877/*
878 * Determine if a certain object on a page is on the freelist. Must hold the
879 * slab lock to guarantee that the chains are in a consistent state.
880 */
881static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
882{
883 int nr = 0;
884 void *fp;
885 void *object = NULL;
886 unsigned long max_objects;
887
888 fp = page->freelist;
889 while (fp && nr <= page->objects) {
890 if (fp == search)
891 return 1;
892 if (!check_valid_pointer(s, page, fp)) {
893 if (object) {
894 object_err(s, page, object,
895 "Freechain corrupt");
896 set_freepointer(s, object, NULL);
897 } else {
898 slab_err(s, page, "Freepointer corrupt");
899 page->freelist = NULL;
900 page->inuse = page->objects;
901 slab_fix(s, "Freelist cleared");
902 return 0;
903 }
904 break;
905 }
906 object = fp;
907 fp = get_freepointer(s, object);
908 nr++;
909 }
910
911 max_objects = order_objects(compound_order(page), s->size, s->reserved);
912 if (max_objects > MAX_OBJS_PER_PAGE)
913 max_objects = MAX_OBJS_PER_PAGE;
914
915 if (page->objects != max_objects) {
916 slab_err(s, page, "Wrong number of objects. Found %d but "
917 "should be %d", page->objects, max_objects);
918 page->objects = max_objects;
919 slab_fix(s, "Number of objects adjusted.");
920 }
921 if (page->inuse != page->objects - nr) {
922 slab_err(s, page, "Wrong object count. Counter is %d but "
923 "counted were %d", page->inuse, page->objects - nr);
924 page->inuse = page->objects - nr;
925 slab_fix(s, "Object count adjusted.");
926 }
927 return search == NULL;
928}
929
930static void trace(struct kmem_cache *s, struct page *page, void *object,
931 int alloc)
932{
933 if (s->flags & SLAB_TRACE) {
934 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
935 s->name,
936 alloc ? "alloc" : "free",
937 object, page->inuse,
938 page->freelist);
939
940 if (!alloc)
941 print_section("Object ", (void *)object,
942 s->object_size);
943
944 dump_stack();
945 }
946}
947
948/*
949 * Hooks for other subsystems that check memory allocations. In a typical
950 * production configuration these hooks all should produce no code at all.
951 */
952static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
953{
954 kmemleak_alloc(ptr, size, 1, flags);
955}
956
957static inline void kfree_hook(const void *x)
958{
959 kmemleak_free(x);
960}
961
962static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
963{
964 flags &= gfp_allowed_mask;
965 lockdep_trace_alloc(flags);
966 might_sleep_if(flags & __GFP_WAIT);
967
968 return should_failslab(s->object_size, flags, s->flags);
969}
970
971static inline void slab_post_alloc_hook(struct kmem_cache *s,
972 gfp_t flags, void *object)
973{
974 flags &= gfp_allowed_mask;
975 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
976 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
977}
978
979static inline void slab_free_hook(struct kmem_cache *s, void *x)
980{
981 kmemleak_free_recursive(x, s->flags);
982
983 /*
984 * Trouble is that we may no longer disable interrupts in the fast path
985 * So in order to make the debug calls that expect irqs to be
986 * disabled we need to disable interrupts temporarily.
987 */
988#if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
989 {
990 unsigned long flags;
991
992 local_irq_save(flags);
993 kmemcheck_slab_free(s, x, s->object_size);
994 debug_check_no_locks_freed(x, s->object_size);
995 local_irq_restore(flags);
996 }
997#endif
998 if (!(s->flags & SLAB_DEBUG_OBJECTS))
999 debug_check_no_obj_freed(x, s->object_size);
1000}
1001
1002/*
1003 * Tracking of fully allocated slabs for debugging purposes.
1004 */
1005static void add_full(struct kmem_cache *s,
1006 struct kmem_cache_node *n, struct page *page)
1007{
1008 if (!(s->flags & SLAB_STORE_USER))
1009 return;
1010
1011 lockdep_assert_held(&n->list_lock);
1012 list_add(&page->lru, &n->full);
1013}
1014
1015static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
1016{
1017 if (!(s->flags & SLAB_STORE_USER))
1018 return;
1019
1020 lockdep_assert_held(&n->list_lock);
1021 list_del(&page->lru);
1022}
1023
1024/* Tracking of the number of slabs for debugging purposes */
1025static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1026{
1027 struct kmem_cache_node *n = get_node(s, node);
1028
1029 return atomic_long_read(&n->nr_slabs);
1030}
1031
1032static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1033{
1034 return atomic_long_read(&n->nr_slabs);
1035}
1036
1037static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1038{
1039 struct kmem_cache_node *n = get_node(s, node);
1040
1041 /*
1042 * May be called early in order to allocate a slab for the
1043 * kmem_cache_node structure. Solve the chicken-egg
1044 * dilemma by deferring the increment of the count during
1045 * bootstrap (see early_kmem_cache_node_alloc).
1046 */
1047 if (likely(n)) {
1048 atomic_long_inc(&n->nr_slabs);
1049 atomic_long_add(objects, &n->total_objects);
1050 }
1051}
1052static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1053{
1054 struct kmem_cache_node *n = get_node(s, node);
1055
1056 atomic_long_dec(&n->nr_slabs);
1057 atomic_long_sub(objects, &n->total_objects);
1058}
1059
1060/* Object debug checks for alloc/free paths */
1061static void setup_object_debug(struct kmem_cache *s, struct page *page,
1062 void *object)
1063{
1064 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1065 return;
1066
1067 init_object(s, object, SLUB_RED_INACTIVE);
1068 init_tracking(s, object);
1069}
1070
1071static noinline int alloc_debug_processing(struct kmem_cache *s,
1072 struct page *page,
1073 void *object, unsigned long addr)
1074{
1075 if (!check_slab(s, page))
1076 goto bad;
1077
1078 if (!check_valid_pointer(s, page, object)) {
1079 object_err(s, page, object, "Freelist Pointer check fails");
1080 goto bad;
1081 }
1082
1083 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1084 goto bad;
1085
1086 /* Success perform special debug activities for allocs */
1087 if (s->flags & SLAB_STORE_USER)
1088 set_track(s, object, TRACK_ALLOC, addr);
1089 trace(s, page, object, 1);
1090 init_object(s, object, SLUB_RED_ACTIVE);
1091 return 1;
1092
1093bad:
1094 if (PageSlab(page)) {
1095 /*
1096 * If this is a slab page then lets do the best we can
1097 * to avoid issues in the future. Marking all objects
1098 * as used avoids touching the remaining objects.
1099 */
1100 slab_fix(s, "Marking all objects used");
1101 page->inuse = page->objects;
1102 page->freelist = NULL;
1103 }
1104 return 0;
1105}
1106
1107static noinline struct kmem_cache_node *free_debug_processing(
1108 struct kmem_cache *s, struct page *page, void *object,
1109 unsigned long addr, unsigned long *flags)
1110{
1111 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1112
1113 spin_lock_irqsave(&n->list_lock, *flags);
1114 slab_lock(page);
1115
1116 if (!check_slab(s, page))
1117 goto fail;
1118
1119 if (!check_valid_pointer(s, page, object)) {
1120 slab_err(s, page, "Invalid object pointer 0x%p", object);
1121 goto fail;
1122 }
1123
1124 if (on_freelist(s, page, object)) {
1125 object_err(s, page, object, "Object already free");
1126 goto fail;
1127 }
1128
1129 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1130 goto out;
1131
1132 if (unlikely(s != page->slab_cache)) {
1133 if (!PageSlab(page)) {
1134 slab_err(s, page, "Attempt to free object(0x%p) "
1135 "outside of slab", object);
1136 } else if (!page->slab_cache) {
1137 printk(KERN_ERR
1138 "SLUB <none>: no slab for object 0x%p.\n",
1139 object);
1140 dump_stack();
1141 } else
1142 object_err(s, page, object,
1143 "page slab pointer corrupt.");
1144 goto fail;
1145 }
1146
1147 if (s->flags & SLAB_STORE_USER)
1148 set_track(s, object, TRACK_FREE, addr);
1149 trace(s, page, object, 0);
1150 init_object(s, object, SLUB_RED_INACTIVE);
1151out:
1152 slab_unlock(page);
1153 /*
1154 * Keep node_lock to preserve integrity
1155 * until the object is actually freed
1156 */
1157 return n;
1158
1159fail:
1160 slab_unlock(page);
1161 spin_unlock_irqrestore(&n->list_lock, *flags);
1162 slab_fix(s, "Object at 0x%p not freed", object);
1163 return NULL;
1164}
1165
1166static int __init setup_slub_debug(char *str)
1167{
1168 slub_debug = DEBUG_DEFAULT_FLAGS;
1169 if (*str++ != '=' || !*str)
1170 /*
1171 * No options specified. Switch on full debugging.
1172 */
1173 goto out;
1174
1175 if (*str == ',')
1176 /*
1177 * No options but restriction on slabs. This means full
1178 * debugging for slabs matching a pattern.
1179 */
1180 goto check_slabs;
1181
1182 if (tolower(*str) == 'o') {
1183 /*
1184 * Avoid enabling debugging on caches if its minimum order
1185 * would increase as a result.
1186 */
1187 disable_higher_order_debug = 1;
1188 goto out;
1189 }
1190
1191 slub_debug = 0;
1192 if (*str == '-')
1193 /*
1194 * Switch off all debugging measures.
1195 */
1196 goto out;
1197
1198 /*
1199 * Determine which debug features should be switched on
1200 */
1201 for (; *str && *str != ','; str++) {
1202 switch (tolower(*str)) {
1203 case 'f':
1204 slub_debug |= SLAB_DEBUG_FREE;
1205 break;
1206 case 'z':
1207 slub_debug |= SLAB_RED_ZONE;
1208 break;
1209 case 'p':
1210 slub_debug |= SLAB_POISON;
1211 break;
1212 case 'u':
1213 slub_debug |= SLAB_STORE_USER;
1214 break;
1215 case 't':
1216 slub_debug |= SLAB_TRACE;
1217 break;
1218 case 'a':
1219 slub_debug |= SLAB_FAILSLAB;
1220 break;
1221 default:
1222 printk(KERN_ERR "slub_debug option '%c' "
1223 "unknown. skipped\n", *str);
1224 }
1225 }
1226
1227check_slabs:
1228 if (*str == ',')
1229 slub_debug_slabs = str + 1;
1230out:
1231 return 1;
1232}
1233
1234__setup("slub_debug", setup_slub_debug);
1235
1236static unsigned long kmem_cache_flags(unsigned long object_size,
1237 unsigned long flags, const char *name,
1238 void (*ctor)(void *))
1239{
1240 /*
1241 * Enable debugging if selected on the kernel commandline.
1242 */
1243 if (slub_debug && (!slub_debug_slabs || (name &&
1244 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1245 flags |= slub_debug;
1246
1247 return flags;
1248}
1249#else
1250static inline void setup_object_debug(struct kmem_cache *s,
1251 struct page *page, void *object) {}
1252
1253static inline int alloc_debug_processing(struct kmem_cache *s,
1254 struct page *page, void *object, unsigned long addr) { return 0; }
1255
1256static inline struct kmem_cache_node *free_debug_processing(
1257 struct kmem_cache *s, struct page *page, void *object,
1258 unsigned long addr, unsigned long *flags) { return NULL; }
1259
1260static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1261 { return 1; }
1262static inline int check_object(struct kmem_cache *s, struct page *page,
1263 void *object, u8 val) { return 1; }
1264static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1265 struct page *page) {}
1266static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1267 struct page *page) {}
1268static inline unsigned long kmem_cache_flags(unsigned long object_size,
1269 unsigned long flags, const char *name,
1270 void (*ctor)(void *))
1271{
1272 return flags;
1273}
1274#define slub_debug 0
1275
1276#define disable_higher_order_debug 0
1277
1278static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1279 { return 0; }
1280static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1281 { return 0; }
1282static inline void inc_slabs_node(struct kmem_cache *s, int node,
1283 int objects) {}
1284static inline void dec_slabs_node(struct kmem_cache *s, int node,
1285 int objects) {}
1286
1287static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1288{
1289 kmemleak_alloc(ptr, size, 1, flags);
1290}
1291
1292static inline void kfree_hook(const void *x)
1293{
1294 kmemleak_free(x);
1295}
1296
1297static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1298 { return 0; }
1299
1300static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1301 void *object)
1302{
1303 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags,
1304 flags & gfp_allowed_mask);
1305}
1306
1307static inline void slab_free_hook(struct kmem_cache *s, void *x)
1308{
1309 kmemleak_free_recursive(x, s->flags);
1310}
1311
1312#endif /* CONFIG_SLUB_DEBUG */
1313
1314/*
1315 * Slab allocation and freeing
1316 */
1317static inline struct page *alloc_slab_page(gfp_t flags, int node,
1318 struct kmem_cache_order_objects oo)
1319{
1320 int order = oo_order(oo);
1321
1322 flags |= __GFP_NOTRACK;
1323
1324 if (node == NUMA_NO_NODE)
1325 return alloc_pages(flags, order);
1326 else
1327 return alloc_pages_exact_node(node, flags, order);
1328}
1329
1330static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1331{
1332 struct page *page;
1333 struct kmem_cache_order_objects oo = s->oo;
1334 gfp_t alloc_gfp;
1335
1336 flags &= gfp_allowed_mask;
1337
1338 if (flags & __GFP_WAIT)
1339 local_irq_enable();
1340
1341 flags |= s->allocflags;
1342
1343 /*
1344 * Let the initial higher-order allocation fail under memory pressure
1345 * so we fall-back to the minimum order allocation.
1346 */
1347 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1348
1349 page = alloc_slab_page(alloc_gfp, node, oo);
1350 if (unlikely(!page)) {
1351 oo = s->min;
1352 alloc_gfp = flags;
1353 /*
1354 * Allocation may have failed due to fragmentation.
1355 * Try a lower order alloc if possible
1356 */
1357 page = alloc_slab_page(alloc_gfp, node, oo);
1358
1359 if (page)
1360 stat(s, ORDER_FALLBACK);
1361 }
1362
1363 if (kmemcheck_enabled && page
1364 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1365 int pages = 1 << oo_order(oo);
1366
1367 kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node);
1368
1369 /*
1370 * Objects from caches that have a constructor don't get
1371 * cleared when they're allocated, so we need to do it here.
1372 */
1373 if (s->ctor)
1374 kmemcheck_mark_uninitialized_pages(page, pages);
1375 else
1376 kmemcheck_mark_unallocated_pages(page, pages);
1377 }
1378
1379 if (flags & __GFP_WAIT)
1380 local_irq_disable();
1381 if (!page)
1382 return NULL;
1383
1384 page->objects = oo_objects(oo);
1385 mod_zone_page_state(page_zone(page),
1386 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1387 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1388 1 << oo_order(oo));
1389
1390 return page;
1391}
1392
1393static void setup_object(struct kmem_cache *s, struct page *page,
1394 void *object)
1395{
1396 setup_object_debug(s, page, object);
1397 if (unlikely(s->ctor))
1398 s->ctor(object);
1399}
1400
1401static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1402{
1403 struct page *page;
1404 void *start;
1405 void *last;
1406 void *p;
1407 int order;
1408
1409 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1410
1411 page = allocate_slab(s,
1412 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1413 if (!page)
1414 goto out;
1415
1416 order = compound_order(page);
1417 inc_slabs_node(s, page_to_nid(page), page->objects);
1418 memcg_bind_pages(s, order);
1419 page->slab_cache = s;
1420 __SetPageSlab(page);
1421 if (page->pfmemalloc)
1422 SetPageSlabPfmemalloc(page);
1423
1424 start = page_address(page);
1425
1426 if (unlikely(s->flags & SLAB_POISON))
1427 memset(start, POISON_INUSE, PAGE_SIZE << order);
1428
1429 last = start;
1430 for_each_object(p, s, start, page->objects) {
1431 setup_object(s, page, last);
1432 set_freepointer(s, last, p);
1433 last = p;
1434 }
1435 setup_object(s, page, last);
1436 set_freepointer(s, last, NULL);
1437
1438 page->freelist = start;
1439 page->inuse = page->objects;
1440 page->frozen = 1;
1441out:
1442 return page;
1443}
1444
1445static void __free_slab(struct kmem_cache *s, struct page *page)
1446{
1447 int order = compound_order(page);
1448 int pages = 1 << order;
1449
1450 if (kmem_cache_debug(s)) {
1451 void *p;
1452
1453 slab_pad_check(s, page);
1454 for_each_object(p, s, page_address(page),
1455 page->objects)
1456 check_object(s, page, p, SLUB_RED_INACTIVE);
1457 }
1458
1459 kmemcheck_free_shadow(page, compound_order(page));
1460
1461 mod_zone_page_state(page_zone(page),
1462 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1463 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1464 -pages);
1465
1466 __ClearPageSlabPfmemalloc(page);
1467 __ClearPageSlab(page);
1468
1469 memcg_release_pages(s, order);
1470 page_mapcount_reset(page);
1471 if (current->reclaim_state)
1472 current->reclaim_state->reclaimed_slab += pages;
1473 __free_memcg_kmem_pages(page, order);
1474}
1475
1476#define need_reserve_slab_rcu \
1477 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1478
1479static void rcu_free_slab(struct rcu_head *h)
1480{
1481 struct page *page;
1482
1483 if (need_reserve_slab_rcu)
1484 page = virt_to_head_page(h);
1485 else
1486 page = container_of((struct list_head *)h, struct page, lru);
1487
1488 __free_slab(page->slab_cache, page);
1489}
1490
1491static void free_slab(struct kmem_cache *s, struct page *page)
1492{
1493 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1494 struct rcu_head *head;
1495
1496 if (need_reserve_slab_rcu) {
1497 int order = compound_order(page);
1498 int offset = (PAGE_SIZE << order) - s->reserved;
1499
1500 VM_BUG_ON(s->reserved != sizeof(*head));
1501 head = page_address(page) + offset;
1502 } else {
1503 /*
1504 * RCU free overloads the RCU head over the LRU
1505 */
1506 head = (void *)&page->lru;
1507 }
1508
1509 call_rcu(head, rcu_free_slab);
1510 } else
1511 __free_slab(s, page);
1512}
1513
1514static void discard_slab(struct kmem_cache *s, struct page *page)
1515{
1516 dec_slabs_node(s, page_to_nid(page), page->objects);
1517 free_slab(s, page);
1518}
1519
1520/*
1521 * Management of partially allocated slabs.
1522 */
1523static inline void
1524__add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1525{
1526 n->nr_partial++;
1527 if (tail == DEACTIVATE_TO_TAIL)
1528 list_add_tail(&page->lru, &n->partial);
1529 else
1530 list_add(&page->lru, &n->partial);
1531}
1532
1533static inline void add_partial(struct kmem_cache_node *n,
1534 struct page *page, int tail)
1535{
1536 lockdep_assert_held(&n->list_lock);
1537 __add_partial(n, page, tail);
1538}
1539
1540static inline void
1541__remove_partial(struct kmem_cache_node *n, struct page *page)
1542{
1543 list_del(&page->lru);
1544 n->nr_partial--;
1545}
1546
1547static inline void remove_partial(struct kmem_cache_node *n,
1548 struct page *page)
1549{
1550 lockdep_assert_held(&n->list_lock);
1551 __remove_partial(n, page);
1552}
1553
1554/*
1555 * Remove slab from the partial list, freeze it and
1556 * return the pointer to the freelist.
1557 *
1558 * Returns a list of objects or NULL if it fails.
1559 */
1560static inline void *acquire_slab(struct kmem_cache *s,
1561 struct kmem_cache_node *n, struct page *page,
1562 int mode, int *objects)
1563{
1564 void *freelist;
1565 unsigned long counters;
1566 struct page new;
1567
1568 lockdep_assert_held(&n->list_lock);
1569
1570 /*
1571 * Zap the freelist and set the frozen bit.
1572 * The old freelist is the list of objects for the
1573 * per cpu allocation list.
1574 */
1575 freelist = page->freelist;
1576 counters = page->counters;
1577 new.counters = counters;
1578 *objects = new.objects - new.inuse;
1579 if (mode) {
1580 new.inuse = page->objects;
1581 new.freelist = NULL;
1582 } else {
1583 new.freelist = freelist;
1584 }
1585
1586 VM_BUG_ON(new.frozen);
1587 new.frozen = 1;
1588
1589 if (!__cmpxchg_double_slab(s, page,
1590 freelist, counters,
1591 new.freelist, new.counters,
1592 "acquire_slab"))
1593 return NULL;
1594
1595 remove_partial(n, page);
1596 WARN_ON(!freelist);
1597 return freelist;
1598}
1599
1600static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1601static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1602
1603/*
1604 * Try to allocate a partial slab from a specific node.
1605 */
1606static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1607 struct kmem_cache_cpu *c, gfp_t flags)
1608{
1609 struct page *page, *page2;
1610 void *object = NULL;
1611 int available = 0;
1612 int objects;
1613
1614 /*
1615 * Racy check. If we mistakenly see no partial slabs then we
1616 * just allocate an empty slab. If we mistakenly try to get a
1617 * partial slab and there is none available then get_partials()
1618 * will return NULL.
1619 */
1620 if (!n || !n->nr_partial)
1621 return NULL;
1622
1623 spin_lock(&n->list_lock);
1624 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1625 void *t;
1626
1627 if (!pfmemalloc_match(page, flags))
1628 continue;
1629
1630 t = acquire_slab(s, n, page, object == NULL, &objects);
1631 if (!t)
1632 break;
1633
1634 available += objects;
1635 if (!object) {
1636 c->page = page;
1637 stat(s, ALLOC_FROM_PARTIAL);
1638 object = t;
1639 } else {
1640 put_cpu_partial(s, page, 0);
1641 stat(s, CPU_PARTIAL_NODE);
1642 }
1643 if (!kmem_cache_has_cpu_partial(s)
1644 || available > s->cpu_partial / 2)
1645 break;
1646
1647 }
1648 spin_unlock(&n->list_lock);
1649 return object;
1650}
1651
1652/*
1653 * Get a page from somewhere. Search in increasing NUMA distances.
1654 */
1655static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1656 struct kmem_cache_cpu *c)
1657{
1658#ifdef CONFIG_NUMA
1659 struct zonelist *zonelist;
1660 struct zoneref *z;
1661 struct zone *zone;
1662 enum zone_type high_zoneidx = gfp_zone(flags);
1663 void *object;
1664 unsigned int cpuset_mems_cookie;
1665
1666 /*
1667 * The defrag ratio allows a configuration of the tradeoffs between
1668 * inter node defragmentation and node local allocations. A lower
1669 * defrag_ratio increases the tendency to do local allocations
1670 * instead of attempting to obtain partial slabs from other nodes.
1671 *
1672 * If the defrag_ratio is set to 0 then kmalloc() always
1673 * returns node local objects. If the ratio is higher then kmalloc()
1674 * may return off node objects because partial slabs are obtained
1675 * from other nodes and filled up.
1676 *
1677 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1678 * defrag_ratio = 1000) then every (well almost) allocation will
1679 * first attempt to defrag slab caches on other nodes. This means
1680 * scanning over all nodes to look for partial slabs which may be
1681 * expensive if we do it every time we are trying to find a slab
1682 * with available objects.
1683 */
1684 if (!s->remote_node_defrag_ratio ||
1685 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1686 return NULL;
1687
1688 do {
1689 cpuset_mems_cookie = read_mems_allowed_begin();
1690 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1691 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1692 struct kmem_cache_node *n;
1693
1694 n = get_node(s, zone_to_nid(zone));
1695
1696 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1697 n->nr_partial > s->min_partial) {
1698 object = get_partial_node(s, n, c, flags);
1699 if (object) {
1700 /*
1701 * Don't check read_mems_allowed_retry()
1702 * here - if mems_allowed was updated in
1703 * parallel, that was a harmless race
1704 * between allocation and the cpuset
1705 * update
1706 */
1707 return object;
1708 }
1709 }
1710 }
1711 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1712#endif
1713 return NULL;
1714}
1715
1716/*
1717 * Get a partial page, lock it and return it.
1718 */
1719static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1720 struct kmem_cache_cpu *c)
1721{
1722 void *object;
1723 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1724
1725 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1726 if (object || node != NUMA_NO_NODE)
1727 return object;
1728
1729 return get_any_partial(s, flags, c);
1730}
1731
1732#ifdef CONFIG_PREEMPT
1733/*
1734 * Calculate the next globally unique transaction for disambiguiation
1735 * during cmpxchg. The transactions start with the cpu number and are then
1736 * incremented by CONFIG_NR_CPUS.
1737 */
1738#define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1739#else
1740/*
1741 * No preemption supported therefore also no need to check for
1742 * different cpus.
1743 */
1744#define TID_STEP 1
1745#endif
1746
1747static inline unsigned long next_tid(unsigned long tid)
1748{
1749 return tid + TID_STEP;
1750}
1751
1752static inline unsigned int tid_to_cpu(unsigned long tid)
1753{
1754 return tid % TID_STEP;
1755}
1756
1757static inline unsigned long tid_to_event(unsigned long tid)
1758{
1759 return tid / TID_STEP;
1760}
1761
1762static inline unsigned int init_tid(int cpu)
1763{
1764 return cpu;
1765}
1766
1767static inline void note_cmpxchg_failure(const char *n,
1768 const struct kmem_cache *s, unsigned long tid)
1769{
1770#ifdef SLUB_DEBUG_CMPXCHG
1771 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1772
1773 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1774
1775#ifdef CONFIG_PREEMPT
1776 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1777 printk("due to cpu change %d -> %d\n",
1778 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1779 else
1780#endif
1781 if (tid_to_event(tid) != tid_to_event(actual_tid))
1782 printk("due to cpu running other code. Event %ld->%ld\n",
1783 tid_to_event(tid), tid_to_event(actual_tid));
1784 else
1785 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1786 actual_tid, tid, next_tid(tid));
1787#endif
1788 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1789}
1790
1791static void init_kmem_cache_cpus(struct kmem_cache *s)
1792{
1793 int cpu;
1794
1795 for_each_possible_cpu(cpu)
1796 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1797}
1798
1799/*
1800 * Remove the cpu slab
1801 */
1802static void deactivate_slab(struct kmem_cache *s, struct page *page,
1803 void *freelist)
1804{
1805 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1806 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1807 int lock = 0;
1808 enum slab_modes l = M_NONE, m = M_NONE;
1809 void *nextfree;
1810 int tail = DEACTIVATE_TO_HEAD;
1811 struct page new;
1812 struct page old;
1813
1814 if (page->freelist) {
1815 stat(s, DEACTIVATE_REMOTE_FREES);
1816 tail = DEACTIVATE_TO_TAIL;
1817 }
1818
1819 /*
1820 * Stage one: Free all available per cpu objects back
1821 * to the page freelist while it is still frozen. Leave the
1822 * last one.
1823 *
1824 * There is no need to take the list->lock because the page
1825 * is still frozen.
1826 */
1827 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1828 void *prior;
1829 unsigned long counters;
1830
1831 do {
1832 prior = page->freelist;
1833 counters = page->counters;
1834 set_freepointer(s, freelist, prior);
1835 new.counters = counters;
1836 new.inuse--;
1837 VM_BUG_ON(!new.frozen);
1838
1839 } while (!__cmpxchg_double_slab(s, page,
1840 prior, counters,
1841 freelist, new.counters,
1842 "drain percpu freelist"));
1843
1844 freelist = nextfree;
1845 }
1846
1847 /*
1848 * Stage two: Ensure that the page is unfrozen while the
1849 * list presence reflects the actual number of objects
1850 * during unfreeze.
1851 *
1852 * We setup the list membership and then perform a cmpxchg
1853 * with the count. If there is a mismatch then the page
1854 * is not unfrozen but the page is on the wrong list.
1855 *
1856 * Then we restart the process which may have to remove
1857 * the page from the list that we just put it on again
1858 * because the number of objects in the slab may have
1859 * changed.
1860 */
1861redo:
1862
1863 old.freelist = page->freelist;
1864 old.counters = page->counters;
1865 VM_BUG_ON(!old.frozen);
1866
1867 /* Determine target state of the slab */
1868 new.counters = old.counters;
1869 if (freelist) {
1870 new.inuse--;
1871 set_freepointer(s, freelist, old.freelist);
1872 new.freelist = freelist;
1873 } else
1874 new.freelist = old.freelist;
1875
1876 new.frozen = 0;
1877
1878 if (!new.inuse && n->nr_partial > s->min_partial)
1879 m = M_FREE;
1880 else if (new.freelist) {
1881 m = M_PARTIAL;
1882 if (!lock) {
1883 lock = 1;
1884 /*
1885 * Taking the spinlock removes the possiblity
1886 * that acquire_slab() will see a slab page that
1887 * is frozen
1888 */
1889 spin_lock(&n->list_lock);
1890 }
1891 } else {
1892 m = M_FULL;
1893 if (kmem_cache_debug(s) && !lock) {
1894 lock = 1;
1895 /*
1896 * This also ensures that the scanning of full
1897 * slabs from diagnostic functions will not see
1898 * any frozen slabs.
1899 */
1900 spin_lock(&n->list_lock);
1901 }
1902 }
1903
1904 if (l != m) {
1905
1906 if (l == M_PARTIAL)
1907
1908 remove_partial(n, page);
1909
1910 else if (l == M_FULL)
1911
1912 remove_full(s, n, page);
1913
1914 if (m == M_PARTIAL) {
1915
1916 add_partial(n, page, tail);
1917 stat(s, tail);
1918
1919 } else if (m == M_FULL) {
1920
1921 stat(s, DEACTIVATE_FULL);
1922 add_full(s, n, page);
1923
1924 }
1925 }
1926
1927 l = m;
1928 if (!__cmpxchg_double_slab(s, page,
1929 old.freelist, old.counters,
1930 new.freelist, new.counters,
1931 "unfreezing slab"))
1932 goto redo;
1933
1934 if (lock)
1935 spin_unlock(&n->list_lock);
1936
1937 if (m == M_FREE) {
1938 stat(s, DEACTIVATE_EMPTY);
1939 discard_slab(s, page);
1940 stat(s, FREE_SLAB);
1941 }
1942}
1943
1944/*
1945 * Unfreeze all the cpu partial slabs.
1946 *
1947 * This function must be called with interrupts disabled
1948 * for the cpu using c (or some other guarantee must be there
1949 * to guarantee no concurrent accesses).
1950 */
1951static void unfreeze_partials(struct kmem_cache *s,
1952 struct kmem_cache_cpu *c)
1953{
1954#ifdef CONFIG_SLUB_CPU_PARTIAL
1955 struct kmem_cache_node *n = NULL, *n2 = NULL;
1956 struct page *page, *discard_page = NULL;
1957
1958 while ((page = c->partial)) {
1959 struct page new;
1960 struct page old;
1961
1962 c->partial = page->next;
1963
1964 n2 = get_node(s, page_to_nid(page));
1965 if (n != n2) {
1966 if (n)
1967 spin_unlock(&n->list_lock);
1968
1969 n = n2;
1970 spin_lock(&n->list_lock);
1971 }
1972
1973 do {
1974
1975 old.freelist = page->freelist;
1976 old.counters = page->counters;
1977 VM_BUG_ON(!old.frozen);
1978
1979 new.counters = old.counters;
1980 new.freelist = old.freelist;
1981
1982 new.frozen = 0;
1983
1984 } while (!__cmpxchg_double_slab(s, page,
1985 old.freelist, old.counters,
1986 new.freelist, new.counters,
1987 "unfreezing slab"));
1988
1989 if (unlikely(!new.inuse && n->nr_partial > s->min_partial)) {
1990 page->next = discard_page;
1991 discard_page = page;
1992 } else {
1993 add_partial(n, page, DEACTIVATE_TO_TAIL);
1994 stat(s, FREE_ADD_PARTIAL);
1995 }
1996 }
1997
1998 if (n)
1999 spin_unlock(&n->list_lock);
2000
2001 while (discard_page) {
2002 page = discard_page;
2003 discard_page = discard_page->next;
2004
2005 stat(s, DEACTIVATE_EMPTY);
2006 discard_slab(s, page);
2007 stat(s, FREE_SLAB);
2008 }
2009#endif
2010}
2011
2012/*
2013 * Put a page that was just frozen (in __slab_free) into a partial page
2014 * slot if available. This is done without interrupts disabled and without
2015 * preemption disabled. The cmpxchg is racy and may put the partial page
2016 * onto a random cpus partial slot.
2017 *
2018 * If we did not find a slot then simply move all the partials to the
2019 * per node partial list.
2020 */
2021static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2022{
2023#ifdef CONFIG_SLUB_CPU_PARTIAL
2024 struct page *oldpage;
2025 int pages;
2026 int pobjects;
2027
2028 do {
2029 pages = 0;
2030 pobjects = 0;
2031 oldpage = this_cpu_read(s->cpu_slab->partial);
2032
2033 if (oldpage) {
2034 pobjects = oldpage->pobjects;
2035 pages = oldpage->pages;
2036 if (drain && pobjects > s->cpu_partial) {
2037 unsigned long flags;
2038 /*
2039 * partial array is full. Move the existing
2040 * set to the per node partial list.
2041 */
2042 local_irq_save(flags);
2043 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2044 local_irq_restore(flags);
2045 oldpage = NULL;
2046 pobjects = 0;
2047 pages = 0;
2048 stat(s, CPU_PARTIAL_DRAIN);
2049 }
2050 }
2051
2052 pages++;
2053 pobjects += page->objects - page->inuse;
2054
2055 page->pages = pages;
2056 page->pobjects = pobjects;
2057 page->next = oldpage;
2058
2059 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2060 != oldpage);
2061#endif
2062}
2063
2064static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2065{
2066 stat(s, CPUSLAB_FLUSH);
2067 deactivate_slab(s, c->page, c->freelist);
2068
2069 c->tid = next_tid(c->tid);
2070 c->page = NULL;
2071 c->freelist = NULL;
2072}
2073
2074/*
2075 * Flush cpu slab.
2076 *
2077 * Called from IPI handler with interrupts disabled.
2078 */
2079static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2080{
2081 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2082
2083 if (likely(c)) {
2084 if (c->page)
2085 flush_slab(s, c);
2086
2087 unfreeze_partials(s, c);
2088 }
2089}
2090
2091static void flush_cpu_slab(void *d)
2092{
2093 struct kmem_cache *s = d;
2094
2095 __flush_cpu_slab(s, smp_processor_id());
2096}
2097
2098static bool has_cpu_slab(int cpu, void *info)
2099{
2100 struct kmem_cache *s = info;
2101 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2102
2103 return c->page || c->partial;
2104}
2105
2106static void flush_all(struct kmem_cache *s)
2107{
2108 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2109}
2110
2111/*
2112 * Check if the objects in a per cpu structure fit numa
2113 * locality expectations.
2114 */
2115static inline int node_match(struct page *page, int node)
2116{
2117#ifdef CONFIG_NUMA
2118 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2119 return 0;
2120#endif
2121 return 1;
2122}
2123
2124static int count_free(struct page *page)
2125{
2126 return page->objects - page->inuse;
2127}
2128
2129static unsigned long count_partial(struct kmem_cache_node *n,
2130 int (*get_count)(struct page *))
2131{
2132 unsigned long flags;
2133 unsigned long x = 0;
2134 struct page *page;
2135
2136 spin_lock_irqsave(&n->list_lock, flags);
2137 list_for_each_entry(page, &n->partial, lru)
2138 x += get_count(page);
2139 spin_unlock_irqrestore(&n->list_lock, flags);
2140 return x;
2141}
2142
2143static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2144{
2145#ifdef CONFIG_SLUB_DEBUG
2146 return atomic_long_read(&n->total_objects);
2147#else
2148 return 0;
2149#endif
2150}
2151
2152static noinline void
2153slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2154{
2155 int node;
2156
2157 printk(KERN_WARNING
2158 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2159 nid, gfpflags);
2160 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
2161 "default order: %d, min order: %d\n", s->name, s->object_size,
2162 s->size, oo_order(s->oo), oo_order(s->min));
2163
2164 if (oo_order(s->min) > get_order(s->object_size))
2165 printk(KERN_WARNING " %s debugging increased min order, use "
2166 "slub_debug=O to disable.\n", s->name);
2167
2168 for_each_online_node(node) {
2169 struct kmem_cache_node *n = get_node(s, node);
2170 unsigned long nr_slabs;
2171 unsigned long nr_objs;
2172 unsigned long nr_free;
2173
2174 if (!n)
2175 continue;
2176
2177 nr_free = count_partial(n, count_free);
2178 nr_slabs = node_nr_slabs(n);
2179 nr_objs = node_nr_objs(n);
2180
2181 printk(KERN_WARNING
2182 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2183 node, nr_slabs, nr_objs, nr_free);
2184 }
2185}
2186
2187static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2188 int node, struct kmem_cache_cpu **pc)
2189{
2190 void *freelist;
2191 struct kmem_cache_cpu *c = *pc;
2192 struct page *page;
2193
2194 freelist = get_partial(s, flags, node, c);
2195
2196 if (freelist)
2197 return freelist;
2198
2199 page = new_slab(s, flags, node);
2200 if (page) {
2201 c = __this_cpu_ptr(s->cpu_slab);
2202 if (c->page)
2203 flush_slab(s, c);
2204
2205 /*
2206 * No other reference to the page yet so we can
2207 * muck around with it freely without cmpxchg
2208 */
2209 freelist = page->freelist;
2210 page->freelist = NULL;
2211
2212 stat(s, ALLOC_SLAB);
2213 c->page = page;
2214 *pc = c;
2215 } else
2216 freelist = NULL;
2217
2218 return freelist;
2219}
2220
2221static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2222{
2223 if (unlikely(PageSlabPfmemalloc(page)))
2224 return gfp_pfmemalloc_allowed(gfpflags);
2225
2226 return true;
2227}
2228
2229/*
2230 * Check the page->freelist of a page and either transfer the freelist to the
2231 * per cpu freelist or deactivate the page.
2232 *
2233 * The page is still frozen if the return value is not NULL.
2234 *
2235 * If this function returns NULL then the page has been unfrozen.
2236 *
2237 * This function must be called with interrupt disabled.
2238 */
2239static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2240{
2241 struct page new;
2242 unsigned long counters;
2243 void *freelist;
2244
2245 do {
2246 freelist = page->freelist;
2247 counters = page->counters;
2248
2249 new.counters = counters;
2250 VM_BUG_ON(!new.frozen);
2251
2252 new.inuse = page->objects;
2253 new.frozen = freelist != NULL;
2254
2255 } while (!__cmpxchg_double_slab(s, page,
2256 freelist, counters,
2257 NULL, new.counters,
2258 "get_freelist"));
2259
2260 return freelist;
2261}
2262
2263/*
2264 * Slow path. The lockless freelist is empty or we need to perform
2265 * debugging duties.
2266 *
2267 * Processing is still very fast if new objects have been freed to the
2268 * regular freelist. In that case we simply take over the regular freelist
2269 * as the lockless freelist and zap the regular freelist.
2270 *
2271 * If that is not working then we fall back to the partial lists. We take the
2272 * first element of the freelist as the object to allocate now and move the
2273 * rest of the freelist to the lockless freelist.
2274 *
2275 * And if we were unable to get a new slab from the partial slab lists then
2276 * we need to allocate a new slab. This is the slowest path since it involves
2277 * a call to the page allocator and the setup of a new slab.
2278 */
2279static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2280 unsigned long addr, struct kmem_cache_cpu *c)
2281{
2282 void *freelist;
2283 struct page *page;
2284 unsigned long flags;
2285
2286 local_irq_save(flags);
2287#ifdef CONFIG_PREEMPT
2288 /*
2289 * We may have been preempted and rescheduled on a different
2290 * cpu before disabling interrupts. Need to reload cpu area
2291 * pointer.
2292 */
2293 c = this_cpu_ptr(s->cpu_slab);
2294#endif
2295
2296 page = c->page;
2297 if (!page)
2298 goto new_slab;
2299redo:
2300
2301 if (unlikely(!node_match(page, node))) {
2302 stat(s, ALLOC_NODE_MISMATCH);
2303 deactivate_slab(s, page, c->freelist);
2304 c->page = NULL;
2305 c->freelist = NULL;
2306 goto new_slab;
2307 }
2308
2309 /*
2310 * By rights, we should be searching for a slab page that was
2311 * PFMEMALLOC but right now, we are losing the pfmemalloc
2312 * information when the page leaves the per-cpu allocator
2313 */
2314 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2315 deactivate_slab(s, page, c->freelist);
2316 c->page = NULL;
2317 c->freelist = NULL;
2318 goto new_slab;
2319 }
2320
2321 /* must check again c->freelist in case of cpu migration or IRQ */
2322 freelist = c->freelist;
2323 if (freelist)
2324 goto load_freelist;
2325
2326 stat(s, ALLOC_SLOWPATH);
2327
2328 freelist = get_freelist(s, page);
2329
2330 if (!freelist) {
2331 c->page = NULL;
2332 stat(s, DEACTIVATE_BYPASS);
2333 goto new_slab;
2334 }
2335
2336 stat(s, ALLOC_REFILL);
2337
2338load_freelist:
2339 /*
2340 * freelist is pointing to the list of objects to be used.
2341 * page is pointing to the page from which the objects are obtained.
2342 * That page must be frozen for per cpu allocations to work.
2343 */
2344 VM_BUG_ON(!c->page->frozen);
2345 c->freelist = get_freepointer(s, freelist);
2346 c->tid = next_tid(c->tid);
2347 local_irq_restore(flags);
2348 return freelist;
2349
2350new_slab:
2351
2352 if (c->partial) {
2353 page = c->page = c->partial;
2354 c->partial = page->next;
2355 stat(s, CPU_PARTIAL_ALLOC);
2356 c->freelist = NULL;
2357 goto redo;
2358 }
2359
2360 freelist = new_slab_objects(s, gfpflags, node, &c);
2361
2362 if (unlikely(!freelist)) {
2363 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2364 slab_out_of_memory(s, gfpflags, node);
2365
2366 local_irq_restore(flags);
2367 return NULL;
2368 }
2369
2370 page = c->page;
2371 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2372 goto load_freelist;
2373
2374 /* Only entered in the debug case */
2375 if (kmem_cache_debug(s) &&
2376 !alloc_debug_processing(s, page, freelist, addr))
2377 goto new_slab; /* Slab failed checks. Next slab needed */
2378
2379 deactivate_slab(s, page, get_freepointer(s, freelist));
2380 c->page = NULL;
2381 c->freelist = NULL;
2382 local_irq_restore(flags);
2383 return freelist;
2384}
2385
2386/*
2387 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2388 * have the fastpath folded into their functions. So no function call
2389 * overhead for requests that can be satisfied on the fastpath.
2390 *
2391 * The fastpath works by first checking if the lockless freelist can be used.
2392 * If not then __slab_alloc is called for slow processing.
2393 *
2394 * Otherwise we can simply pick the next object from the lockless free list.
2395 */
2396static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2397 gfp_t gfpflags, int node, unsigned long addr)
2398{
2399 void **object;
2400 struct kmem_cache_cpu *c;
2401 struct page *page;
2402 unsigned long tid;
2403
2404 if (slab_pre_alloc_hook(s, gfpflags))
2405 return NULL;
2406
2407 s = memcg_kmem_get_cache(s, gfpflags);
2408redo:
2409 /*
2410 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2411 * enabled. We may switch back and forth between cpus while
2412 * reading from one cpu area. That does not matter as long
2413 * as we end up on the original cpu again when doing the cmpxchg.
2414 *
2415 * Preemption is disabled for the retrieval of the tid because that
2416 * must occur from the current processor. We cannot allow rescheduling
2417 * on a different processor between the determination of the pointer
2418 * and the retrieval of the tid.
2419 */
2420 preempt_disable();
2421 c = __this_cpu_ptr(s->cpu_slab);
2422
2423 /*
2424 * The transaction ids are globally unique per cpu and per operation on
2425 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2426 * occurs on the right processor and that there was no operation on the
2427 * linked list in between.
2428 */
2429 tid = c->tid;
2430 preempt_enable();
2431
2432 object = c->freelist;
2433 page = c->page;
2434 if (unlikely(!object || !node_match(page, node)))
2435 object = __slab_alloc(s, gfpflags, node, addr, c);
2436
2437 else {
2438 void *next_object = get_freepointer_safe(s, object);
2439
2440 /*
2441 * The cmpxchg will only match if there was no additional
2442 * operation and if we are on the right processor.
2443 *
2444 * The cmpxchg does the following atomically (without lock
2445 * semantics!)
2446 * 1. Relocate first pointer to the current per cpu area.
2447 * 2. Verify that tid and freelist have not been changed
2448 * 3. If they were not changed replace tid and freelist
2449 *
2450 * Since this is without lock semantics the protection is only
2451 * against code executing on this cpu *not* from access by
2452 * other cpus.
2453 */
2454 if (unlikely(!this_cpu_cmpxchg_double(
2455 s->cpu_slab->freelist, s->cpu_slab->tid,
2456 object, tid,
2457 next_object, next_tid(tid)))) {
2458
2459 note_cmpxchg_failure("slab_alloc", s, tid);
2460 goto redo;
2461 }
2462 prefetch_freepointer(s, next_object);
2463 stat(s, ALLOC_FASTPATH);
2464 }
2465
2466 if (unlikely(gfpflags & __GFP_ZERO) && object)
2467 memset(object, 0, s->object_size);
2468
2469 slab_post_alloc_hook(s, gfpflags, object);
2470
2471 return object;
2472}
2473
2474static __always_inline void *slab_alloc(struct kmem_cache *s,
2475 gfp_t gfpflags, unsigned long addr)
2476{
2477 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2478}
2479
2480void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2481{
2482 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2483
2484 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2485 s->size, gfpflags);
2486
2487 return ret;
2488}
2489EXPORT_SYMBOL(kmem_cache_alloc);
2490
2491#ifdef CONFIG_TRACING
2492void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2493{
2494 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2495 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2496 return ret;
2497}
2498EXPORT_SYMBOL(kmem_cache_alloc_trace);
2499#endif
2500
2501#ifdef CONFIG_NUMA
2502void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2503{
2504 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2505
2506 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2507 s->object_size, s->size, gfpflags, node);
2508
2509 return ret;
2510}
2511EXPORT_SYMBOL(kmem_cache_alloc_node);
2512
2513#ifdef CONFIG_TRACING
2514void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2515 gfp_t gfpflags,
2516 int node, size_t size)
2517{
2518 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2519
2520 trace_kmalloc_node(_RET_IP_, ret,
2521 size, s->size, gfpflags, node);
2522 return ret;
2523}
2524EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2525#endif
2526#endif
2527
2528/*
2529 * Slow patch handling. This may still be called frequently since objects
2530 * have a longer lifetime than the cpu slabs in most processing loads.
2531 *
2532 * So we still attempt to reduce cache line usage. Just take the slab
2533 * lock and free the item. If there is no additional partial page
2534 * handling required then we can return immediately.
2535 */
2536static void __slab_free(struct kmem_cache *s, struct page *page,
2537 void *x, unsigned long addr)
2538{
2539 void *prior;
2540 void **object = (void *)x;
2541 int was_frozen;
2542 struct page new;
2543 unsigned long counters;
2544 struct kmem_cache_node *n = NULL;
2545 unsigned long uninitialized_var(flags);
2546
2547 stat(s, FREE_SLOWPATH);
2548
2549 if (kmem_cache_debug(s) &&
2550 !(n = free_debug_processing(s, page, x, addr, &flags)))
2551 return;
2552
2553 do {
2554 if (unlikely(n)) {
2555 spin_unlock_irqrestore(&n->list_lock, flags);
2556 n = NULL;
2557 }
2558 prior = page->freelist;
2559 counters = page->counters;
2560 set_freepointer(s, object, prior);
2561 new.counters = counters;
2562 was_frozen = new.frozen;
2563 new.inuse--;
2564 if ((!new.inuse || !prior) && !was_frozen) {
2565
2566 if (kmem_cache_has_cpu_partial(s) && !prior) {
2567
2568 /*
2569 * Slab was on no list before and will be
2570 * partially empty
2571 * We can defer the list move and instead
2572 * freeze it.
2573 */
2574 new.frozen = 1;
2575
2576 } else { /* Needs to be taken off a list */
2577
2578 n = get_node(s, page_to_nid(page));
2579 /*
2580 * Speculatively acquire the list_lock.
2581 * If the cmpxchg does not succeed then we may
2582 * drop the list_lock without any processing.
2583 *
2584 * Otherwise the list_lock will synchronize with
2585 * other processors updating the list of slabs.
2586 */
2587 spin_lock_irqsave(&n->list_lock, flags);
2588
2589 }
2590 }
2591
2592 } while (!cmpxchg_double_slab(s, page,
2593 prior, counters,
2594 object, new.counters,
2595 "__slab_free"));
2596
2597 if (likely(!n)) {
2598
2599 /*
2600 * If we just froze the page then put it onto the
2601 * per cpu partial list.
2602 */
2603 if (new.frozen && !was_frozen) {
2604 put_cpu_partial(s, page, 1);
2605 stat(s, CPU_PARTIAL_FREE);
2606 }
2607 /*
2608 * The list lock was not taken therefore no list
2609 * activity can be necessary.
2610 */
2611 if (was_frozen)
2612 stat(s, FREE_FROZEN);
2613 return;
2614 }
2615
2616 if (unlikely(!new.inuse && n->nr_partial > s->min_partial))
2617 goto slab_empty;
2618
2619 /*
2620 * Objects left in the slab. If it was not on the partial list before
2621 * then add it.
2622 */
2623 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2624 if (kmem_cache_debug(s))
2625 remove_full(s, n, page);
2626 add_partial(n, page, DEACTIVATE_TO_TAIL);
2627 stat(s, FREE_ADD_PARTIAL);
2628 }
2629 spin_unlock_irqrestore(&n->list_lock, flags);
2630 return;
2631
2632slab_empty:
2633 if (prior) {
2634 /*
2635 * Slab on the partial list.
2636 */
2637 remove_partial(n, page);
2638 stat(s, FREE_REMOVE_PARTIAL);
2639 } else {
2640 /* Slab must be on the full list */
2641 remove_full(s, n, page);
2642 }
2643
2644 spin_unlock_irqrestore(&n->list_lock, flags);
2645 stat(s, FREE_SLAB);
2646 discard_slab(s, page);
2647}
2648
2649/*
2650 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2651 * can perform fastpath freeing without additional function calls.
2652 *
2653 * The fastpath is only possible if we are freeing to the current cpu slab
2654 * of this processor. This typically the case if we have just allocated
2655 * the item before.
2656 *
2657 * If fastpath is not possible then fall back to __slab_free where we deal
2658 * with all sorts of special processing.
2659 */
2660static __always_inline void slab_free(struct kmem_cache *s,
2661 struct page *page, void *x, unsigned long addr)
2662{
2663 void **object = (void *)x;
2664 struct kmem_cache_cpu *c;
2665 unsigned long tid;
2666
2667 slab_free_hook(s, x);
2668
2669redo:
2670 /*
2671 * Determine the currently cpus per cpu slab.
2672 * The cpu may change afterward. However that does not matter since
2673 * data is retrieved via this pointer. If we are on the same cpu
2674 * during the cmpxchg then the free will succedd.
2675 */
2676 preempt_disable();
2677 c = __this_cpu_ptr(s->cpu_slab);
2678
2679 tid = c->tid;
2680 preempt_enable();
2681
2682 if (likely(page == c->page)) {
2683 set_freepointer(s, object, c->freelist);
2684
2685 if (unlikely(!this_cpu_cmpxchg_double(
2686 s->cpu_slab->freelist, s->cpu_slab->tid,
2687 c->freelist, tid,
2688 object, next_tid(tid)))) {
2689
2690 note_cmpxchg_failure("slab_free", s, tid);
2691 goto redo;
2692 }
2693 stat(s, FREE_FASTPATH);
2694 } else
2695 __slab_free(s, page, x, addr);
2696
2697}
2698
2699void kmem_cache_free(struct kmem_cache *s, void *x)
2700{
2701 s = cache_from_obj(s, x);
2702 if (!s)
2703 return;
2704 slab_free(s, virt_to_head_page(x), x, _RET_IP_);
2705 trace_kmem_cache_free(_RET_IP_, x);
2706}
2707EXPORT_SYMBOL(kmem_cache_free);
2708
2709/*
2710 * Object placement in a slab is made very easy because we always start at
2711 * offset 0. If we tune the size of the object to the alignment then we can
2712 * get the required alignment by putting one properly sized object after
2713 * another.
2714 *
2715 * Notice that the allocation order determines the sizes of the per cpu
2716 * caches. Each processor has always one slab available for allocations.
2717 * Increasing the allocation order reduces the number of times that slabs
2718 * must be moved on and off the partial lists and is therefore a factor in
2719 * locking overhead.
2720 */
2721
2722/*
2723 * Mininum / Maximum order of slab pages. This influences locking overhead
2724 * and slab fragmentation. A higher order reduces the number of partial slabs
2725 * and increases the number of allocations possible without having to
2726 * take the list_lock.
2727 */
2728static int slub_min_order;
2729static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2730static int slub_min_objects;
2731
2732/*
2733 * Merge control. If this is set then no merging of slab caches will occur.
2734 * (Could be removed. This was introduced to pacify the merge skeptics.)
2735 */
2736static int slub_nomerge;
2737
2738/*
2739 * Calculate the order of allocation given an slab object size.
2740 *
2741 * The order of allocation has significant impact on performance and other
2742 * system components. Generally order 0 allocations should be preferred since
2743 * order 0 does not cause fragmentation in the page allocator. Larger objects
2744 * be problematic to put into order 0 slabs because there may be too much
2745 * unused space left. We go to a higher order if more than 1/16th of the slab
2746 * would be wasted.
2747 *
2748 * In order to reach satisfactory performance we must ensure that a minimum
2749 * number of objects is in one slab. Otherwise we may generate too much
2750 * activity on the partial lists which requires taking the list_lock. This is
2751 * less a concern for large slabs though which are rarely used.
2752 *
2753 * slub_max_order specifies the order where we begin to stop considering the
2754 * number of objects in a slab as critical. If we reach slub_max_order then
2755 * we try to keep the page order as low as possible. So we accept more waste
2756 * of space in favor of a small page order.
2757 *
2758 * Higher order allocations also allow the placement of more objects in a
2759 * slab and thereby reduce object handling overhead. If the user has
2760 * requested a higher mininum order then we start with that one instead of
2761 * the smallest order which will fit the object.
2762 */
2763static inline int slab_order(int size, int min_objects,
2764 int max_order, int fract_leftover, int reserved)
2765{
2766 int order;
2767 int rem;
2768 int min_order = slub_min_order;
2769
2770 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2771 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2772
2773 for (order = max(min_order,
2774 fls(min_objects * size - 1) - PAGE_SHIFT);
2775 order <= max_order; order++) {
2776
2777 unsigned long slab_size = PAGE_SIZE << order;
2778
2779 if (slab_size < min_objects * size + reserved)
2780 continue;
2781
2782 rem = (slab_size - reserved) % size;
2783
2784 if (rem <= slab_size / fract_leftover)
2785 break;
2786
2787 }
2788
2789 return order;
2790}
2791
2792static inline int calculate_order(int size, int reserved)
2793{
2794 int order;
2795 int min_objects;
2796 int fraction;
2797 int max_objects;
2798
2799 /*
2800 * Attempt to find best configuration for a slab. This
2801 * works by first attempting to generate a layout with
2802 * the best configuration and backing off gradually.
2803 *
2804 * First we reduce the acceptable waste in a slab. Then
2805 * we reduce the minimum objects required in a slab.
2806 */
2807 min_objects = slub_min_objects;
2808 if (!min_objects)
2809 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2810 max_objects = order_objects(slub_max_order, size, reserved);
2811 min_objects = min(min_objects, max_objects);
2812
2813 while (min_objects > 1) {
2814 fraction = 16;
2815 while (fraction >= 4) {
2816 order = slab_order(size, min_objects,
2817 slub_max_order, fraction, reserved);
2818 if (order <= slub_max_order)
2819 return order;
2820 fraction /= 2;
2821 }
2822 min_objects--;
2823 }
2824
2825 /*
2826 * We were unable to place multiple objects in a slab. Now
2827 * lets see if we can place a single object there.
2828 */
2829 order = slab_order(size, 1, slub_max_order, 1, reserved);
2830 if (order <= slub_max_order)
2831 return order;
2832
2833 /*
2834 * Doh this slab cannot be placed using slub_max_order.
2835 */
2836 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2837 if (order < MAX_ORDER)
2838 return order;
2839 return -ENOSYS;
2840}
2841
2842static void
2843init_kmem_cache_node(struct kmem_cache_node *n)
2844{
2845 n->nr_partial = 0;
2846 spin_lock_init(&n->list_lock);
2847 INIT_LIST_HEAD(&n->partial);
2848#ifdef CONFIG_SLUB_DEBUG
2849 atomic_long_set(&n->nr_slabs, 0);
2850 atomic_long_set(&n->total_objects, 0);
2851 INIT_LIST_HEAD(&n->full);
2852#endif
2853}
2854
2855static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2856{
2857 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2858 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
2859
2860 /*
2861 * Must align to double word boundary for the double cmpxchg
2862 * instructions to work; see __pcpu_double_call_return_bool().
2863 */
2864 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2865 2 * sizeof(void *));
2866
2867 if (!s->cpu_slab)
2868 return 0;
2869
2870 init_kmem_cache_cpus(s);
2871
2872 return 1;
2873}
2874
2875static struct kmem_cache *kmem_cache_node;
2876
2877/*
2878 * No kmalloc_node yet so do it by hand. We know that this is the first
2879 * slab on the node for this slabcache. There are no concurrent accesses
2880 * possible.
2881 *
2882 * Note that this function only works on the kmem_cache_node
2883 * when allocating for the kmem_cache_node. This is used for bootstrapping
2884 * memory on a fresh node that has no slab structures yet.
2885 */
2886static void early_kmem_cache_node_alloc(int node)
2887{
2888 struct page *page;
2889 struct kmem_cache_node *n;
2890
2891 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2892
2893 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2894
2895 BUG_ON(!page);
2896 if (page_to_nid(page) != node) {
2897 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2898 "node %d\n", node);
2899 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2900 "in order to be able to continue\n");
2901 }
2902
2903 n = page->freelist;
2904 BUG_ON(!n);
2905 page->freelist = get_freepointer(kmem_cache_node, n);
2906 page->inuse = 1;
2907 page->frozen = 0;
2908 kmem_cache_node->node[node] = n;
2909#ifdef CONFIG_SLUB_DEBUG
2910 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2911 init_tracking(kmem_cache_node, n);
2912#endif
2913 init_kmem_cache_node(n);
2914 inc_slabs_node(kmem_cache_node, node, page->objects);
2915
2916 /*
2917 * No locks need to be taken here as it has just been
2918 * initialized and there is no concurrent access.
2919 */
2920 __add_partial(n, page, DEACTIVATE_TO_HEAD);
2921}
2922
2923static void free_kmem_cache_nodes(struct kmem_cache *s)
2924{
2925 int node;
2926
2927 for_each_node_state(node, N_NORMAL_MEMORY) {
2928 struct kmem_cache_node *n = s->node[node];
2929
2930 if (n)
2931 kmem_cache_free(kmem_cache_node, n);
2932
2933 s->node[node] = NULL;
2934 }
2935}
2936
2937static int init_kmem_cache_nodes(struct kmem_cache *s)
2938{
2939 int node;
2940
2941 for_each_node_state(node, N_NORMAL_MEMORY) {
2942 struct kmem_cache_node *n;
2943
2944 if (slab_state == DOWN) {
2945 early_kmem_cache_node_alloc(node);
2946 continue;
2947 }
2948 n = kmem_cache_alloc_node(kmem_cache_node,
2949 GFP_KERNEL, node);
2950
2951 if (!n) {
2952 free_kmem_cache_nodes(s);
2953 return 0;
2954 }
2955
2956 s->node[node] = n;
2957 init_kmem_cache_node(n);
2958 }
2959 return 1;
2960}
2961
2962static void set_min_partial(struct kmem_cache *s, unsigned long min)
2963{
2964 if (min < MIN_PARTIAL)
2965 min = MIN_PARTIAL;
2966 else if (min > MAX_PARTIAL)
2967 min = MAX_PARTIAL;
2968 s->min_partial = min;
2969}
2970
2971/*
2972 * calculate_sizes() determines the order and the distribution of data within
2973 * a slab object.
2974 */
2975static int calculate_sizes(struct kmem_cache *s, int forced_order)
2976{
2977 unsigned long flags = s->flags;
2978 unsigned long size = s->object_size;
2979 int order;
2980
2981 /*
2982 * Round up object size to the next word boundary. We can only
2983 * place the free pointer at word boundaries and this determines
2984 * the possible location of the free pointer.
2985 */
2986 size = ALIGN(size, sizeof(void *));
2987
2988#ifdef CONFIG_SLUB_DEBUG
2989 /*
2990 * Determine if we can poison the object itself. If the user of
2991 * the slab may touch the object after free or before allocation
2992 * then we should never poison the object itself.
2993 */
2994 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2995 !s->ctor)
2996 s->flags |= __OBJECT_POISON;
2997 else
2998 s->flags &= ~__OBJECT_POISON;
2999
3000
3001 /*
3002 * If we are Redzoning then check if there is some space between the
3003 * end of the object and the free pointer. If not then add an
3004 * additional word to have some bytes to store Redzone information.
3005 */
3006 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3007 size += sizeof(void *);
3008#endif
3009
3010 /*
3011 * With that we have determined the number of bytes in actual use
3012 * by the object. This is the potential offset to the free pointer.
3013 */
3014 s->inuse = size;
3015
3016 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
3017 s->ctor)) {
3018 /*
3019 * Relocate free pointer after the object if it is not
3020 * permitted to overwrite the first word of the object on
3021 * kmem_cache_free.
3022 *
3023 * This is the case if we do RCU, have a constructor or
3024 * destructor or are poisoning the objects.
3025 */
3026 s->offset = size;
3027 size += sizeof(void *);
3028 }
3029
3030#ifdef CONFIG_SLUB_DEBUG
3031 if (flags & SLAB_STORE_USER)
3032 /*
3033 * Need to store information about allocs and frees after
3034 * the object.
3035 */
3036 size += 2 * sizeof(struct track);
3037
3038 if (flags & SLAB_RED_ZONE)
3039 /*
3040 * Add some empty padding so that we can catch
3041 * overwrites from earlier objects rather than let
3042 * tracking information or the free pointer be
3043 * corrupted if a user writes before the start
3044 * of the object.
3045 */
3046 size += sizeof(void *);
3047#endif
3048
3049 /*
3050 * SLUB stores one object immediately after another beginning from
3051 * offset 0. In order to align the objects we have to simply size
3052 * each object to conform to the alignment.
3053 */
3054 size = ALIGN(size, s->align);
3055 s->size = size;
3056 if (forced_order >= 0)
3057 order = forced_order;
3058 else
3059 order = calculate_order(size, s->reserved);
3060
3061 if (order < 0)
3062 return 0;
3063
3064 s->allocflags = 0;
3065 if (order)
3066 s->allocflags |= __GFP_COMP;
3067
3068 if (s->flags & SLAB_CACHE_DMA)
3069 s->allocflags |= GFP_DMA;
3070
3071 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3072 s->allocflags |= __GFP_RECLAIMABLE;
3073
3074 /*
3075 * Determine the number of objects per slab
3076 */
3077 s->oo = oo_make(order, size, s->reserved);
3078 s->min = oo_make(get_order(size), size, s->reserved);
3079 if (oo_objects(s->oo) > oo_objects(s->max))
3080 s->max = s->oo;
3081
3082 return !!oo_objects(s->oo);
3083}
3084
3085static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3086{
3087 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3088 s->reserved = 0;
3089
3090 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3091 s->reserved = sizeof(struct rcu_head);
3092
3093 if (!calculate_sizes(s, -1))
3094 goto error;
3095 if (disable_higher_order_debug) {
3096 /*
3097 * Disable debugging flags that store metadata if the min slab
3098 * order increased.
3099 */
3100 if (get_order(s->size) > get_order(s->object_size)) {
3101 s->flags &= ~DEBUG_METADATA_FLAGS;
3102 s->offset = 0;
3103 if (!calculate_sizes(s, -1))
3104 goto error;
3105 }
3106 }
3107
3108#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3109 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3110 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3111 /* Enable fast mode */
3112 s->flags |= __CMPXCHG_DOUBLE;
3113#endif
3114
3115 /*
3116 * The larger the object size is, the more pages we want on the partial
3117 * list to avoid pounding the page allocator excessively.
3118 */
3119 set_min_partial(s, ilog2(s->size) / 2);
3120
3121 /*
3122 * cpu_partial determined the maximum number of objects kept in the
3123 * per cpu partial lists of a processor.
3124 *
3125 * Per cpu partial lists mainly contain slabs that just have one
3126 * object freed. If they are used for allocation then they can be
3127 * filled up again with minimal effort. The slab will never hit the
3128 * per node partial lists and therefore no locking will be required.
3129 *
3130 * This setting also determines
3131 *
3132 * A) The number of objects from per cpu partial slabs dumped to the
3133 * per node list when we reach the limit.
3134 * B) The number of objects in cpu partial slabs to extract from the
3135 * per node list when we run out of per cpu objects. We only fetch
3136 * 50% to keep some capacity around for frees.
3137 */
3138 if (!kmem_cache_has_cpu_partial(s))
3139 s->cpu_partial = 0;
3140 else if (s->size >= PAGE_SIZE)
3141 s->cpu_partial = 2;
3142 else if (s->size >= 1024)
3143 s->cpu_partial = 6;
3144 else if (s->size >= 256)
3145 s->cpu_partial = 13;
3146 else
3147 s->cpu_partial = 30;
3148
3149#ifdef CONFIG_NUMA
3150 s->remote_node_defrag_ratio = 1000;
3151#endif
3152 if (!init_kmem_cache_nodes(s))
3153 goto error;
3154
3155 if (alloc_kmem_cache_cpus(s))
3156 return 0;
3157
3158 free_kmem_cache_nodes(s);
3159error:
3160 if (flags & SLAB_PANIC)
3161 panic("Cannot create slab %s size=%lu realsize=%u "
3162 "order=%u offset=%u flags=%lx\n",
3163 s->name, (unsigned long)s->size, s->size,
3164 oo_order(s->oo), s->offset, flags);
3165 return -EINVAL;
3166}
3167
3168static void list_slab_objects(struct kmem_cache *s, struct page *page,
3169 const char *text)
3170{
3171#ifdef CONFIG_SLUB_DEBUG
3172 void *addr = page_address(page);
3173 void *p;
3174 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3175 sizeof(long), GFP_ATOMIC);
3176 if (!map)
3177 return;
3178 slab_err(s, page, text, s->name);
3179 slab_lock(page);
3180
3181 get_map(s, page, map);
3182 for_each_object(p, s, addr, page->objects) {
3183
3184 if (!test_bit(slab_index(p, s, addr), map)) {
3185 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3186 p, p - addr);
3187 print_tracking(s, p);
3188 }
3189 }
3190 slab_unlock(page);
3191 kfree(map);
3192#endif
3193}
3194
3195/*
3196 * Attempt to free all partial slabs on a node.
3197 * This is called from kmem_cache_close(). We must be the last thread
3198 * using the cache and therefore we do not need to lock anymore.
3199 */
3200static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3201{
3202 struct page *page, *h;
3203
3204 list_for_each_entry_safe(page, h, &n->partial, lru) {
3205 if (!page->inuse) {
3206 __remove_partial(n, page);
3207 discard_slab(s, page);
3208 } else {
3209 list_slab_objects(s, page,
3210 "Objects remaining in %s on kmem_cache_close()");
3211 }
3212 }
3213}
3214
3215/*
3216 * Release all resources used by a slab cache.
3217 */
3218static inline int kmem_cache_close(struct kmem_cache *s)
3219{
3220 int node;
3221
3222 flush_all(s);
3223 /* Attempt to free all objects */
3224 for_each_node_state(node, N_NORMAL_MEMORY) {
3225 struct kmem_cache_node *n = get_node(s, node);
3226
3227 free_partial(s, n);
3228 if (n->nr_partial || slabs_node(s, node))
3229 return 1;
3230 }
3231 free_percpu(s->cpu_slab);
3232 free_kmem_cache_nodes(s);
3233 return 0;
3234}
3235
3236int __kmem_cache_shutdown(struct kmem_cache *s)
3237{
3238 return kmem_cache_close(s);
3239}
3240
3241/********************************************************************
3242 * Kmalloc subsystem
3243 *******************************************************************/
3244
3245static int __init setup_slub_min_order(char *str)
3246{
3247 get_option(&str, &slub_min_order);
3248
3249 return 1;
3250}
3251
3252__setup("slub_min_order=", setup_slub_min_order);
3253
3254static int __init setup_slub_max_order(char *str)
3255{
3256 get_option(&str, &slub_max_order);
3257 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3258
3259 return 1;
3260}
3261
3262__setup("slub_max_order=", setup_slub_max_order);
3263
3264static int __init setup_slub_min_objects(char *str)
3265{
3266 get_option(&str, &slub_min_objects);
3267
3268 return 1;
3269}
3270
3271__setup("slub_min_objects=", setup_slub_min_objects);
3272
3273static int __init setup_slub_nomerge(char *str)
3274{
3275 slub_nomerge = 1;
3276 return 1;
3277}
3278
3279__setup("slub_nomerge", setup_slub_nomerge);
3280
3281void *__kmalloc(size_t size, gfp_t flags)
3282{
3283 struct kmem_cache *s;
3284 void *ret;
3285
3286 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3287 return kmalloc_large(size, flags);
3288
3289 s = kmalloc_slab(size, flags);
3290
3291 if (unlikely(ZERO_OR_NULL_PTR(s)))
3292 return s;
3293
3294 ret = slab_alloc(s, flags, _RET_IP_);
3295
3296 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3297
3298 return ret;
3299}
3300EXPORT_SYMBOL(__kmalloc);
3301
3302#ifdef CONFIG_NUMA
3303static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3304{
3305 struct page *page;
3306 void *ptr = NULL;
3307
3308 flags |= __GFP_COMP | __GFP_NOTRACK | __GFP_KMEMCG;
3309 page = alloc_pages_node(node, flags, get_order(size));
3310 if (page)
3311 ptr = page_address(page);
3312
3313 kmalloc_large_node_hook(ptr, size, flags);
3314 return ptr;
3315}
3316
3317void *__kmalloc_node(size_t size, gfp_t flags, int node)
3318{
3319 struct kmem_cache *s;
3320 void *ret;
3321
3322 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3323 ret = kmalloc_large_node(size, flags, node);
3324
3325 trace_kmalloc_node(_RET_IP_, ret,
3326 size, PAGE_SIZE << get_order(size),
3327 flags, node);
3328
3329 return ret;
3330 }
3331
3332 s = kmalloc_slab(size, flags);
3333
3334 if (unlikely(ZERO_OR_NULL_PTR(s)))
3335 return s;
3336
3337 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3338
3339 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3340
3341 return ret;
3342}
3343EXPORT_SYMBOL(__kmalloc_node);
3344#endif
3345
3346size_t ksize(const void *object)
3347{
3348 struct page *page;
3349
3350 if (unlikely(object == ZERO_SIZE_PTR))
3351 return 0;
3352
3353 page = virt_to_head_page(object);
3354
3355 if (unlikely(!PageSlab(page))) {
3356 WARN_ON(!PageCompound(page));
3357 return PAGE_SIZE << compound_order(page);
3358 }
3359
3360 return slab_ksize(page->slab_cache);
3361}
3362EXPORT_SYMBOL(ksize);
3363
3364void kfree(const void *x)
3365{
3366 struct page *page;
3367 void *object = (void *)x;
3368
3369 trace_kfree(_RET_IP_, x);
3370
3371 if (unlikely(ZERO_OR_NULL_PTR(x)))
3372 return;
3373
3374 page = virt_to_head_page(x);
3375 if (unlikely(!PageSlab(page))) {
3376 BUG_ON(!PageCompound(page));
3377 kfree_hook(x);
3378 __free_memcg_kmem_pages(page, compound_order(page));
3379 return;
3380 }
3381 slab_free(page->slab_cache, page, object, _RET_IP_);
3382}
3383EXPORT_SYMBOL(kfree);
3384
3385/*
3386 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3387 * the remaining slabs by the number of items in use. The slabs with the
3388 * most items in use come first. New allocations will then fill those up
3389 * and thus they can be removed from the partial lists.
3390 *
3391 * The slabs with the least items are placed last. This results in them
3392 * being allocated from last increasing the chance that the last objects
3393 * are freed in them.
3394 */
3395int kmem_cache_shrink(struct kmem_cache *s)
3396{
3397 int node;
3398 int i;
3399 struct kmem_cache_node *n;
3400 struct page *page;
3401 struct page *t;
3402 int objects = oo_objects(s->max);
3403 struct list_head *slabs_by_inuse =
3404 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3405 unsigned long flags;
3406
3407 if (!slabs_by_inuse)
3408 return -ENOMEM;
3409
3410 flush_all(s);
3411 for_each_node_state(node, N_NORMAL_MEMORY) {
3412 n = get_node(s, node);
3413
3414 if (!n->nr_partial)
3415 continue;
3416
3417 for (i = 0; i < objects; i++)
3418 INIT_LIST_HEAD(slabs_by_inuse + i);
3419
3420 spin_lock_irqsave(&n->list_lock, flags);
3421
3422 /*
3423 * Build lists indexed by the items in use in each slab.
3424 *
3425 * Note that concurrent frees may occur while we hold the
3426 * list_lock. page->inuse here is the upper limit.
3427 */
3428 list_for_each_entry_safe(page, t, &n->partial, lru) {
3429 list_move(&page->lru, slabs_by_inuse + page->inuse);
3430 if (!page->inuse)
3431 n->nr_partial--;
3432 }
3433
3434 /*
3435 * Rebuild the partial list with the slabs filled up most
3436 * first and the least used slabs at the end.
3437 */
3438 for (i = objects - 1; i > 0; i--)
3439 list_splice(slabs_by_inuse + i, n->partial.prev);
3440
3441 spin_unlock_irqrestore(&n->list_lock, flags);
3442
3443 /* Release empty slabs */
3444 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3445 discard_slab(s, page);
3446 }
3447
3448 kfree(slabs_by_inuse);
3449 return 0;
3450}
3451EXPORT_SYMBOL(kmem_cache_shrink);
3452
3453static int slab_mem_going_offline_callback(void *arg)
3454{
3455 struct kmem_cache *s;
3456
3457 mutex_lock(&slab_mutex);
3458 list_for_each_entry(s, &slab_caches, list)
3459 kmem_cache_shrink(s);
3460 mutex_unlock(&slab_mutex);
3461
3462 return 0;
3463}
3464
3465static void slab_mem_offline_callback(void *arg)
3466{
3467 struct kmem_cache_node *n;
3468 struct kmem_cache *s;
3469 struct memory_notify *marg = arg;
3470 int offline_node;
3471
3472 offline_node = marg->status_change_nid_normal;
3473
3474 /*
3475 * If the node still has available memory. we need kmem_cache_node
3476 * for it yet.
3477 */
3478 if (offline_node < 0)
3479 return;
3480
3481 mutex_lock(&slab_mutex);
3482 list_for_each_entry(s, &slab_caches, list) {
3483 n = get_node(s, offline_node);
3484 if (n) {
3485 /*
3486 * if n->nr_slabs > 0, slabs still exist on the node
3487 * that is going down. We were unable to free them,
3488 * and offline_pages() function shouldn't call this
3489 * callback. So, we must fail.
3490 */
3491 BUG_ON(slabs_node(s, offline_node));
3492
3493 s->node[offline_node] = NULL;
3494 kmem_cache_free(kmem_cache_node, n);
3495 }
3496 }
3497 mutex_unlock(&slab_mutex);
3498}
3499
3500static int slab_mem_going_online_callback(void *arg)
3501{
3502 struct kmem_cache_node *n;
3503 struct kmem_cache *s;
3504 struct memory_notify *marg = arg;
3505 int nid = marg->status_change_nid_normal;
3506 int ret = 0;
3507
3508 /*
3509 * If the node's memory is already available, then kmem_cache_node is
3510 * already created. Nothing to do.
3511 */
3512 if (nid < 0)
3513 return 0;
3514
3515 /*
3516 * We are bringing a node online. No memory is available yet. We must
3517 * allocate a kmem_cache_node structure in order to bring the node
3518 * online.
3519 */
3520 mutex_lock(&slab_mutex);
3521 list_for_each_entry(s, &slab_caches, list) {
3522 /*
3523 * XXX: kmem_cache_alloc_node will fallback to other nodes
3524 * since memory is not yet available from the node that
3525 * is brought up.
3526 */
3527 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3528 if (!n) {
3529 ret = -ENOMEM;
3530 goto out;
3531 }
3532 init_kmem_cache_node(n);
3533 s->node[nid] = n;
3534 }
3535out:
3536 mutex_unlock(&slab_mutex);
3537 return ret;
3538}
3539
3540static int slab_memory_callback(struct notifier_block *self,
3541 unsigned long action, void *arg)
3542{
3543 int ret = 0;
3544
3545 switch (action) {
3546 case MEM_GOING_ONLINE:
3547 ret = slab_mem_going_online_callback(arg);
3548 break;
3549 case MEM_GOING_OFFLINE:
3550 ret = slab_mem_going_offline_callback(arg);
3551 break;
3552 case MEM_OFFLINE:
3553 case MEM_CANCEL_ONLINE:
3554 slab_mem_offline_callback(arg);
3555 break;
3556 case MEM_ONLINE:
3557 case MEM_CANCEL_OFFLINE:
3558 break;
3559 }
3560 if (ret)
3561 ret = notifier_from_errno(ret);
3562 else
3563 ret = NOTIFY_OK;
3564 return ret;
3565}
3566
3567static struct notifier_block slab_memory_callback_nb = {
3568 .notifier_call = slab_memory_callback,
3569 .priority = SLAB_CALLBACK_PRI,
3570};
3571
3572/********************************************************************
3573 * Basic setup of slabs
3574 *******************************************************************/
3575
3576/*
3577 * Used for early kmem_cache structures that were allocated using
3578 * the page allocator. Allocate them properly then fix up the pointers
3579 * that may be pointing to the wrong kmem_cache structure.
3580 */
3581
3582static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3583{
3584 int node;
3585 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3586
3587 memcpy(s, static_cache, kmem_cache->object_size);
3588
3589 /*
3590 * This runs very early, and only the boot processor is supposed to be
3591 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3592 * IPIs around.
3593 */
3594 __flush_cpu_slab(s, smp_processor_id());
3595 for_each_node_state(node, N_NORMAL_MEMORY) {
3596 struct kmem_cache_node *n = get_node(s, node);
3597 struct page *p;
3598
3599 if (n) {
3600 list_for_each_entry(p, &n->partial, lru)
3601 p->slab_cache = s;
3602
3603#ifdef CONFIG_SLUB_DEBUG
3604 list_for_each_entry(p, &n->full, lru)
3605 p->slab_cache = s;
3606#endif
3607 }
3608 }
3609 list_add(&s->list, &slab_caches);
3610 return s;
3611}
3612
3613void __init kmem_cache_init(void)
3614{
3615 static __initdata struct kmem_cache boot_kmem_cache,
3616 boot_kmem_cache_node;
3617
3618 if (debug_guardpage_minorder())
3619 slub_max_order = 0;
3620
3621 kmem_cache_node = &boot_kmem_cache_node;
3622 kmem_cache = &boot_kmem_cache;
3623
3624 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3625 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3626
3627 register_hotmemory_notifier(&slab_memory_callback_nb);
3628
3629 /* Able to allocate the per node structures */
3630 slab_state = PARTIAL;
3631
3632 create_boot_cache(kmem_cache, "kmem_cache",
3633 offsetof(struct kmem_cache, node) +
3634 nr_node_ids * sizeof(struct kmem_cache_node *),
3635 SLAB_HWCACHE_ALIGN);
3636
3637 kmem_cache = bootstrap(&boot_kmem_cache);
3638
3639 /*
3640 * Allocate kmem_cache_node properly from the kmem_cache slab.
3641 * kmem_cache_node is separately allocated so no need to
3642 * update any list pointers.
3643 */
3644 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3645
3646 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3647 create_kmalloc_caches(0);
3648
3649#ifdef CONFIG_SMP
3650 register_cpu_notifier(&slab_notifier);
3651#endif
3652
3653 printk(KERN_INFO
3654 "SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d,"
3655 " CPUs=%d, Nodes=%d\n",
3656 cache_line_size(),
3657 slub_min_order, slub_max_order, slub_min_objects,
3658 nr_cpu_ids, nr_node_ids);
3659}
3660
3661void __init kmem_cache_init_late(void)
3662{
3663}
3664
3665/*
3666 * Find a mergeable slab cache
3667 */
3668static int slab_unmergeable(struct kmem_cache *s)
3669{
3670 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3671 return 1;
3672
3673 if (!is_root_cache(s))
3674 return 1;
3675
3676 if (s->ctor)
3677 return 1;
3678
3679 /*
3680 * We may have set a slab to be unmergeable during bootstrap.
3681 */
3682 if (s->refcount < 0)
3683 return 1;
3684
3685 return 0;
3686}
3687
3688static struct kmem_cache *find_mergeable(size_t size, size_t align,
3689 unsigned long flags, const char *name, void (*ctor)(void *))
3690{
3691 struct kmem_cache *s;
3692
3693 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3694 return NULL;
3695
3696 if (ctor)
3697 return NULL;
3698
3699 size = ALIGN(size, sizeof(void *));
3700 align = calculate_alignment(flags, align, size);
3701 size = ALIGN(size, align);
3702 flags = kmem_cache_flags(size, flags, name, NULL);
3703
3704 list_for_each_entry(s, &slab_caches, list) {
3705 if (slab_unmergeable(s))
3706 continue;
3707
3708 if (size > s->size)
3709 continue;
3710
3711 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3712 continue;
3713 /*
3714 * Check if alignment is compatible.
3715 * Courtesy of Adrian Drzewiecki
3716 */
3717 if ((s->size & ~(align - 1)) != s->size)
3718 continue;
3719
3720 if (s->size - size >= sizeof(void *))
3721 continue;
3722
3723 return s;
3724 }
3725 return NULL;
3726}
3727
3728struct kmem_cache *
3729__kmem_cache_alias(const char *name, size_t size, size_t align,
3730 unsigned long flags, void (*ctor)(void *))
3731{
3732 struct kmem_cache *s;
3733
3734 s = find_mergeable(size, align, flags, name, ctor);
3735 if (s) {
3736 int i;
3737 struct kmem_cache *c;
3738
3739 s->refcount++;
3740
3741 /*
3742 * Adjust the object sizes so that we clear
3743 * the complete object on kzalloc.
3744 */
3745 s->object_size = max(s->object_size, (int)size);
3746 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3747
3748 for_each_memcg_cache_index(i) {
3749 c = cache_from_memcg_idx(s, i);
3750 if (!c)
3751 continue;
3752 c->object_size = s->object_size;
3753 c->inuse = max_t(int, c->inuse,
3754 ALIGN(size, sizeof(void *)));
3755 }
3756
3757 if (sysfs_slab_alias(s, name)) {
3758 s->refcount--;
3759 s = NULL;
3760 }
3761 }
3762
3763 return s;
3764}
3765
3766int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3767{
3768 int err;
3769
3770 err = kmem_cache_open(s, flags);
3771 if (err)
3772 return err;
3773
3774 /* Mutex is not taken during early boot */
3775 if (slab_state <= UP)
3776 return 0;
3777
3778 memcg_propagate_slab_attrs(s);
3779 err = sysfs_slab_add(s);
3780 if (err)
3781 kmem_cache_close(s);
3782
3783 return err;
3784}
3785
3786#ifdef CONFIG_SMP
3787/*
3788 * Use the cpu notifier to insure that the cpu slabs are flushed when
3789 * necessary.
3790 */
3791static int slab_cpuup_callback(struct notifier_block *nfb,
3792 unsigned long action, void *hcpu)
3793{
3794 long cpu = (long)hcpu;
3795 struct kmem_cache *s;
3796 unsigned long flags;
3797
3798 switch (action) {
3799 case CPU_UP_CANCELED:
3800 case CPU_UP_CANCELED_FROZEN:
3801 case CPU_DEAD:
3802 case CPU_DEAD_FROZEN:
3803 mutex_lock(&slab_mutex);
3804 list_for_each_entry(s, &slab_caches, list) {
3805 local_irq_save(flags);
3806 __flush_cpu_slab(s, cpu);
3807 local_irq_restore(flags);
3808 }
3809 mutex_unlock(&slab_mutex);
3810 break;
3811 default:
3812 break;
3813 }
3814 return NOTIFY_OK;
3815}
3816
3817static struct notifier_block slab_notifier = {
3818 .notifier_call = slab_cpuup_callback
3819};
3820
3821#endif
3822
3823void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3824{
3825 struct kmem_cache *s;
3826 void *ret;
3827
3828 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3829 return kmalloc_large(size, gfpflags);
3830
3831 s = kmalloc_slab(size, gfpflags);
3832
3833 if (unlikely(ZERO_OR_NULL_PTR(s)))
3834 return s;
3835
3836 ret = slab_alloc(s, gfpflags, caller);
3837
3838 /* Honor the call site pointer we received. */
3839 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3840
3841 return ret;
3842}
3843
3844#ifdef CONFIG_NUMA
3845void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3846 int node, unsigned long caller)
3847{
3848 struct kmem_cache *s;
3849 void *ret;
3850
3851 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3852 ret = kmalloc_large_node(size, gfpflags, node);
3853
3854 trace_kmalloc_node(caller, ret,
3855 size, PAGE_SIZE << get_order(size),
3856 gfpflags, node);
3857
3858 return ret;
3859 }
3860
3861 s = kmalloc_slab(size, gfpflags);
3862
3863 if (unlikely(ZERO_OR_NULL_PTR(s)))
3864 return s;
3865
3866 ret = slab_alloc_node(s, gfpflags, node, caller);
3867
3868 /* Honor the call site pointer we received. */
3869 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3870
3871 return ret;
3872}
3873#endif
3874
3875#ifdef CONFIG_SYSFS
3876static int count_inuse(struct page *page)
3877{
3878 return page->inuse;
3879}
3880
3881static int count_total(struct page *page)
3882{
3883 return page->objects;
3884}
3885#endif
3886
3887#ifdef CONFIG_SLUB_DEBUG
3888static int validate_slab(struct kmem_cache *s, struct page *page,
3889 unsigned long *map)
3890{
3891 void *p;
3892 void *addr = page_address(page);
3893
3894 if (!check_slab(s, page) ||
3895 !on_freelist(s, page, NULL))
3896 return 0;
3897
3898 /* Now we know that a valid freelist exists */
3899 bitmap_zero(map, page->objects);
3900
3901 get_map(s, page, map);
3902 for_each_object(p, s, addr, page->objects) {
3903 if (test_bit(slab_index(p, s, addr), map))
3904 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3905 return 0;
3906 }
3907
3908 for_each_object(p, s, addr, page->objects)
3909 if (!test_bit(slab_index(p, s, addr), map))
3910 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3911 return 0;
3912 return 1;
3913}
3914
3915static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3916 unsigned long *map)
3917{
3918 slab_lock(page);
3919 validate_slab(s, page, map);
3920 slab_unlock(page);
3921}
3922
3923static int validate_slab_node(struct kmem_cache *s,
3924 struct kmem_cache_node *n, unsigned long *map)
3925{
3926 unsigned long count = 0;
3927 struct page *page;
3928 unsigned long flags;
3929
3930 spin_lock_irqsave(&n->list_lock, flags);
3931
3932 list_for_each_entry(page, &n->partial, lru) {
3933 validate_slab_slab(s, page, map);
3934 count++;
3935 }
3936 if (count != n->nr_partial)
3937 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3938 "counter=%ld\n", s->name, count, n->nr_partial);
3939
3940 if (!(s->flags & SLAB_STORE_USER))
3941 goto out;
3942
3943 list_for_each_entry(page, &n->full, lru) {
3944 validate_slab_slab(s, page, map);
3945 count++;
3946 }
3947 if (count != atomic_long_read(&n->nr_slabs))
3948 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3949 "counter=%ld\n", s->name, count,
3950 atomic_long_read(&n->nr_slabs));
3951
3952out:
3953 spin_unlock_irqrestore(&n->list_lock, flags);
3954 return count;
3955}
3956
3957static long validate_slab_cache(struct kmem_cache *s)
3958{
3959 int node;
3960 unsigned long count = 0;
3961 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3962 sizeof(unsigned long), GFP_KERNEL);
3963
3964 if (!map)
3965 return -ENOMEM;
3966
3967 flush_all(s);
3968 for_each_node_state(node, N_NORMAL_MEMORY) {
3969 struct kmem_cache_node *n = get_node(s, node);
3970
3971 count += validate_slab_node(s, n, map);
3972 }
3973 kfree(map);
3974 return count;
3975}
3976/*
3977 * Generate lists of code addresses where slabcache objects are allocated
3978 * and freed.
3979 */
3980
3981struct location {
3982 unsigned long count;
3983 unsigned long addr;
3984 long long sum_time;
3985 long min_time;
3986 long max_time;
3987 long min_pid;
3988 long max_pid;
3989 DECLARE_BITMAP(cpus, NR_CPUS);
3990 nodemask_t nodes;
3991};
3992
3993struct loc_track {
3994 unsigned long max;
3995 unsigned long count;
3996 struct location *loc;
3997};
3998
3999static void free_loc_track(struct loc_track *t)
4000{
4001 if (t->max)
4002 free_pages((unsigned long)t->loc,
4003 get_order(sizeof(struct location) * t->max));
4004}
4005
4006static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4007{
4008 struct location *l;
4009 int order;
4010
4011 order = get_order(sizeof(struct location) * max);
4012
4013 l = (void *)__get_free_pages(flags, order);
4014 if (!l)
4015 return 0;
4016
4017 if (t->count) {
4018 memcpy(l, t->loc, sizeof(struct location) * t->count);
4019 free_loc_track(t);
4020 }
4021 t->max = max;
4022 t->loc = l;
4023 return 1;
4024}
4025
4026static int add_location(struct loc_track *t, struct kmem_cache *s,
4027 const struct track *track)
4028{
4029 long start, end, pos;
4030 struct location *l;
4031 unsigned long caddr;
4032 unsigned long age = jiffies - track->when;
4033
4034 start = -1;
4035 end = t->count;
4036
4037 for ( ; ; ) {
4038 pos = start + (end - start + 1) / 2;
4039
4040 /*
4041 * There is nothing at "end". If we end up there
4042 * we need to add something to before end.
4043 */
4044 if (pos == end)
4045 break;
4046
4047 caddr = t->loc[pos].addr;
4048 if (track->addr == caddr) {
4049
4050 l = &t->loc[pos];
4051 l->count++;
4052 if (track->when) {
4053 l->sum_time += age;
4054 if (age < l->min_time)
4055 l->min_time = age;
4056 if (age > l->max_time)
4057 l->max_time = age;
4058
4059 if (track->pid < l->min_pid)
4060 l->min_pid = track->pid;
4061 if (track->pid > l->max_pid)
4062 l->max_pid = track->pid;
4063
4064 cpumask_set_cpu(track->cpu,
4065 to_cpumask(l->cpus));
4066 }
4067 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4068 return 1;
4069 }
4070
4071 if (track->addr < caddr)
4072 end = pos;
4073 else
4074 start = pos;
4075 }
4076
4077 /*
4078 * Not found. Insert new tracking element.
4079 */
4080 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4081 return 0;
4082
4083 l = t->loc + pos;
4084 if (pos < t->count)
4085 memmove(l + 1, l,
4086 (t->count - pos) * sizeof(struct location));
4087 t->count++;
4088 l->count = 1;
4089 l->addr = track->addr;
4090 l->sum_time = age;
4091 l->min_time = age;
4092 l->max_time = age;
4093 l->min_pid = track->pid;
4094 l->max_pid = track->pid;
4095 cpumask_clear(to_cpumask(l->cpus));
4096 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4097 nodes_clear(l->nodes);
4098 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4099 return 1;
4100}
4101
4102static void process_slab(struct loc_track *t, struct kmem_cache *s,
4103 struct page *page, enum track_item alloc,
4104 unsigned long *map)
4105{
4106 void *addr = page_address(page);
4107 void *p;
4108
4109 bitmap_zero(map, page->objects);
4110 get_map(s, page, map);
4111
4112 for_each_object(p, s, addr, page->objects)
4113 if (!test_bit(slab_index(p, s, addr), map))
4114 add_location(t, s, get_track(s, p, alloc));
4115}
4116
4117static int list_locations(struct kmem_cache *s, char *buf,
4118 enum track_item alloc)
4119{
4120 int len = 0;
4121 unsigned long i;
4122 struct loc_track t = { 0, 0, NULL };
4123 int node;
4124 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4125 sizeof(unsigned long), GFP_KERNEL);
4126
4127 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4128 GFP_TEMPORARY)) {
4129 kfree(map);
4130 return sprintf(buf, "Out of memory\n");
4131 }
4132 /* Push back cpu slabs */
4133 flush_all(s);
4134
4135 for_each_node_state(node, N_NORMAL_MEMORY) {
4136 struct kmem_cache_node *n = get_node(s, node);
4137 unsigned long flags;
4138 struct page *page;
4139
4140 if (!atomic_long_read(&n->nr_slabs))
4141 continue;
4142
4143 spin_lock_irqsave(&n->list_lock, flags);
4144 list_for_each_entry(page, &n->partial, lru)
4145 process_slab(&t, s, page, alloc, map);
4146 list_for_each_entry(page, &n->full, lru)
4147 process_slab(&t, s, page, alloc, map);
4148 spin_unlock_irqrestore(&n->list_lock, flags);
4149 }
4150
4151 for (i = 0; i < t.count; i++) {
4152 struct location *l = &t.loc[i];
4153
4154 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4155 break;
4156 len += sprintf(buf + len, "%7ld ", l->count);
4157
4158 if (l->addr)
4159 len += sprintf(buf + len, "%pS", (void *)l->addr);
4160 else
4161 len += sprintf(buf + len, "<not-available>");
4162
4163 if (l->sum_time != l->min_time) {
4164 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4165 l->min_time,
4166 (long)div_u64(l->sum_time, l->count),
4167 l->max_time);
4168 } else
4169 len += sprintf(buf + len, " age=%ld",
4170 l->min_time);
4171
4172 if (l->min_pid != l->max_pid)
4173 len += sprintf(buf + len, " pid=%ld-%ld",
4174 l->min_pid, l->max_pid);
4175 else
4176 len += sprintf(buf + len, " pid=%ld",
4177 l->min_pid);
4178
4179 if (num_online_cpus() > 1 &&
4180 !cpumask_empty(to_cpumask(l->cpus)) &&
4181 len < PAGE_SIZE - 60) {
4182 len += sprintf(buf + len, " cpus=");
4183 len += cpulist_scnprintf(buf + len,
4184 PAGE_SIZE - len - 50,
4185 to_cpumask(l->cpus));
4186 }
4187
4188 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4189 len < PAGE_SIZE - 60) {
4190 len += sprintf(buf + len, " nodes=");
4191 len += nodelist_scnprintf(buf + len,
4192 PAGE_SIZE - len - 50,
4193 l->nodes);
4194 }
4195
4196 len += sprintf(buf + len, "\n");
4197 }
4198
4199 free_loc_track(&t);
4200 kfree(map);
4201 if (!t.count)
4202 len += sprintf(buf, "No data\n");
4203 return len;
4204}
4205#endif
4206
4207#ifdef SLUB_RESILIENCY_TEST
4208static void resiliency_test(void)
4209{
4210 u8 *p;
4211
4212 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4213
4214 printk(KERN_ERR "SLUB resiliency testing\n");
4215 printk(KERN_ERR "-----------------------\n");
4216 printk(KERN_ERR "A. Corruption after allocation\n");
4217
4218 p = kzalloc(16, GFP_KERNEL);
4219 p[16] = 0x12;
4220 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4221 " 0x12->0x%p\n\n", p + 16);
4222
4223 validate_slab_cache(kmalloc_caches[4]);
4224
4225 /* Hmmm... The next two are dangerous */
4226 p = kzalloc(32, GFP_KERNEL);
4227 p[32 + sizeof(void *)] = 0x34;
4228 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4229 " 0x34 -> -0x%p\n", p);
4230 printk(KERN_ERR
4231 "If allocated object is overwritten then not detectable\n\n");
4232
4233 validate_slab_cache(kmalloc_caches[5]);
4234 p = kzalloc(64, GFP_KERNEL);
4235 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4236 *p = 0x56;
4237 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4238 p);
4239 printk(KERN_ERR
4240 "If allocated object is overwritten then not detectable\n\n");
4241 validate_slab_cache(kmalloc_caches[6]);
4242
4243 printk(KERN_ERR "\nB. Corruption after free\n");
4244 p = kzalloc(128, GFP_KERNEL);
4245 kfree(p);
4246 *p = 0x78;
4247 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4248 validate_slab_cache(kmalloc_caches[7]);
4249
4250 p = kzalloc(256, GFP_KERNEL);
4251 kfree(p);
4252 p[50] = 0x9a;
4253 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4254 p);
4255 validate_slab_cache(kmalloc_caches[8]);
4256
4257 p = kzalloc(512, GFP_KERNEL);
4258 kfree(p);
4259 p[512] = 0xab;
4260 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4261 validate_slab_cache(kmalloc_caches[9]);
4262}
4263#else
4264#ifdef CONFIG_SYSFS
4265static void resiliency_test(void) {};
4266#endif
4267#endif
4268
4269#ifdef CONFIG_SYSFS
4270enum slab_stat_type {
4271 SL_ALL, /* All slabs */
4272 SL_PARTIAL, /* Only partially allocated slabs */
4273 SL_CPU, /* Only slabs used for cpu caches */
4274 SL_OBJECTS, /* Determine allocated objects not slabs */
4275 SL_TOTAL /* Determine object capacity not slabs */
4276};
4277
4278#define SO_ALL (1 << SL_ALL)
4279#define SO_PARTIAL (1 << SL_PARTIAL)
4280#define SO_CPU (1 << SL_CPU)
4281#define SO_OBJECTS (1 << SL_OBJECTS)
4282#define SO_TOTAL (1 << SL_TOTAL)
4283
4284static ssize_t show_slab_objects(struct kmem_cache *s,
4285 char *buf, unsigned long flags)
4286{
4287 unsigned long total = 0;
4288 int node;
4289 int x;
4290 unsigned long *nodes;
4291
4292 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4293 if (!nodes)
4294 return -ENOMEM;
4295
4296 if (flags & SO_CPU) {
4297 int cpu;
4298
4299 for_each_possible_cpu(cpu) {
4300 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4301 cpu);
4302 int node;
4303 struct page *page;
4304
4305 page = ACCESS_ONCE(c->page);
4306 if (!page)
4307 continue;
4308
4309 node = page_to_nid(page);
4310 if (flags & SO_TOTAL)
4311 x = page->objects;
4312 else if (flags & SO_OBJECTS)
4313 x = page->inuse;
4314 else
4315 x = 1;
4316
4317 total += x;
4318 nodes[node] += x;
4319
4320 page = ACCESS_ONCE(c->partial);
4321 if (page) {
4322 node = page_to_nid(page);
4323 if (flags & SO_TOTAL)
4324 WARN_ON_ONCE(1);
4325 else if (flags & SO_OBJECTS)
4326 WARN_ON_ONCE(1);
4327 else
4328 x = page->pages;
4329 total += x;
4330 nodes[node] += x;
4331 }
4332 }
4333 }
4334
4335 lock_memory_hotplug();
4336#ifdef CONFIG_SLUB_DEBUG
4337 if (flags & SO_ALL) {
4338 for_each_node_state(node, N_NORMAL_MEMORY) {
4339 struct kmem_cache_node *n = get_node(s, node);
4340
4341 if (flags & SO_TOTAL)
4342 x = atomic_long_read(&n->total_objects);
4343 else if (flags & SO_OBJECTS)
4344 x = atomic_long_read(&n->total_objects) -
4345 count_partial(n, count_free);
4346 else
4347 x = atomic_long_read(&n->nr_slabs);
4348 total += x;
4349 nodes[node] += x;
4350 }
4351
4352 } else
4353#endif
4354 if (flags & SO_PARTIAL) {
4355 for_each_node_state(node, N_NORMAL_MEMORY) {
4356 struct kmem_cache_node *n = get_node(s, node);
4357
4358 if (flags & SO_TOTAL)
4359 x = count_partial(n, count_total);
4360 else if (flags & SO_OBJECTS)
4361 x = count_partial(n, count_inuse);
4362 else
4363 x = n->nr_partial;
4364 total += x;
4365 nodes[node] += x;
4366 }
4367 }
4368 x = sprintf(buf, "%lu", total);
4369#ifdef CONFIG_NUMA
4370 for_each_node_state(node, N_NORMAL_MEMORY)
4371 if (nodes[node])
4372 x += sprintf(buf + x, " N%d=%lu",
4373 node, nodes[node]);
4374#endif
4375 unlock_memory_hotplug();
4376 kfree(nodes);
4377 return x + sprintf(buf + x, "\n");
4378}
4379
4380#ifdef CONFIG_SLUB_DEBUG
4381static int any_slab_objects(struct kmem_cache *s)
4382{
4383 int node;
4384
4385 for_each_online_node(node) {
4386 struct kmem_cache_node *n = get_node(s, node);
4387
4388 if (!n)
4389 continue;
4390
4391 if (atomic_long_read(&n->total_objects))
4392 return 1;
4393 }
4394 return 0;
4395}
4396#endif
4397
4398#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4399#define to_slab(n) container_of(n, struct kmem_cache, kobj)
4400
4401struct slab_attribute {
4402 struct attribute attr;
4403 ssize_t (*show)(struct kmem_cache *s, char *buf);
4404 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4405};
4406
4407#define SLAB_ATTR_RO(_name) \
4408 static struct slab_attribute _name##_attr = \
4409 __ATTR(_name, 0400, _name##_show, NULL)
4410
4411#define SLAB_ATTR(_name) \
4412 static struct slab_attribute _name##_attr = \
4413 __ATTR(_name, 0600, _name##_show, _name##_store)
4414
4415static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4416{
4417 return sprintf(buf, "%d\n", s->size);
4418}
4419SLAB_ATTR_RO(slab_size);
4420
4421static ssize_t align_show(struct kmem_cache *s, char *buf)
4422{
4423 return sprintf(buf, "%d\n", s->align);
4424}
4425SLAB_ATTR_RO(align);
4426
4427static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4428{
4429 return sprintf(buf, "%d\n", s->object_size);
4430}
4431SLAB_ATTR_RO(object_size);
4432
4433static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4434{
4435 return sprintf(buf, "%d\n", oo_objects(s->oo));
4436}
4437SLAB_ATTR_RO(objs_per_slab);
4438
4439static ssize_t order_store(struct kmem_cache *s,
4440 const char *buf, size_t length)
4441{
4442 unsigned long order;
4443 int err;
4444
4445 err = kstrtoul(buf, 10, &order);
4446 if (err)
4447 return err;
4448
4449 if (order > slub_max_order || order < slub_min_order)
4450 return -EINVAL;
4451
4452 calculate_sizes(s, order);
4453 return length;
4454}
4455
4456static ssize_t order_show(struct kmem_cache *s, char *buf)
4457{
4458 return sprintf(buf, "%d\n", oo_order(s->oo));
4459}
4460SLAB_ATTR(order);
4461
4462static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4463{
4464 return sprintf(buf, "%lu\n", s->min_partial);
4465}
4466
4467static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4468 size_t length)
4469{
4470 unsigned long min;
4471 int err;
4472
4473 err = kstrtoul(buf, 10, &min);
4474 if (err)
4475 return err;
4476
4477 set_min_partial(s, min);
4478 return length;
4479}
4480SLAB_ATTR(min_partial);
4481
4482static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4483{
4484 return sprintf(buf, "%u\n", s->cpu_partial);
4485}
4486
4487static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4488 size_t length)
4489{
4490 unsigned long objects;
4491 int err;
4492
4493 err = kstrtoul(buf, 10, &objects);
4494 if (err)
4495 return err;
4496 if (objects && !kmem_cache_has_cpu_partial(s))
4497 return -EINVAL;
4498
4499 s->cpu_partial = objects;
4500 flush_all(s);
4501 return length;
4502}
4503SLAB_ATTR(cpu_partial);
4504
4505static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4506{
4507 if (!s->ctor)
4508 return 0;
4509 return sprintf(buf, "%pS\n", s->ctor);
4510}
4511SLAB_ATTR_RO(ctor);
4512
4513static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4514{
4515 return sprintf(buf, "%d\n", s->refcount - 1);
4516}
4517SLAB_ATTR_RO(aliases);
4518
4519static ssize_t partial_show(struct kmem_cache *s, char *buf)
4520{
4521 return show_slab_objects(s, buf, SO_PARTIAL);
4522}
4523SLAB_ATTR_RO(partial);
4524
4525static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4526{
4527 return show_slab_objects(s, buf, SO_CPU);
4528}
4529SLAB_ATTR_RO(cpu_slabs);
4530
4531static ssize_t objects_show(struct kmem_cache *s, char *buf)
4532{
4533 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4534}
4535SLAB_ATTR_RO(objects);
4536
4537static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4538{
4539 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4540}
4541SLAB_ATTR_RO(objects_partial);
4542
4543static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4544{
4545 int objects = 0;
4546 int pages = 0;
4547 int cpu;
4548 int len;
4549
4550 for_each_online_cpu(cpu) {
4551 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4552
4553 if (page) {
4554 pages += page->pages;
4555 objects += page->pobjects;
4556 }
4557 }
4558
4559 len = sprintf(buf, "%d(%d)", objects, pages);
4560
4561#ifdef CONFIG_SMP
4562 for_each_online_cpu(cpu) {
4563 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4564
4565 if (page && len < PAGE_SIZE - 20)
4566 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4567 page->pobjects, page->pages);
4568 }
4569#endif
4570 return len + sprintf(buf + len, "\n");
4571}
4572SLAB_ATTR_RO(slabs_cpu_partial);
4573
4574static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4575{
4576 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4577}
4578
4579static ssize_t reclaim_account_store(struct kmem_cache *s,
4580 const char *buf, size_t length)
4581{
4582 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4583 if (buf[0] == '1')
4584 s->flags |= SLAB_RECLAIM_ACCOUNT;
4585 return length;
4586}
4587SLAB_ATTR(reclaim_account);
4588
4589static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4590{
4591 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4592}
4593SLAB_ATTR_RO(hwcache_align);
4594
4595#ifdef CONFIG_ZONE_DMA
4596static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4597{
4598 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4599}
4600SLAB_ATTR_RO(cache_dma);
4601#endif
4602
4603static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4604{
4605 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4606}
4607SLAB_ATTR_RO(destroy_by_rcu);
4608
4609static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4610{
4611 return sprintf(buf, "%d\n", s->reserved);
4612}
4613SLAB_ATTR_RO(reserved);
4614
4615#ifdef CONFIG_SLUB_DEBUG
4616static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4617{
4618 return show_slab_objects(s, buf, SO_ALL);
4619}
4620SLAB_ATTR_RO(slabs);
4621
4622static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4623{
4624 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4625}
4626SLAB_ATTR_RO(total_objects);
4627
4628static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4629{
4630 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4631}
4632
4633static ssize_t sanity_checks_store(struct kmem_cache *s,
4634 const char *buf, size_t length)
4635{
4636 s->flags &= ~SLAB_DEBUG_FREE;
4637 if (buf[0] == '1') {
4638 s->flags &= ~__CMPXCHG_DOUBLE;
4639 s->flags |= SLAB_DEBUG_FREE;
4640 }
4641 return length;
4642}
4643SLAB_ATTR(sanity_checks);
4644
4645static ssize_t trace_show(struct kmem_cache *s, char *buf)
4646{
4647 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4648}
4649
4650static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4651 size_t length)
4652{
4653 s->flags &= ~SLAB_TRACE;
4654 if (buf[0] == '1') {
4655 s->flags &= ~__CMPXCHG_DOUBLE;
4656 s->flags |= SLAB_TRACE;
4657 }
4658 return length;
4659}
4660SLAB_ATTR(trace);
4661
4662static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4663{
4664 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4665}
4666
4667static ssize_t red_zone_store(struct kmem_cache *s,
4668 const char *buf, size_t length)
4669{
4670 if (any_slab_objects(s))
4671 return -EBUSY;
4672
4673 s->flags &= ~SLAB_RED_ZONE;
4674 if (buf[0] == '1') {
4675 s->flags &= ~__CMPXCHG_DOUBLE;
4676 s->flags |= SLAB_RED_ZONE;
4677 }
4678 calculate_sizes(s, -1);
4679 return length;
4680}
4681SLAB_ATTR(red_zone);
4682
4683static ssize_t poison_show(struct kmem_cache *s, char *buf)
4684{
4685 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4686}
4687
4688static ssize_t poison_store(struct kmem_cache *s,
4689 const char *buf, size_t length)
4690{
4691 if (any_slab_objects(s))
4692 return -EBUSY;
4693
4694 s->flags &= ~SLAB_POISON;
4695 if (buf[0] == '1') {
4696 s->flags &= ~__CMPXCHG_DOUBLE;
4697 s->flags |= SLAB_POISON;
4698 }
4699 calculate_sizes(s, -1);
4700 return length;
4701}
4702SLAB_ATTR(poison);
4703
4704static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4705{
4706 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4707}
4708
4709static ssize_t store_user_store(struct kmem_cache *s,
4710 const char *buf, size_t length)
4711{
4712 if (any_slab_objects(s))
4713 return -EBUSY;
4714
4715 s->flags &= ~SLAB_STORE_USER;
4716 if (buf[0] == '1') {
4717 s->flags &= ~__CMPXCHG_DOUBLE;
4718 s->flags |= SLAB_STORE_USER;
4719 }
4720 calculate_sizes(s, -1);
4721 return length;
4722}
4723SLAB_ATTR(store_user);
4724
4725static ssize_t validate_show(struct kmem_cache *s, char *buf)
4726{
4727 return 0;
4728}
4729
4730static ssize_t validate_store(struct kmem_cache *s,
4731 const char *buf, size_t length)
4732{
4733 int ret = -EINVAL;
4734
4735 if (buf[0] == '1') {
4736 ret = validate_slab_cache(s);
4737 if (ret >= 0)
4738 ret = length;
4739 }
4740 return ret;
4741}
4742SLAB_ATTR(validate);
4743
4744static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4745{
4746 if (!(s->flags & SLAB_STORE_USER))
4747 return -ENOSYS;
4748 return list_locations(s, buf, TRACK_ALLOC);
4749}
4750SLAB_ATTR_RO(alloc_calls);
4751
4752static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4753{
4754 if (!(s->flags & SLAB_STORE_USER))
4755 return -ENOSYS;
4756 return list_locations(s, buf, TRACK_FREE);
4757}
4758SLAB_ATTR_RO(free_calls);
4759#endif /* CONFIG_SLUB_DEBUG */
4760
4761#ifdef CONFIG_FAILSLAB
4762static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4763{
4764 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4765}
4766
4767static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4768 size_t length)
4769{
4770 s->flags &= ~SLAB_FAILSLAB;
4771 if (buf[0] == '1')
4772 s->flags |= SLAB_FAILSLAB;
4773 return length;
4774}
4775SLAB_ATTR(failslab);
4776#endif
4777
4778static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4779{
4780 return 0;
4781}
4782
4783static ssize_t shrink_store(struct kmem_cache *s,
4784 const char *buf, size_t length)
4785{
4786 if (buf[0] == '1') {
4787 int rc = kmem_cache_shrink(s);
4788
4789 if (rc)
4790 return rc;
4791 } else
4792 return -EINVAL;
4793 return length;
4794}
4795SLAB_ATTR(shrink);
4796
4797#ifdef CONFIG_NUMA
4798static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4799{
4800 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4801}
4802
4803static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4804 const char *buf, size_t length)
4805{
4806 unsigned long ratio;
4807 int err;
4808
4809 err = kstrtoul(buf, 10, &ratio);
4810 if (err)
4811 return err;
4812
4813 if (ratio <= 100)
4814 s->remote_node_defrag_ratio = ratio * 10;
4815
4816 return length;
4817}
4818SLAB_ATTR(remote_node_defrag_ratio);
4819#endif
4820
4821#ifdef CONFIG_SLUB_STATS
4822static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4823{
4824 unsigned long sum = 0;
4825 int cpu;
4826 int len;
4827 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4828
4829 if (!data)
4830 return -ENOMEM;
4831
4832 for_each_online_cpu(cpu) {
4833 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4834
4835 data[cpu] = x;
4836 sum += x;
4837 }
4838
4839 len = sprintf(buf, "%lu", sum);
4840
4841#ifdef CONFIG_SMP
4842 for_each_online_cpu(cpu) {
4843 if (data[cpu] && len < PAGE_SIZE - 20)
4844 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4845 }
4846#endif
4847 kfree(data);
4848 return len + sprintf(buf + len, "\n");
4849}
4850
4851static void clear_stat(struct kmem_cache *s, enum stat_item si)
4852{
4853 int cpu;
4854
4855 for_each_online_cpu(cpu)
4856 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4857}
4858
4859#define STAT_ATTR(si, text) \
4860static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4861{ \
4862 return show_stat(s, buf, si); \
4863} \
4864static ssize_t text##_store(struct kmem_cache *s, \
4865 const char *buf, size_t length) \
4866{ \
4867 if (buf[0] != '0') \
4868 return -EINVAL; \
4869 clear_stat(s, si); \
4870 return length; \
4871} \
4872SLAB_ATTR(text); \
4873
4874STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4875STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4876STAT_ATTR(FREE_FASTPATH, free_fastpath);
4877STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4878STAT_ATTR(FREE_FROZEN, free_frozen);
4879STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4880STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4881STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4882STAT_ATTR(ALLOC_SLAB, alloc_slab);
4883STAT_ATTR(ALLOC_REFILL, alloc_refill);
4884STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
4885STAT_ATTR(FREE_SLAB, free_slab);
4886STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4887STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4888STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4889STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4890STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4891STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4892STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
4893STAT_ATTR(ORDER_FALLBACK, order_fallback);
4894STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4895STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4896STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
4897STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
4898STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
4899STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
4900#endif
4901
4902static struct attribute *slab_attrs[] = {
4903 &slab_size_attr.attr,
4904 &object_size_attr.attr,
4905 &objs_per_slab_attr.attr,
4906 &order_attr.attr,
4907 &min_partial_attr.attr,
4908 &cpu_partial_attr.attr,
4909 &objects_attr.attr,
4910 &objects_partial_attr.attr,
4911 &partial_attr.attr,
4912 &cpu_slabs_attr.attr,
4913 &ctor_attr.attr,
4914 &aliases_attr.attr,
4915 &align_attr.attr,
4916 &hwcache_align_attr.attr,
4917 &reclaim_account_attr.attr,
4918 &destroy_by_rcu_attr.attr,
4919 &shrink_attr.attr,
4920 &reserved_attr.attr,
4921 &slabs_cpu_partial_attr.attr,
4922#ifdef CONFIG_SLUB_DEBUG
4923 &total_objects_attr.attr,
4924 &slabs_attr.attr,
4925 &sanity_checks_attr.attr,
4926 &trace_attr.attr,
4927 &red_zone_attr.attr,
4928 &poison_attr.attr,
4929 &store_user_attr.attr,
4930 &validate_attr.attr,
4931 &alloc_calls_attr.attr,
4932 &free_calls_attr.attr,
4933#endif
4934#ifdef CONFIG_ZONE_DMA
4935 &cache_dma_attr.attr,
4936#endif
4937#ifdef CONFIG_NUMA
4938 &remote_node_defrag_ratio_attr.attr,
4939#endif
4940#ifdef CONFIG_SLUB_STATS
4941 &alloc_fastpath_attr.attr,
4942 &alloc_slowpath_attr.attr,
4943 &free_fastpath_attr.attr,
4944 &free_slowpath_attr.attr,
4945 &free_frozen_attr.attr,
4946 &free_add_partial_attr.attr,
4947 &free_remove_partial_attr.attr,
4948 &alloc_from_partial_attr.attr,
4949 &alloc_slab_attr.attr,
4950 &alloc_refill_attr.attr,
4951 &alloc_node_mismatch_attr.attr,
4952 &free_slab_attr.attr,
4953 &cpuslab_flush_attr.attr,
4954 &deactivate_full_attr.attr,
4955 &deactivate_empty_attr.attr,
4956 &deactivate_to_head_attr.attr,
4957 &deactivate_to_tail_attr.attr,
4958 &deactivate_remote_frees_attr.attr,
4959 &deactivate_bypass_attr.attr,
4960 &order_fallback_attr.attr,
4961 &cmpxchg_double_fail_attr.attr,
4962 &cmpxchg_double_cpu_fail_attr.attr,
4963 &cpu_partial_alloc_attr.attr,
4964 &cpu_partial_free_attr.attr,
4965 &cpu_partial_node_attr.attr,
4966 &cpu_partial_drain_attr.attr,
4967#endif
4968#ifdef CONFIG_FAILSLAB
4969 &failslab_attr.attr,
4970#endif
4971
4972 NULL
4973};
4974
4975static struct attribute_group slab_attr_group = {
4976 .attrs = slab_attrs,
4977};
4978
4979static ssize_t slab_attr_show(struct kobject *kobj,
4980 struct attribute *attr,
4981 char *buf)
4982{
4983 struct slab_attribute *attribute;
4984 struct kmem_cache *s;
4985 int err;
4986
4987 attribute = to_slab_attr(attr);
4988 s = to_slab(kobj);
4989
4990 if (!attribute->show)
4991 return -EIO;
4992
4993 err = attribute->show(s, buf);
4994
4995 return err;
4996}
4997
4998static ssize_t slab_attr_store(struct kobject *kobj,
4999 struct attribute *attr,
5000 const char *buf, size_t len)
5001{
5002 struct slab_attribute *attribute;
5003 struct kmem_cache *s;
5004 int err;
5005
5006 attribute = to_slab_attr(attr);
5007 s = to_slab(kobj);
5008
5009 if (!attribute->store)
5010 return -EIO;
5011
5012 err = attribute->store(s, buf, len);
5013#ifdef CONFIG_MEMCG_KMEM
5014 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5015 int i;
5016
5017 mutex_lock(&slab_mutex);
5018 if (s->max_attr_size < len)
5019 s->max_attr_size = len;
5020
5021 /*
5022 * This is a best effort propagation, so this function's return
5023 * value will be determined by the parent cache only. This is
5024 * basically because not all attributes will have a well
5025 * defined semantics for rollbacks - most of the actions will
5026 * have permanent effects.
5027 *
5028 * Returning the error value of any of the children that fail
5029 * is not 100 % defined, in the sense that users seeing the
5030 * error code won't be able to know anything about the state of
5031 * the cache.
5032 *
5033 * Only returning the error code for the parent cache at least
5034 * has well defined semantics. The cache being written to
5035 * directly either failed or succeeded, in which case we loop
5036 * through the descendants with best-effort propagation.
5037 */
5038 for_each_memcg_cache_index(i) {
5039 struct kmem_cache *c = cache_from_memcg_idx(s, i);
5040 if (c)
5041 attribute->store(c, buf, len);
5042 }
5043 mutex_unlock(&slab_mutex);
5044 }
5045#endif
5046 return err;
5047}
5048
5049static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5050{
5051#ifdef CONFIG_MEMCG_KMEM
5052 int i;
5053 char *buffer = NULL;
5054 struct kmem_cache *root_cache;
5055
5056 if (is_root_cache(s))
5057 return;
5058
5059 root_cache = s->memcg_params->root_cache;
5060
5061 /*
5062 * This mean this cache had no attribute written. Therefore, no point
5063 * in copying default values around
5064 */
5065 if (!root_cache->max_attr_size)
5066 return;
5067
5068 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5069 char mbuf[64];
5070 char *buf;
5071 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5072
5073 if (!attr || !attr->store || !attr->show)
5074 continue;
5075
5076 /*
5077 * It is really bad that we have to allocate here, so we will
5078 * do it only as a fallback. If we actually allocate, though,
5079 * we can just use the allocated buffer until the end.
5080 *
5081 * Most of the slub attributes will tend to be very small in
5082 * size, but sysfs allows buffers up to a page, so they can
5083 * theoretically happen.
5084 */
5085 if (buffer)
5086 buf = buffer;
5087 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5088 buf = mbuf;
5089 else {
5090 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5091 if (WARN_ON(!buffer))
5092 continue;
5093 buf = buffer;
5094 }
5095
5096 attr->show(root_cache, buf);
5097 attr->store(s, buf, strlen(buf));
5098 }
5099
5100 if (buffer)
5101 free_page((unsigned long)buffer);
5102#endif
5103}
5104
5105static void kmem_cache_release(struct kobject *k)
5106{
5107 slab_kmem_cache_release(to_slab(k));
5108}
5109
5110static const struct sysfs_ops slab_sysfs_ops = {
5111 .show = slab_attr_show,
5112 .store = slab_attr_store,
5113};
5114
5115static struct kobj_type slab_ktype = {
5116 .sysfs_ops = &slab_sysfs_ops,
5117 .release = kmem_cache_release,
5118};
5119
5120static int uevent_filter(struct kset *kset, struct kobject *kobj)
5121{
5122 struct kobj_type *ktype = get_ktype(kobj);
5123
5124 if (ktype == &slab_ktype)
5125 return 1;
5126 return 0;
5127}
5128
5129static const struct kset_uevent_ops slab_uevent_ops = {
5130 .filter = uevent_filter,
5131};
5132
5133static struct kset *slab_kset;
5134
5135static inline struct kset *cache_kset(struct kmem_cache *s)
5136{
5137#ifdef CONFIG_MEMCG_KMEM
5138 if (!is_root_cache(s))
5139 return s->memcg_params->root_cache->memcg_kset;
5140#endif
5141 return slab_kset;
5142}
5143
5144#define ID_STR_LENGTH 64
5145
5146/* Create a unique string id for a slab cache:
5147 *
5148 * Format :[flags-]size
5149 */
5150static char *create_unique_id(struct kmem_cache *s)
5151{
5152 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5153 char *p = name;
5154
5155 BUG_ON(!name);
5156
5157 *p++ = ':';
5158 /*
5159 * First flags affecting slabcache operations. We will only
5160 * get here for aliasable slabs so we do not need to support
5161 * too many flags. The flags here must cover all flags that
5162 * are matched during merging to guarantee that the id is
5163 * unique.
5164 */
5165 if (s->flags & SLAB_CACHE_DMA)
5166 *p++ = 'd';
5167 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5168 *p++ = 'a';
5169 if (s->flags & SLAB_DEBUG_FREE)
5170 *p++ = 'F';
5171 if (!(s->flags & SLAB_NOTRACK))
5172 *p++ = 't';
5173 if (p != name + 1)
5174 *p++ = '-';
5175 p += sprintf(p, "%07d", s->size);
5176
5177#ifdef CONFIG_MEMCG_KMEM
5178 if (!is_root_cache(s))
5179 p += sprintf(p, "-%08d",
5180 memcg_cache_id(s->memcg_params->memcg));
5181#endif
5182
5183 BUG_ON(p > name + ID_STR_LENGTH - 1);
5184 return name;
5185}
5186
5187static int sysfs_slab_add(struct kmem_cache *s)
5188{
5189 int err;
5190 const char *name;
5191 int unmergeable = slab_unmergeable(s);
5192
5193 if (unmergeable) {
5194 /*
5195 * Slabcache can never be merged so we can use the name proper.
5196 * This is typically the case for debug situations. In that
5197 * case we can catch duplicate names easily.
5198 */
5199 sysfs_remove_link(&slab_kset->kobj, s->name);
5200 name = s->name;
5201 } else {
5202 /*
5203 * Create a unique name for the slab as a target
5204 * for the symlinks.
5205 */
5206 name = create_unique_id(s);
5207 }
5208
5209 s->kobj.kset = cache_kset(s);
5210 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5211 if (err)
5212 goto out_put_kobj;
5213
5214 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5215 if (err)
5216 goto out_del_kobj;
5217
5218#ifdef CONFIG_MEMCG_KMEM
5219 if (is_root_cache(s)) {
5220 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5221 if (!s->memcg_kset) {
5222 err = -ENOMEM;
5223 goto out_del_kobj;
5224 }
5225 }
5226#endif
5227
5228 kobject_uevent(&s->kobj, KOBJ_ADD);
5229 if (!unmergeable) {
5230 /* Setup first alias */
5231 sysfs_slab_alias(s, s->name);
5232 }
5233out:
5234 if (!unmergeable)
5235 kfree(name);
5236 return err;
5237out_del_kobj:
5238 kobject_del(&s->kobj);
5239out_put_kobj:
5240 kobject_put(&s->kobj);
5241 goto out;
5242}
5243
5244void sysfs_slab_remove(struct kmem_cache *s)
5245{
5246 if (slab_state < FULL)
5247 /*
5248 * Sysfs has not been setup yet so no need to remove the
5249 * cache from sysfs.
5250 */
5251 return;
5252
5253#ifdef CONFIG_MEMCG_KMEM
5254 kset_unregister(s->memcg_kset);
5255#endif
5256 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5257 kobject_del(&s->kobj);
5258 kobject_put(&s->kobj);
5259}
5260
5261/*
5262 * Need to buffer aliases during bootup until sysfs becomes
5263 * available lest we lose that information.
5264 */
5265struct saved_alias {
5266 struct kmem_cache *s;
5267 const char *name;
5268 struct saved_alias *next;
5269};
5270
5271static struct saved_alias *alias_list;
5272
5273static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5274{
5275 struct saved_alias *al;
5276
5277 if (slab_state == FULL) {
5278 /*
5279 * If we have a leftover link then remove it.
5280 */
5281 sysfs_remove_link(&slab_kset->kobj, name);
5282 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5283 }
5284
5285 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5286 if (!al)
5287 return -ENOMEM;
5288
5289 al->s = s;
5290 al->name = name;
5291 al->next = alias_list;
5292 alias_list = al;
5293 return 0;
5294}
5295
5296static int __init slab_sysfs_init(void)
5297{
5298 struct kmem_cache *s;
5299 int err;
5300
5301 mutex_lock(&slab_mutex);
5302
5303 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5304 if (!slab_kset) {
5305 mutex_unlock(&slab_mutex);
5306 printk(KERN_ERR "Cannot register slab subsystem.\n");
5307 return -ENOSYS;
5308 }
5309
5310 slab_state = FULL;
5311
5312 list_for_each_entry(s, &slab_caches, list) {
5313 err = sysfs_slab_add(s);
5314 if (err)
5315 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5316 " to sysfs\n", s->name);
5317 }
5318
5319 while (alias_list) {
5320 struct saved_alias *al = alias_list;
5321
5322 alias_list = alias_list->next;
5323 err = sysfs_slab_alias(al->s, al->name);
5324 if (err)
5325 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5326 " %s to sysfs\n", al->name);
5327 kfree(al);
5328 }
5329
5330 mutex_unlock(&slab_mutex);
5331 resiliency_test();
5332 return 0;
5333}
5334
5335__initcall(slab_sysfs_init);
5336#endif /* CONFIG_SYSFS */
5337
5338/*
5339 * The /proc/slabinfo ABI
5340 */
5341#ifdef CONFIG_SLABINFO
5342void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5343{
5344 unsigned long nr_slabs = 0;
5345 unsigned long nr_objs = 0;
5346 unsigned long nr_free = 0;
5347 int node;
5348
5349 for_each_online_node(node) {
5350 struct kmem_cache_node *n = get_node(s, node);
5351
5352 if (!n)
5353 continue;
5354
5355 nr_slabs += node_nr_slabs(n);
5356 nr_objs += node_nr_objs(n);
5357 nr_free += count_partial(n, count_free);
5358 }
5359
5360 sinfo->active_objs = nr_objs - nr_free;
5361 sinfo->num_objs = nr_objs;
5362 sinfo->active_slabs = nr_slabs;
5363 sinfo->num_slabs = nr_slabs;
5364 sinfo->objects_per_slab = oo_objects(s->oo);
5365 sinfo->cache_order = oo_order(s->oo);
5366}
5367
5368void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5369{
5370}
5371
5372ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5373 size_t count, loff_t *ppos)
5374{
5375 return -EIO;
5376}
5377#endif /* CONFIG_SLABINFO */