]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blob - mm/slub.c
Merge branch 'drm-next-4.7' of git://people.freedesktop.org/~agd5f/linux into drm...
[mirror_ubuntu-artful-kernel.git] / mm / slub.c
1 /*
2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
4 *
5 * The allocator synchronizes using per slab locks or atomic operatios
6 * and only uses a centralized lock to manage a pool of partial slabs.
7 *
8 * (C) 2007 SGI, Christoph Lameter
9 * (C) 2011 Linux Foundation, Christoph Lameter
10 */
11
12 #include <linux/mm.h>
13 #include <linux/swap.h> /* struct reclaim_state */
14 #include <linux/module.h>
15 #include <linux/bit_spinlock.h>
16 #include <linux/interrupt.h>
17 #include <linux/bitops.h>
18 #include <linux/slab.h>
19 #include "slab.h"
20 #include <linux/proc_fs.h>
21 #include <linux/notifier.h>
22 #include <linux/seq_file.h>
23 #include <linux/kasan.h>
24 #include <linux/kmemcheck.h>
25 #include <linux/cpu.h>
26 #include <linux/cpuset.h>
27 #include <linux/mempolicy.h>
28 #include <linux/ctype.h>
29 #include <linux/debugobjects.h>
30 #include <linux/kallsyms.h>
31 #include <linux/memory.h>
32 #include <linux/math64.h>
33 #include <linux/fault-inject.h>
34 #include <linux/stacktrace.h>
35 #include <linux/prefetch.h>
36 #include <linux/memcontrol.h>
37
38 #include <trace/events/kmem.h>
39
40 #include "internal.h"
41
42 /*
43 * Lock order:
44 * 1. slab_mutex (Global Mutex)
45 * 2. node->list_lock
46 * 3. slab_lock(page) (Only on some arches and for debugging)
47 *
48 * slab_mutex
49 *
50 * The role of the slab_mutex is to protect the list of all the slabs
51 * and to synchronize major metadata changes to slab cache structures.
52 *
53 * The slab_lock is only used for debugging and on arches that do not
54 * have the ability to do a cmpxchg_double. It only protects the second
55 * double word in the page struct. Meaning
56 * A. page->freelist -> List of object free in a page
57 * B. page->counters -> Counters of objects
58 * C. page->frozen -> frozen state
59 *
60 * If a slab is frozen then it is exempt from list management. It is not
61 * on any list. The processor that froze the slab is the one who can
62 * perform list operations on the page. Other processors may put objects
63 * onto the freelist but the processor that froze the slab is the only
64 * one that can retrieve the objects from the page's freelist.
65 *
66 * The list_lock protects the partial and full list on each node and
67 * the partial slab counter. If taken then no new slabs may be added or
68 * removed from the lists nor make the number of partial slabs be modified.
69 * (Note that the total number of slabs is an atomic value that may be
70 * modified without taking the list lock).
71 *
72 * The list_lock is a centralized lock and thus we avoid taking it as
73 * much as possible. As long as SLUB does not have to handle partial
74 * slabs, operations can continue without any centralized lock. F.e.
75 * allocating a long series of objects that fill up slabs does not require
76 * the list lock.
77 * Interrupts are disabled during allocation and deallocation in order to
78 * make the slab allocator safe to use in the context of an irq. In addition
79 * interrupts are disabled to ensure that the processor does not change
80 * while handling per_cpu slabs, due to kernel preemption.
81 *
82 * SLUB assigns one slab for allocation to each processor.
83 * Allocations only occur from these slabs called cpu slabs.
84 *
85 * Slabs with free elements are kept on a partial list and during regular
86 * operations no list for full slabs is used. If an object in a full slab is
87 * freed then the slab will show up again on the partial lists.
88 * We track full slabs for debugging purposes though because otherwise we
89 * cannot scan all objects.
90 *
91 * Slabs are freed when they become empty. Teardown and setup is
92 * minimal so we rely on the page allocators per cpu caches for
93 * fast frees and allocs.
94 *
95 * Overloading of page flags that are otherwise used for LRU management.
96 *
97 * PageActive The slab is frozen and exempt from list processing.
98 * This means that the slab is dedicated to a purpose
99 * such as satisfying allocations for a specific
100 * processor. Objects may be freed in the slab while
101 * it is frozen but slab_free will then skip the usual
102 * list operations. It is up to the processor holding
103 * the slab to integrate the slab into the slab lists
104 * when the slab is no longer needed.
105 *
106 * One use of this flag is to mark slabs that are
107 * used for allocations. Then such a slab becomes a cpu
108 * slab. The cpu slab may be equipped with an additional
109 * freelist that allows lockless access to
110 * free objects in addition to the regular freelist
111 * that requires the slab lock.
112 *
113 * PageError Slab requires special handling due to debug
114 * options set. This moves slab handling out of
115 * the fast path and disables lockless freelists.
116 */
117
118 static inline int kmem_cache_debug(struct kmem_cache *s)
119 {
120 #ifdef CONFIG_SLUB_DEBUG
121 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
122 #else
123 return 0;
124 #endif
125 }
126
127 static inline void *fixup_red_left(struct kmem_cache *s, void *p)
128 {
129 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE)
130 p += s->red_left_pad;
131
132 return p;
133 }
134
135 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
136 {
137 #ifdef CONFIG_SLUB_CPU_PARTIAL
138 return !kmem_cache_debug(s);
139 #else
140 return false;
141 #endif
142 }
143
144 /*
145 * Issues still to be resolved:
146 *
147 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
148 *
149 * - Variable sizing of the per node arrays
150 */
151
152 /* Enable to test recovery from slab corruption on boot */
153 #undef SLUB_RESILIENCY_TEST
154
155 /* Enable to log cmpxchg failures */
156 #undef SLUB_DEBUG_CMPXCHG
157
158 /*
159 * Mininum number of partial slabs. These will be left on the partial
160 * lists even if they are empty. kmem_cache_shrink may reclaim them.
161 */
162 #define MIN_PARTIAL 5
163
164 /*
165 * Maximum number of desirable partial slabs.
166 * The existence of more partial slabs makes kmem_cache_shrink
167 * sort the partial list by the number of objects in use.
168 */
169 #define MAX_PARTIAL 10
170
171 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
172 SLAB_POISON | SLAB_STORE_USER)
173
174 /*
175 * These debug flags cannot use CMPXCHG because there might be consistency
176 * issues when checking or reading debug information
177 */
178 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
179 SLAB_TRACE)
180
181
182 /*
183 * Debugging flags that require metadata to be stored in the slab. These get
184 * disabled when slub_debug=O is used and a cache's min order increases with
185 * metadata.
186 */
187 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
188
189 #define OO_SHIFT 16
190 #define OO_MASK ((1 << OO_SHIFT) - 1)
191 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
192
193 /* Internal SLUB flags */
194 #define __OBJECT_POISON 0x80000000UL /* Poison object */
195 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
196
197 #ifdef CONFIG_SMP
198 static struct notifier_block slab_notifier;
199 #endif
200
201 /*
202 * Tracking user of a slab.
203 */
204 #define TRACK_ADDRS_COUNT 16
205 struct track {
206 unsigned long addr; /* Called from address */
207 #ifdef CONFIG_STACKTRACE
208 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
209 #endif
210 int cpu; /* Was running on cpu */
211 int pid; /* Pid context */
212 unsigned long when; /* When did the operation occur */
213 };
214
215 enum track_item { TRACK_ALLOC, TRACK_FREE };
216
217 #ifdef CONFIG_SYSFS
218 static int sysfs_slab_add(struct kmem_cache *);
219 static int sysfs_slab_alias(struct kmem_cache *, const char *);
220 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
221 #else
222 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
223 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
224 { return 0; }
225 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
226 #endif
227
228 static inline void stat(const struct kmem_cache *s, enum stat_item si)
229 {
230 #ifdef CONFIG_SLUB_STATS
231 /*
232 * The rmw is racy on a preemptible kernel but this is acceptable, so
233 * avoid this_cpu_add()'s irq-disable overhead.
234 */
235 raw_cpu_inc(s->cpu_slab->stat[si]);
236 #endif
237 }
238
239 /********************************************************************
240 * Core slab cache functions
241 *******************************************************************/
242
243 static inline void *get_freepointer(struct kmem_cache *s, void *object)
244 {
245 return *(void **)(object + s->offset);
246 }
247
248 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
249 {
250 prefetch(object + s->offset);
251 }
252
253 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
254 {
255 void *p;
256
257 if (!debug_pagealloc_enabled())
258 return get_freepointer(s, object);
259
260 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
261 return p;
262 }
263
264 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
265 {
266 *(void **)(object + s->offset) = fp;
267 }
268
269 /* Loop over all objects in a slab */
270 #define for_each_object(__p, __s, __addr, __objects) \
271 for (__p = fixup_red_left(__s, __addr); \
272 __p < (__addr) + (__objects) * (__s)->size; \
273 __p += (__s)->size)
274
275 #define for_each_object_idx(__p, __idx, __s, __addr, __objects) \
276 for (__p = fixup_red_left(__s, __addr), __idx = 1; \
277 __idx <= __objects; \
278 __p += (__s)->size, __idx++)
279
280 /* Determine object index from a given position */
281 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
282 {
283 return (p - addr) / s->size;
284 }
285
286 static inline int order_objects(int order, unsigned long size, int reserved)
287 {
288 return ((PAGE_SIZE << order) - reserved) / size;
289 }
290
291 static inline struct kmem_cache_order_objects oo_make(int order,
292 unsigned long size, int reserved)
293 {
294 struct kmem_cache_order_objects x = {
295 (order << OO_SHIFT) + order_objects(order, size, reserved)
296 };
297
298 return x;
299 }
300
301 static inline int oo_order(struct kmem_cache_order_objects x)
302 {
303 return x.x >> OO_SHIFT;
304 }
305
306 static inline int oo_objects(struct kmem_cache_order_objects x)
307 {
308 return x.x & OO_MASK;
309 }
310
311 /*
312 * Per slab locking using the pagelock
313 */
314 static __always_inline void slab_lock(struct page *page)
315 {
316 VM_BUG_ON_PAGE(PageTail(page), page);
317 bit_spin_lock(PG_locked, &page->flags);
318 }
319
320 static __always_inline void slab_unlock(struct page *page)
321 {
322 VM_BUG_ON_PAGE(PageTail(page), page);
323 __bit_spin_unlock(PG_locked, &page->flags);
324 }
325
326 static inline void set_page_slub_counters(struct page *page, unsigned long counters_new)
327 {
328 struct page tmp;
329 tmp.counters = counters_new;
330 /*
331 * page->counters can cover frozen/inuse/objects as well
332 * as page->_refcount. If we assign to ->counters directly
333 * we run the risk of losing updates to page->_refcount, so
334 * be careful and only assign to the fields we need.
335 */
336 page->frozen = tmp.frozen;
337 page->inuse = tmp.inuse;
338 page->objects = tmp.objects;
339 }
340
341 /* Interrupts must be disabled (for the fallback code to work right) */
342 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
343 void *freelist_old, unsigned long counters_old,
344 void *freelist_new, unsigned long counters_new,
345 const char *n)
346 {
347 VM_BUG_ON(!irqs_disabled());
348 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
349 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
350 if (s->flags & __CMPXCHG_DOUBLE) {
351 if (cmpxchg_double(&page->freelist, &page->counters,
352 freelist_old, counters_old,
353 freelist_new, counters_new))
354 return true;
355 } else
356 #endif
357 {
358 slab_lock(page);
359 if (page->freelist == freelist_old &&
360 page->counters == counters_old) {
361 page->freelist = freelist_new;
362 set_page_slub_counters(page, counters_new);
363 slab_unlock(page);
364 return true;
365 }
366 slab_unlock(page);
367 }
368
369 cpu_relax();
370 stat(s, CMPXCHG_DOUBLE_FAIL);
371
372 #ifdef SLUB_DEBUG_CMPXCHG
373 pr_info("%s %s: cmpxchg double redo ", n, s->name);
374 #endif
375
376 return false;
377 }
378
379 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
380 void *freelist_old, unsigned long counters_old,
381 void *freelist_new, unsigned long counters_new,
382 const char *n)
383 {
384 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
385 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
386 if (s->flags & __CMPXCHG_DOUBLE) {
387 if (cmpxchg_double(&page->freelist, &page->counters,
388 freelist_old, counters_old,
389 freelist_new, counters_new))
390 return true;
391 } else
392 #endif
393 {
394 unsigned long flags;
395
396 local_irq_save(flags);
397 slab_lock(page);
398 if (page->freelist == freelist_old &&
399 page->counters == counters_old) {
400 page->freelist = freelist_new;
401 set_page_slub_counters(page, counters_new);
402 slab_unlock(page);
403 local_irq_restore(flags);
404 return true;
405 }
406 slab_unlock(page);
407 local_irq_restore(flags);
408 }
409
410 cpu_relax();
411 stat(s, CMPXCHG_DOUBLE_FAIL);
412
413 #ifdef SLUB_DEBUG_CMPXCHG
414 pr_info("%s %s: cmpxchg double redo ", n, s->name);
415 #endif
416
417 return false;
418 }
419
420 #ifdef CONFIG_SLUB_DEBUG
421 /*
422 * Determine a map of object in use on a page.
423 *
424 * Node listlock must be held to guarantee that the page does
425 * not vanish from under us.
426 */
427 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
428 {
429 void *p;
430 void *addr = page_address(page);
431
432 for (p = page->freelist; p; p = get_freepointer(s, p))
433 set_bit(slab_index(p, s, addr), map);
434 }
435
436 static inline int size_from_object(struct kmem_cache *s)
437 {
438 if (s->flags & SLAB_RED_ZONE)
439 return s->size - s->red_left_pad;
440
441 return s->size;
442 }
443
444 static inline void *restore_red_left(struct kmem_cache *s, void *p)
445 {
446 if (s->flags & SLAB_RED_ZONE)
447 p -= s->red_left_pad;
448
449 return p;
450 }
451
452 /*
453 * Debug settings:
454 */
455 #if defined(CONFIG_SLUB_DEBUG_ON)
456 static int slub_debug = DEBUG_DEFAULT_FLAGS;
457 #elif defined(CONFIG_KASAN)
458 static int slub_debug = SLAB_STORE_USER;
459 #else
460 static int slub_debug;
461 #endif
462
463 static char *slub_debug_slabs;
464 static int disable_higher_order_debug;
465
466 /*
467 * slub is about to manipulate internal object metadata. This memory lies
468 * outside the range of the allocated object, so accessing it would normally
469 * be reported by kasan as a bounds error. metadata_access_enable() is used
470 * to tell kasan that these accesses are OK.
471 */
472 static inline void metadata_access_enable(void)
473 {
474 kasan_disable_current();
475 }
476
477 static inline void metadata_access_disable(void)
478 {
479 kasan_enable_current();
480 }
481
482 /*
483 * Object debugging
484 */
485
486 /* Verify that a pointer has an address that is valid within a slab page */
487 static inline int check_valid_pointer(struct kmem_cache *s,
488 struct page *page, void *object)
489 {
490 void *base;
491
492 if (!object)
493 return 1;
494
495 base = page_address(page);
496 object = restore_red_left(s, object);
497 if (object < base || object >= base + page->objects * s->size ||
498 (object - base) % s->size) {
499 return 0;
500 }
501
502 return 1;
503 }
504
505 static void print_section(char *text, u8 *addr, unsigned int length)
506 {
507 metadata_access_enable();
508 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
509 length, 1);
510 metadata_access_disable();
511 }
512
513 static struct track *get_track(struct kmem_cache *s, void *object,
514 enum track_item alloc)
515 {
516 struct track *p;
517
518 if (s->offset)
519 p = object + s->offset + sizeof(void *);
520 else
521 p = object + s->inuse;
522
523 return p + alloc;
524 }
525
526 static void set_track(struct kmem_cache *s, void *object,
527 enum track_item alloc, unsigned long addr)
528 {
529 struct track *p = get_track(s, object, alloc);
530
531 if (addr) {
532 #ifdef CONFIG_STACKTRACE
533 struct stack_trace trace;
534 int i;
535
536 trace.nr_entries = 0;
537 trace.max_entries = TRACK_ADDRS_COUNT;
538 trace.entries = p->addrs;
539 trace.skip = 3;
540 metadata_access_enable();
541 save_stack_trace(&trace);
542 metadata_access_disable();
543
544 /* See rant in lockdep.c */
545 if (trace.nr_entries != 0 &&
546 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
547 trace.nr_entries--;
548
549 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
550 p->addrs[i] = 0;
551 #endif
552 p->addr = addr;
553 p->cpu = smp_processor_id();
554 p->pid = current->pid;
555 p->when = jiffies;
556 } else
557 memset(p, 0, sizeof(struct track));
558 }
559
560 static void init_tracking(struct kmem_cache *s, void *object)
561 {
562 if (!(s->flags & SLAB_STORE_USER))
563 return;
564
565 set_track(s, object, TRACK_FREE, 0UL);
566 set_track(s, object, TRACK_ALLOC, 0UL);
567 }
568
569 static void print_track(const char *s, struct track *t)
570 {
571 if (!t->addr)
572 return;
573
574 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
575 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
576 #ifdef CONFIG_STACKTRACE
577 {
578 int i;
579 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
580 if (t->addrs[i])
581 pr_err("\t%pS\n", (void *)t->addrs[i]);
582 else
583 break;
584 }
585 #endif
586 }
587
588 static void print_tracking(struct kmem_cache *s, void *object)
589 {
590 if (!(s->flags & SLAB_STORE_USER))
591 return;
592
593 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
594 print_track("Freed", get_track(s, object, TRACK_FREE));
595 }
596
597 static void print_page_info(struct page *page)
598 {
599 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
600 page, page->objects, page->inuse, page->freelist, page->flags);
601
602 }
603
604 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
605 {
606 struct va_format vaf;
607 va_list args;
608
609 va_start(args, fmt);
610 vaf.fmt = fmt;
611 vaf.va = &args;
612 pr_err("=============================================================================\n");
613 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
614 pr_err("-----------------------------------------------------------------------------\n\n");
615
616 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
617 va_end(args);
618 }
619
620 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
621 {
622 struct va_format vaf;
623 va_list args;
624
625 va_start(args, fmt);
626 vaf.fmt = fmt;
627 vaf.va = &args;
628 pr_err("FIX %s: %pV\n", s->name, &vaf);
629 va_end(args);
630 }
631
632 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
633 {
634 unsigned int off; /* Offset of last byte */
635 u8 *addr = page_address(page);
636
637 print_tracking(s, p);
638
639 print_page_info(page);
640
641 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
642 p, p - addr, get_freepointer(s, p));
643
644 if (s->flags & SLAB_RED_ZONE)
645 print_section("Redzone ", p - s->red_left_pad, s->red_left_pad);
646 else if (p > addr + 16)
647 print_section("Bytes b4 ", p - 16, 16);
648
649 print_section("Object ", p, min_t(unsigned long, s->object_size,
650 PAGE_SIZE));
651 if (s->flags & SLAB_RED_ZONE)
652 print_section("Redzone ", p + s->object_size,
653 s->inuse - s->object_size);
654
655 if (s->offset)
656 off = s->offset + sizeof(void *);
657 else
658 off = s->inuse;
659
660 if (s->flags & SLAB_STORE_USER)
661 off += 2 * sizeof(struct track);
662
663 if (off != size_from_object(s))
664 /* Beginning of the filler is the free pointer */
665 print_section("Padding ", p + off, size_from_object(s) - off);
666
667 dump_stack();
668 }
669
670 void object_err(struct kmem_cache *s, struct page *page,
671 u8 *object, char *reason)
672 {
673 slab_bug(s, "%s", reason);
674 print_trailer(s, page, object);
675 }
676
677 static void slab_err(struct kmem_cache *s, struct page *page,
678 const char *fmt, ...)
679 {
680 va_list args;
681 char buf[100];
682
683 va_start(args, fmt);
684 vsnprintf(buf, sizeof(buf), fmt, args);
685 va_end(args);
686 slab_bug(s, "%s", buf);
687 print_page_info(page);
688 dump_stack();
689 }
690
691 static void init_object(struct kmem_cache *s, void *object, u8 val)
692 {
693 u8 *p = object;
694
695 if (s->flags & SLAB_RED_ZONE)
696 memset(p - s->red_left_pad, val, s->red_left_pad);
697
698 if (s->flags & __OBJECT_POISON) {
699 memset(p, POISON_FREE, s->object_size - 1);
700 p[s->object_size - 1] = POISON_END;
701 }
702
703 if (s->flags & SLAB_RED_ZONE)
704 memset(p + s->object_size, val, s->inuse - s->object_size);
705 }
706
707 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
708 void *from, void *to)
709 {
710 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
711 memset(from, data, to - from);
712 }
713
714 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
715 u8 *object, char *what,
716 u8 *start, unsigned int value, unsigned int bytes)
717 {
718 u8 *fault;
719 u8 *end;
720
721 metadata_access_enable();
722 fault = memchr_inv(start, value, bytes);
723 metadata_access_disable();
724 if (!fault)
725 return 1;
726
727 end = start + bytes;
728 while (end > fault && end[-1] == value)
729 end--;
730
731 slab_bug(s, "%s overwritten", what);
732 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
733 fault, end - 1, fault[0], value);
734 print_trailer(s, page, object);
735
736 restore_bytes(s, what, value, fault, end);
737 return 0;
738 }
739
740 /*
741 * Object layout:
742 *
743 * object address
744 * Bytes of the object to be managed.
745 * If the freepointer may overlay the object then the free
746 * pointer is the first word of the object.
747 *
748 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
749 * 0xa5 (POISON_END)
750 *
751 * object + s->object_size
752 * Padding to reach word boundary. This is also used for Redzoning.
753 * Padding is extended by another word if Redzoning is enabled and
754 * object_size == inuse.
755 *
756 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
757 * 0xcc (RED_ACTIVE) for objects in use.
758 *
759 * object + s->inuse
760 * Meta data starts here.
761 *
762 * A. Free pointer (if we cannot overwrite object on free)
763 * B. Tracking data for SLAB_STORE_USER
764 * C. Padding to reach required alignment boundary or at mininum
765 * one word if debugging is on to be able to detect writes
766 * before the word boundary.
767 *
768 * Padding is done using 0x5a (POISON_INUSE)
769 *
770 * object + s->size
771 * Nothing is used beyond s->size.
772 *
773 * If slabcaches are merged then the object_size and inuse boundaries are mostly
774 * ignored. And therefore no slab options that rely on these boundaries
775 * may be used with merged slabcaches.
776 */
777
778 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
779 {
780 unsigned long off = s->inuse; /* The end of info */
781
782 if (s->offset)
783 /* Freepointer is placed after the object. */
784 off += sizeof(void *);
785
786 if (s->flags & SLAB_STORE_USER)
787 /* We also have user information there */
788 off += 2 * sizeof(struct track);
789
790 if (size_from_object(s) == off)
791 return 1;
792
793 return check_bytes_and_report(s, page, p, "Object padding",
794 p + off, POISON_INUSE, size_from_object(s) - off);
795 }
796
797 /* Check the pad bytes at the end of a slab page */
798 static int slab_pad_check(struct kmem_cache *s, struct page *page)
799 {
800 u8 *start;
801 u8 *fault;
802 u8 *end;
803 int length;
804 int remainder;
805
806 if (!(s->flags & SLAB_POISON))
807 return 1;
808
809 start = page_address(page);
810 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
811 end = start + length;
812 remainder = length % s->size;
813 if (!remainder)
814 return 1;
815
816 metadata_access_enable();
817 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
818 metadata_access_disable();
819 if (!fault)
820 return 1;
821 while (end > fault && end[-1] == POISON_INUSE)
822 end--;
823
824 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
825 print_section("Padding ", end - remainder, remainder);
826
827 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
828 return 0;
829 }
830
831 static int check_object(struct kmem_cache *s, struct page *page,
832 void *object, u8 val)
833 {
834 u8 *p = object;
835 u8 *endobject = object + s->object_size;
836
837 if (s->flags & SLAB_RED_ZONE) {
838 if (!check_bytes_and_report(s, page, object, "Redzone",
839 object - s->red_left_pad, val, s->red_left_pad))
840 return 0;
841
842 if (!check_bytes_and_report(s, page, object, "Redzone",
843 endobject, val, s->inuse - s->object_size))
844 return 0;
845 } else {
846 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
847 check_bytes_and_report(s, page, p, "Alignment padding",
848 endobject, POISON_INUSE,
849 s->inuse - s->object_size);
850 }
851 }
852
853 if (s->flags & SLAB_POISON) {
854 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
855 (!check_bytes_and_report(s, page, p, "Poison", p,
856 POISON_FREE, s->object_size - 1) ||
857 !check_bytes_and_report(s, page, p, "Poison",
858 p + s->object_size - 1, POISON_END, 1)))
859 return 0;
860 /*
861 * check_pad_bytes cleans up on its own.
862 */
863 check_pad_bytes(s, page, p);
864 }
865
866 if (!s->offset && val == SLUB_RED_ACTIVE)
867 /*
868 * Object and freepointer overlap. Cannot check
869 * freepointer while object is allocated.
870 */
871 return 1;
872
873 /* Check free pointer validity */
874 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
875 object_err(s, page, p, "Freepointer corrupt");
876 /*
877 * No choice but to zap it and thus lose the remainder
878 * of the free objects in this slab. May cause
879 * another error because the object count is now wrong.
880 */
881 set_freepointer(s, p, NULL);
882 return 0;
883 }
884 return 1;
885 }
886
887 static int check_slab(struct kmem_cache *s, struct page *page)
888 {
889 int maxobj;
890
891 VM_BUG_ON(!irqs_disabled());
892
893 if (!PageSlab(page)) {
894 slab_err(s, page, "Not a valid slab page");
895 return 0;
896 }
897
898 maxobj = order_objects(compound_order(page), s->size, s->reserved);
899 if (page->objects > maxobj) {
900 slab_err(s, page, "objects %u > max %u",
901 page->objects, maxobj);
902 return 0;
903 }
904 if (page->inuse > page->objects) {
905 slab_err(s, page, "inuse %u > max %u",
906 page->inuse, page->objects);
907 return 0;
908 }
909 /* Slab_pad_check fixes things up after itself */
910 slab_pad_check(s, page);
911 return 1;
912 }
913
914 /*
915 * Determine if a certain object on a page is on the freelist. Must hold the
916 * slab lock to guarantee that the chains are in a consistent state.
917 */
918 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
919 {
920 int nr = 0;
921 void *fp;
922 void *object = NULL;
923 int max_objects;
924
925 fp = page->freelist;
926 while (fp && nr <= page->objects) {
927 if (fp == search)
928 return 1;
929 if (!check_valid_pointer(s, page, fp)) {
930 if (object) {
931 object_err(s, page, object,
932 "Freechain corrupt");
933 set_freepointer(s, object, NULL);
934 } else {
935 slab_err(s, page, "Freepointer corrupt");
936 page->freelist = NULL;
937 page->inuse = page->objects;
938 slab_fix(s, "Freelist cleared");
939 return 0;
940 }
941 break;
942 }
943 object = fp;
944 fp = get_freepointer(s, object);
945 nr++;
946 }
947
948 max_objects = order_objects(compound_order(page), s->size, s->reserved);
949 if (max_objects > MAX_OBJS_PER_PAGE)
950 max_objects = MAX_OBJS_PER_PAGE;
951
952 if (page->objects != max_objects) {
953 slab_err(s, page, "Wrong number of objects. Found %d but should be %d",
954 page->objects, max_objects);
955 page->objects = max_objects;
956 slab_fix(s, "Number of objects adjusted.");
957 }
958 if (page->inuse != page->objects - nr) {
959 slab_err(s, page, "Wrong object count. Counter is %d but counted were %d",
960 page->inuse, page->objects - nr);
961 page->inuse = page->objects - nr;
962 slab_fix(s, "Object count adjusted.");
963 }
964 return search == NULL;
965 }
966
967 static void trace(struct kmem_cache *s, struct page *page, void *object,
968 int alloc)
969 {
970 if (s->flags & SLAB_TRACE) {
971 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
972 s->name,
973 alloc ? "alloc" : "free",
974 object, page->inuse,
975 page->freelist);
976
977 if (!alloc)
978 print_section("Object ", (void *)object,
979 s->object_size);
980
981 dump_stack();
982 }
983 }
984
985 /*
986 * Tracking of fully allocated slabs for debugging purposes.
987 */
988 static void add_full(struct kmem_cache *s,
989 struct kmem_cache_node *n, struct page *page)
990 {
991 if (!(s->flags & SLAB_STORE_USER))
992 return;
993
994 lockdep_assert_held(&n->list_lock);
995 list_add(&page->lru, &n->full);
996 }
997
998 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
999 {
1000 if (!(s->flags & SLAB_STORE_USER))
1001 return;
1002
1003 lockdep_assert_held(&n->list_lock);
1004 list_del(&page->lru);
1005 }
1006
1007 /* Tracking of the number of slabs for debugging purposes */
1008 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1009 {
1010 struct kmem_cache_node *n = get_node(s, node);
1011
1012 return atomic_long_read(&n->nr_slabs);
1013 }
1014
1015 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1016 {
1017 return atomic_long_read(&n->nr_slabs);
1018 }
1019
1020 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1021 {
1022 struct kmem_cache_node *n = get_node(s, node);
1023
1024 /*
1025 * May be called early in order to allocate a slab for the
1026 * kmem_cache_node structure. Solve the chicken-egg
1027 * dilemma by deferring the increment of the count during
1028 * bootstrap (see early_kmem_cache_node_alloc).
1029 */
1030 if (likely(n)) {
1031 atomic_long_inc(&n->nr_slabs);
1032 atomic_long_add(objects, &n->total_objects);
1033 }
1034 }
1035 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1036 {
1037 struct kmem_cache_node *n = get_node(s, node);
1038
1039 atomic_long_dec(&n->nr_slabs);
1040 atomic_long_sub(objects, &n->total_objects);
1041 }
1042
1043 /* Object debug checks for alloc/free paths */
1044 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1045 void *object)
1046 {
1047 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1048 return;
1049
1050 init_object(s, object, SLUB_RED_INACTIVE);
1051 init_tracking(s, object);
1052 }
1053
1054 static inline int alloc_consistency_checks(struct kmem_cache *s,
1055 struct page *page,
1056 void *object, unsigned long addr)
1057 {
1058 if (!check_slab(s, page))
1059 return 0;
1060
1061 if (!check_valid_pointer(s, page, object)) {
1062 object_err(s, page, object, "Freelist Pointer check fails");
1063 return 0;
1064 }
1065
1066 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1067 return 0;
1068
1069 return 1;
1070 }
1071
1072 static noinline int alloc_debug_processing(struct kmem_cache *s,
1073 struct page *page,
1074 void *object, unsigned long addr)
1075 {
1076 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1077 if (!alloc_consistency_checks(s, page, object, addr))
1078 goto bad;
1079 }
1080
1081 /* Success perform special debug activities for allocs */
1082 if (s->flags & SLAB_STORE_USER)
1083 set_track(s, object, TRACK_ALLOC, addr);
1084 trace(s, page, object, 1);
1085 init_object(s, object, SLUB_RED_ACTIVE);
1086 return 1;
1087
1088 bad:
1089 if (PageSlab(page)) {
1090 /*
1091 * If this is a slab page then lets do the best we can
1092 * to avoid issues in the future. Marking all objects
1093 * as used avoids touching the remaining objects.
1094 */
1095 slab_fix(s, "Marking all objects used");
1096 page->inuse = page->objects;
1097 page->freelist = NULL;
1098 }
1099 return 0;
1100 }
1101
1102 static inline int free_consistency_checks(struct kmem_cache *s,
1103 struct page *page, void *object, unsigned long addr)
1104 {
1105 if (!check_valid_pointer(s, page, object)) {
1106 slab_err(s, page, "Invalid object pointer 0x%p", object);
1107 return 0;
1108 }
1109
1110 if (on_freelist(s, page, object)) {
1111 object_err(s, page, object, "Object already free");
1112 return 0;
1113 }
1114
1115 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1116 return 0;
1117
1118 if (unlikely(s != page->slab_cache)) {
1119 if (!PageSlab(page)) {
1120 slab_err(s, page, "Attempt to free object(0x%p) outside of slab",
1121 object);
1122 } else if (!page->slab_cache) {
1123 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1124 object);
1125 dump_stack();
1126 } else
1127 object_err(s, page, object,
1128 "page slab pointer corrupt.");
1129 return 0;
1130 }
1131 return 1;
1132 }
1133
1134 /* Supports checking bulk free of a constructed freelist */
1135 static noinline int free_debug_processing(
1136 struct kmem_cache *s, struct page *page,
1137 void *head, void *tail, int bulk_cnt,
1138 unsigned long addr)
1139 {
1140 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1141 void *object = head;
1142 int cnt = 0;
1143 unsigned long uninitialized_var(flags);
1144 int ret = 0;
1145
1146 spin_lock_irqsave(&n->list_lock, flags);
1147 slab_lock(page);
1148
1149 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1150 if (!check_slab(s, page))
1151 goto out;
1152 }
1153
1154 next_object:
1155 cnt++;
1156
1157 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1158 if (!free_consistency_checks(s, page, object, addr))
1159 goto out;
1160 }
1161
1162 if (s->flags & SLAB_STORE_USER)
1163 set_track(s, object, TRACK_FREE, addr);
1164 trace(s, page, object, 0);
1165 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1166 init_object(s, object, SLUB_RED_INACTIVE);
1167
1168 /* Reached end of constructed freelist yet? */
1169 if (object != tail) {
1170 object = get_freepointer(s, object);
1171 goto next_object;
1172 }
1173 ret = 1;
1174
1175 out:
1176 if (cnt != bulk_cnt)
1177 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1178 bulk_cnt, cnt);
1179
1180 slab_unlock(page);
1181 spin_unlock_irqrestore(&n->list_lock, flags);
1182 if (!ret)
1183 slab_fix(s, "Object at 0x%p not freed", object);
1184 return ret;
1185 }
1186
1187 static int __init setup_slub_debug(char *str)
1188 {
1189 slub_debug = DEBUG_DEFAULT_FLAGS;
1190 if (*str++ != '=' || !*str)
1191 /*
1192 * No options specified. Switch on full debugging.
1193 */
1194 goto out;
1195
1196 if (*str == ',')
1197 /*
1198 * No options but restriction on slabs. This means full
1199 * debugging for slabs matching a pattern.
1200 */
1201 goto check_slabs;
1202
1203 slub_debug = 0;
1204 if (*str == '-')
1205 /*
1206 * Switch off all debugging measures.
1207 */
1208 goto out;
1209
1210 /*
1211 * Determine which debug features should be switched on
1212 */
1213 for (; *str && *str != ','; str++) {
1214 switch (tolower(*str)) {
1215 case 'f':
1216 slub_debug |= SLAB_CONSISTENCY_CHECKS;
1217 break;
1218 case 'z':
1219 slub_debug |= SLAB_RED_ZONE;
1220 break;
1221 case 'p':
1222 slub_debug |= SLAB_POISON;
1223 break;
1224 case 'u':
1225 slub_debug |= SLAB_STORE_USER;
1226 break;
1227 case 't':
1228 slub_debug |= SLAB_TRACE;
1229 break;
1230 case 'a':
1231 slub_debug |= SLAB_FAILSLAB;
1232 break;
1233 case 'o':
1234 /*
1235 * Avoid enabling debugging on caches if its minimum
1236 * order would increase as a result.
1237 */
1238 disable_higher_order_debug = 1;
1239 break;
1240 default:
1241 pr_err("slub_debug option '%c' unknown. skipped\n",
1242 *str);
1243 }
1244 }
1245
1246 check_slabs:
1247 if (*str == ',')
1248 slub_debug_slabs = str + 1;
1249 out:
1250 return 1;
1251 }
1252
1253 __setup("slub_debug", setup_slub_debug);
1254
1255 unsigned long kmem_cache_flags(unsigned long object_size,
1256 unsigned long flags, const char *name,
1257 void (*ctor)(void *))
1258 {
1259 /*
1260 * Enable debugging if selected on the kernel commandline.
1261 */
1262 if (slub_debug && (!slub_debug_slabs || (name &&
1263 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1264 flags |= slub_debug;
1265
1266 return flags;
1267 }
1268 #else /* !CONFIG_SLUB_DEBUG */
1269 static inline void setup_object_debug(struct kmem_cache *s,
1270 struct page *page, void *object) {}
1271
1272 static inline int alloc_debug_processing(struct kmem_cache *s,
1273 struct page *page, void *object, unsigned long addr) { return 0; }
1274
1275 static inline int free_debug_processing(
1276 struct kmem_cache *s, struct page *page,
1277 void *head, void *tail, int bulk_cnt,
1278 unsigned long addr) { return 0; }
1279
1280 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1281 { return 1; }
1282 static inline int check_object(struct kmem_cache *s, struct page *page,
1283 void *object, u8 val) { return 1; }
1284 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1285 struct page *page) {}
1286 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1287 struct page *page) {}
1288 unsigned long kmem_cache_flags(unsigned long object_size,
1289 unsigned long flags, const char *name,
1290 void (*ctor)(void *))
1291 {
1292 return flags;
1293 }
1294 #define slub_debug 0
1295
1296 #define disable_higher_order_debug 0
1297
1298 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1299 { return 0; }
1300 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1301 { return 0; }
1302 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1303 int objects) {}
1304 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1305 int objects) {}
1306
1307 #endif /* CONFIG_SLUB_DEBUG */
1308
1309 /*
1310 * Hooks for other subsystems that check memory allocations. In a typical
1311 * production configuration these hooks all should produce no code at all.
1312 */
1313 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1314 {
1315 kmemleak_alloc(ptr, size, 1, flags);
1316 kasan_kmalloc_large(ptr, size, flags);
1317 }
1318
1319 static inline void kfree_hook(const void *x)
1320 {
1321 kmemleak_free(x);
1322 kasan_kfree_large(x);
1323 }
1324
1325 static inline void slab_free_hook(struct kmem_cache *s, void *x)
1326 {
1327 kmemleak_free_recursive(x, s->flags);
1328
1329 /*
1330 * Trouble is that we may no longer disable interrupts in the fast path
1331 * So in order to make the debug calls that expect irqs to be
1332 * disabled we need to disable interrupts temporarily.
1333 */
1334 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1335 {
1336 unsigned long flags;
1337
1338 local_irq_save(flags);
1339 kmemcheck_slab_free(s, x, s->object_size);
1340 debug_check_no_locks_freed(x, s->object_size);
1341 local_irq_restore(flags);
1342 }
1343 #endif
1344 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1345 debug_check_no_obj_freed(x, s->object_size);
1346
1347 kasan_slab_free(s, x);
1348 }
1349
1350 static inline void slab_free_freelist_hook(struct kmem_cache *s,
1351 void *head, void *tail)
1352 {
1353 /*
1354 * Compiler cannot detect this function can be removed if slab_free_hook()
1355 * evaluates to nothing. Thus, catch all relevant config debug options here.
1356 */
1357 #if defined(CONFIG_KMEMCHECK) || \
1358 defined(CONFIG_LOCKDEP) || \
1359 defined(CONFIG_DEBUG_KMEMLEAK) || \
1360 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1361 defined(CONFIG_KASAN)
1362
1363 void *object = head;
1364 void *tail_obj = tail ? : head;
1365
1366 do {
1367 slab_free_hook(s, object);
1368 } while ((object != tail_obj) &&
1369 (object = get_freepointer(s, object)));
1370 #endif
1371 }
1372
1373 static void setup_object(struct kmem_cache *s, struct page *page,
1374 void *object)
1375 {
1376 setup_object_debug(s, page, object);
1377 if (unlikely(s->ctor)) {
1378 kasan_unpoison_object_data(s, object);
1379 s->ctor(object);
1380 kasan_poison_object_data(s, object);
1381 }
1382 }
1383
1384 /*
1385 * Slab allocation and freeing
1386 */
1387 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1388 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1389 {
1390 struct page *page;
1391 int order = oo_order(oo);
1392
1393 flags |= __GFP_NOTRACK;
1394
1395 if (node == NUMA_NO_NODE)
1396 page = alloc_pages(flags, order);
1397 else
1398 page = __alloc_pages_node(node, flags, order);
1399
1400 if (page && memcg_charge_slab(page, flags, order, s)) {
1401 __free_pages(page, order);
1402 page = NULL;
1403 }
1404
1405 return page;
1406 }
1407
1408 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1409 {
1410 struct page *page;
1411 struct kmem_cache_order_objects oo = s->oo;
1412 gfp_t alloc_gfp;
1413 void *start, *p;
1414 int idx, order;
1415
1416 flags &= gfp_allowed_mask;
1417
1418 if (gfpflags_allow_blocking(flags))
1419 local_irq_enable();
1420
1421 flags |= s->allocflags;
1422
1423 /*
1424 * Let the initial higher-order allocation fail under memory pressure
1425 * so we fall-back to the minimum order allocation.
1426 */
1427 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1428 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1429 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
1430
1431 page = alloc_slab_page(s, alloc_gfp, node, oo);
1432 if (unlikely(!page)) {
1433 oo = s->min;
1434 alloc_gfp = flags;
1435 /*
1436 * Allocation may have failed due to fragmentation.
1437 * Try a lower order alloc if possible
1438 */
1439 page = alloc_slab_page(s, alloc_gfp, node, oo);
1440 if (unlikely(!page))
1441 goto out;
1442 stat(s, ORDER_FALLBACK);
1443 }
1444
1445 if (kmemcheck_enabled &&
1446 !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1447 int pages = 1 << oo_order(oo);
1448
1449 kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node);
1450
1451 /*
1452 * Objects from caches that have a constructor don't get
1453 * cleared when they're allocated, so we need to do it here.
1454 */
1455 if (s->ctor)
1456 kmemcheck_mark_uninitialized_pages(page, pages);
1457 else
1458 kmemcheck_mark_unallocated_pages(page, pages);
1459 }
1460
1461 page->objects = oo_objects(oo);
1462
1463 order = compound_order(page);
1464 page->slab_cache = s;
1465 __SetPageSlab(page);
1466 if (page_is_pfmemalloc(page))
1467 SetPageSlabPfmemalloc(page);
1468
1469 start = page_address(page);
1470
1471 if (unlikely(s->flags & SLAB_POISON))
1472 memset(start, POISON_INUSE, PAGE_SIZE << order);
1473
1474 kasan_poison_slab(page);
1475
1476 for_each_object_idx(p, idx, s, start, page->objects) {
1477 setup_object(s, page, p);
1478 if (likely(idx < page->objects))
1479 set_freepointer(s, p, p + s->size);
1480 else
1481 set_freepointer(s, p, NULL);
1482 }
1483
1484 page->freelist = fixup_red_left(s, start);
1485 page->inuse = page->objects;
1486 page->frozen = 1;
1487
1488 out:
1489 if (gfpflags_allow_blocking(flags))
1490 local_irq_disable();
1491 if (!page)
1492 return NULL;
1493
1494 mod_zone_page_state(page_zone(page),
1495 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1496 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1497 1 << oo_order(oo));
1498
1499 inc_slabs_node(s, page_to_nid(page), page->objects);
1500
1501 return page;
1502 }
1503
1504 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1505 {
1506 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1507 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
1508 BUG();
1509 }
1510
1511 return allocate_slab(s,
1512 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1513 }
1514
1515 static void __free_slab(struct kmem_cache *s, struct page *page)
1516 {
1517 int order = compound_order(page);
1518 int pages = 1 << order;
1519
1520 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1521 void *p;
1522
1523 slab_pad_check(s, page);
1524 for_each_object(p, s, page_address(page),
1525 page->objects)
1526 check_object(s, page, p, SLUB_RED_INACTIVE);
1527 }
1528
1529 kmemcheck_free_shadow(page, compound_order(page));
1530
1531 mod_zone_page_state(page_zone(page),
1532 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1533 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1534 -pages);
1535
1536 __ClearPageSlabPfmemalloc(page);
1537 __ClearPageSlab(page);
1538
1539 page_mapcount_reset(page);
1540 if (current->reclaim_state)
1541 current->reclaim_state->reclaimed_slab += pages;
1542 memcg_uncharge_slab(page, order, s);
1543 __free_pages(page, order);
1544 }
1545
1546 #define need_reserve_slab_rcu \
1547 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1548
1549 static void rcu_free_slab(struct rcu_head *h)
1550 {
1551 struct page *page;
1552
1553 if (need_reserve_slab_rcu)
1554 page = virt_to_head_page(h);
1555 else
1556 page = container_of((struct list_head *)h, struct page, lru);
1557
1558 __free_slab(page->slab_cache, page);
1559 }
1560
1561 static void free_slab(struct kmem_cache *s, struct page *page)
1562 {
1563 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1564 struct rcu_head *head;
1565
1566 if (need_reserve_slab_rcu) {
1567 int order = compound_order(page);
1568 int offset = (PAGE_SIZE << order) - s->reserved;
1569
1570 VM_BUG_ON(s->reserved != sizeof(*head));
1571 head = page_address(page) + offset;
1572 } else {
1573 head = &page->rcu_head;
1574 }
1575
1576 call_rcu(head, rcu_free_slab);
1577 } else
1578 __free_slab(s, page);
1579 }
1580
1581 static void discard_slab(struct kmem_cache *s, struct page *page)
1582 {
1583 dec_slabs_node(s, page_to_nid(page), page->objects);
1584 free_slab(s, page);
1585 }
1586
1587 /*
1588 * Management of partially allocated slabs.
1589 */
1590 static inline void
1591 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1592 {
1593 n->nr_partial++;
1594 if (tail == DEACTIVATE_TO_TAIL)
1595 list_add_tail(&page->lru, &n->partial);
1596 else
1597 list_add(&page->lru, &n->partial);
1598 }
1599
1600 static inline void add_partial(struct kmem_cache_node *n,
1601 struct page *page, int tail)
1602 {
1603 lockdep_assert_held(&n->list_lock);
1604 __add_partial(n, page, tail);
1605 }
1606
1607 static inline void remove_partial(struct kmem_cache_node *n,
1608 struct page *page)
1609 {
1610 lockdep_assert_held(&n->list_lock);
1611 list_del(&page->lru);
1612 n->nr_partial--;
1613 }
1614
1615 /*
1616 * Remove slab from the partial list, freeze it and
1617 * return the pointer to the freelist.
1618 *
1619 * Returns a list of objects or NULL if it fails.
1620 */
1621 static inline void *acquire_slab(struct kmem_cache *s,
1622 struct kmem_cache_node *n, struct page *page,
1623 int mode, int *objects)
1624 {
1625 void *freelist;
1626 unsigned long counters;
1627 struct page new;
1628
1629 lockdep_assert_held(&n->list_lock);
1630
1631 /*
1632 * Zap the freelist and set the frozen bit.
1633 * The old freelist is the list of objects for the
1634 * per cpu allocation list.
1635 */
1636 freelist = page->freelist;
1637 counters = page->counters;
1638 new.counters = counters;
1639 *objects = new.objects - new.inuse;
1640 if (mode) {
1641 new.inuse = page->objects;
1642 new.freelist = NULL;
1643 } else {
1644 new.freelist = freelist;
1645 }
1646
1647 VM_BUG_ON(new.frozen);
1648 new.frozen = 1;
1649
1650 if (!__cmpxchg_double_slab(s, page,
1651 freelist, counters,
1652 new.freelist, new.counters,
1653 "acquire_slab"))
1654 return NULL;
1655
1656 remove_partial(n, page);
1657 WARN_ON(!freelist);
1658 return freelist;
1659 }
1660
1661 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1662 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1663
1664 /*
1665 * Try to allocate a partial slab from a specific node.
1666 */
1667 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1668 struct kmem_cache_cpu *c, gfp_t flags)
1669 {
1670 struct page *page, *page2;
1671 void *object = NULL;
1672 int available = 0;
1673 int objects;
1674
1675 /*
1676 * Racy check. If we mistakenly see no partial slabs then we
1677 * just allocate an empty slab. If we mistakenly try to get a
1678 * partial slab and there is none available then get_partials()
1679 * will return NULL.
1680 */
1681 if (!n || !n->nr_partial)
1682 return NULL;
1683
1684 spin_lock(&n->list_lock);
1685 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1686 void *t;
1687
1688 if (!pfmemalloc_match(page, flags))
1689 continue;
1690
1691 t = acquire_slab(s, n, page, object == NULL, &objects);
1692 if (!t)
1693 break;
1694
1695 available += objects;
1696 if (!object) {
1697 c->page = page;
1698 stat(s, ALLOC_FROM_PARTIAL);
1699 object = t;
1700 } else {
1701 put_cpu_partial(s, page, 0);
1702 stat(s, CPU_PARTIAL_NODE);
1703 }
1704 if (!kmem_cache_has_cpu_partial(s)
1705 || available > s->cpu_partial / 2)
1706 break;
1707
1708 }
1709 spin_unlock(&n->list_lock);
1710 return object;
1711 }
1712
1713 /*
1714 * Get a page from somewhere. Search in increasing NUMA distances.
1715 */
1716 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1717 struct kmem_cache_cpu *c)
1718 {
1719 #ifdef CONFIG_NUMA
1720 struct zonelist *zonelist;
1721 struct zoneref *z;
1722 struct zone *zone;
1723 enum zone_type high_zoneidx = gfp_zone(flags);
1724 void *object;
1725 unsigned int cpuset_mems_cookie;
1726
1727 /*
1728 * The defrag ratio allows a configuration of the tradeoffs between
1729 * inter node defragmentation and node local allocations. A lower
1730 * defrag_ratio increases the tendency to do local allocations
1731 * instead of attempting to obtain partial slabs from other nodes.
1732 *
1733 * If the defrag_ratio is set to 0 then kmalloc() always
1734 * returns node local objects. If the ratio is higher then kmalloc()
1735 * may return off node objects because partial slabs are obtained
1736 * from other nodes and filled up.
1737 *
1738 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
1739 * (which makes defrag_ratio = 1000) then every (well almost)
1740 * allocation will first attempt to defrag slab caches on other nodes.
1741 * This means scanning over all nodes to look for partial slabs which
1742 * may be expensive if we do it every time we are trying to find a slab
1743 * with available objects.
1744 */
1745 if (!s->remote_node_defrag_ratio ||
1746 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1747 return NULL;
1748
1749 do {
1750 cpuset_mems_cookie = read_mems_allowed_begin();
1751 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1752 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1753 struct kmem_cache_node *n;
1754
1755 n = get_node(s, zone_to_nid(zone));
1756
1757 if (n && cpuset_zone_allowed(zone, flags) &&
1758 n->nr_partial > s->min_partial) {
1759 object = get_partial_node(s, n, c, flags);
1760 if (object) {
1761 /*
1762 * Don't check read_mems_allowed_retry()
1763 * here - if mems_allowed was updated in
1764 * parallel, that was a harmless race
1765 * between allocation and the cpuset
1766 * update
1767 */
1768 return object;
1769 }
1770 }
1771 }
1772 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1773 #endif
1774 return NULL;
1775 }
1776
1777 /*
1778 * Get a partial page, lock it and return it.
1779 */
1780 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1781 struct kmem_cache_cpu *c)
1782 {
1783 void *object;
1784 int searchnode = node;
1785
1786 if (node == NUMA_NO_NODE)
1787 searchnode = numa_mem_id();
1788 else if (!node_present_pages(node))
1789 searchnode = node_to_mem_node(node);
1790
1791 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1792 if (object || node != NUMA_NO_NODE)
1793 return object;
1794
1795 return get_any_partial(s, flags, c);
1796 }
1797
1798 #ifdef CONFIG_PREEMPT
1799 /*
1800 * Calculate the next globally unique transaction for disambiguiation
1801 * during cmpxchg. The transactions start with the cpu number and are then
1802 * incremented by CONFIG_NR_CPUS.
1803 */
1804 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1805 #else
1806 /*
1807 * No preemption supported therefore also no need to check for
1808 * different cpus.
1809 */
1810 #define TID_STEP 1
1811 #endif
1812
1813 static inline unsigned long next_tid(unsigned long tid)
1814 {
1815 return tid + TID_STEP;
1816 }
1817
1818 static inline unsigned int tid_to_cpu(unsigned long tid)
1819 {
1820 return tid % TID_STEP;
1821 }
1822
1823 static inline unsigned long tid_to_event(unsigned long tid)
1824 {
1825 return tid / TID_STEP;
1826 }
1827
1828 static inline unsigned int init_tid(int cpu)
1829 {
1830 return cpu;
1831 }
1832
1833 static inline void note_cmpxchg_failure(const char *n,
1834 const struct kmem_cache *s, unsigned long tid)
1835 {
1836 #ifdef SLUB_DEBUG_CMPXCHG
1837 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1838
1839 pr_info("%s %s: cmpxchg redo ", n, s->name);
1840
1841 #ifdef CONFIG_PREEMPT
1842 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1843 pr_warn("due to cpu change %d -> %d\n",
1844 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1845 else
1846 #endif
1847 if (tid_to_event(tid) != tid_to_event(actual_tid))
1848 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1849 tid_to_event(tid), tid_to_event(actual_tid));
1850 else
1851 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1852 actual_tid, tid, next_tid(tid));
1853 #endif
1854 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1855 }
1856
1857 static void init_kmem_cache_cpus(struct kmem_cache *s)
1858 {
1859 int cpu;
1860
1861 for_each_possible_cpu(cpu)
1862 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1863 }
1864
1865 /*
1866 * Remove the cpu slab
1867 */
1868 static void deactivate_slab(struct kmem_cache *s, struct page *page,
1869 void *freelist)
1870 {
1871 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1872 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1873 int lock = 0;
1874 enum slab_modes l = M_NONE, m = M_NONE;
1875 void *nextfree;
1876 int tail = DEACTIVATE_TO_HEAD;
1877 struct page new;
1878 struct page old;
1879
1880 if (page->freelist) {
1881 stat(s, DEACTIVATE_REMOTE_FREES);
1882 tail = DEACTIVATE_TO_TAIL;
1883 }
1884
1885 /*
1886 * Stage one: Free all available per cpu objects back
1887 * to the page freelist while it is still frozen. Leave the
1888 * last one.
1889 *
1890 * There is no need to take the list->lock because the page
1891 * is still frozen.
1892 */
1893 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1894 void *prior;
1895 unsigned long counters;
1896
1897 do {
1898 prior = page->freelist;
1899 counters = page->counters;
1900 set_freepointer(s, freelist, prior);
1901 new.counters = counters;
1902 new.inuse--;
1903 VM_BUG_ON(!new.frozen);
1904
1905 } while (!__cmpxchg_double_slab(s, page,
1906 prior, counters,
1907 freelist, new.counters,
1908 "drain percpu freelist"));
1909
1910 freelist = nextfree;
1911 }
1912
1913 /*
1914 * Stage two: Ensure that the page is unfrozen while the
1915 * list presence reflects the actual number of objects
1916 * during unfreeze.
1917 *
1918 * We setup the list membership and then perform a cmpxchg
1919 * with the count. If there is a mismatch then the page
1920 * is not unfrozen but the page is on the wrong list.
1921 *
1922 * Then we restart the process which may have to remove
1923 * the page from the list that we just put it on again
1924 * because the number of objects in the slab may have
1925 * changed.
1926 */
1927 redo:
1928
1929 old.freelist = page->freelist;
1930 old.counters = page->counters;
1931 VM_BUG_ON(!old.frozen);
1932
1933 /* Determine target state of the slab */
1934 new.counters = old.counters;
1935 if (freelist) {
1936 new.inuse--;
1937 set_freepointer(s, freelist, old.freelist);
1938 new.freelist = freelist;
1939 } else
1940 new.freelist = old.freelist;
1941
1942 new.frozen = 0;
1943
1944 if (!new.inuse && n->nr_partial >= s->min_partial)
1945 m = M_FREE;
1946 else if (new.freelist) {
1947 m = M_PARTIAL;
1948 if (!lock) {
1949 lock = 1;
1950 /*
1951 * Taking the spinlock removes the possiblity
1952 * that acquire_slab() will see a slab page that
1953 * is frozen
1954 */
1955 spin_lock(&n->list_lock);
1956 }
1957 } else {
1958 m = M_FULL;
1959 if (kmem_cache_debug(s) && !lock) {
1960 lock = 1;
1961 /*
1962 * This also ensures that the scanning of full
1963 * slabs from diagnostic functions will not see
1964 * any frozen slabs.
1965 */
1966 spin_lock(&n->list_lock);
1967 }
1968 }
1969
1970 if (l != m) {
1971
1972 if (l == M_PARTIAL)
1973
1974 remove_partial(n, page);
1975
1976 else if (l == M_FULL)
1977
1978 remove_full(s, n, page);
1979
1980 if (m == M_PARTIAL) {
1981
1982 add_partial(n, page, tail);
1983 stat(s, tail);
1984
1985 } else if (m == M_FULL) {
1986
1987 stat(s, DEACTIVATE_FULL);
1988 add_full(s, n, page);
1989
1990 }
1991 }
1992
1993 l = m;
1994 if (!__cmpxchg_double_slab(s, page,
1995 old.freelist, old.counters,
1996 new.freelist, new.counters,
1997 "unfreezing slab"))
1998 goto redo;
1999
2000 if (lock)
2001 spin_unlock(&n->list_lock);
2002
2003 if (m == M_FREE) {
2004 stat(s, DEACTIVATE_EMPTY);
2005 discard_slab(s, page);
2006 stat(s, FREE_SLAB);
2007 }
2008 }
2009
2010 /*
2011 * Unfreeze all the cpu partial slabs.
2012 *
2013 * This function must be called with interrupts disabled
2014 * for the cpu using c (or some other guarantee must be there
2015 * to guarantee no concurrent accesses).
2016 */
2017 static void unfreeze_partials(struct kmem_cache *s,
2018 struct kmem_cache_cpu *c)
2019 {
2020 #ifdef CONFIG_SLUB_CPU_PARTIAL
2021 struct kmem_cache_node *n = NULL, *n2 = NULL;
2022 struct page *page, *discard_page = NULL;
2023
2024 while ((page = c->partial)) {
2025 struct page new;
2026 struct page old;
2027
2028 c->partial = page->next;
2029
2030 n2 = get_node(s, page_to_nid(page));
2031 if (n != n2) {
2032 if (n)
2033 spin_unlock(&n->list_lock);
2034
2035 n = n2;
2036 spin_lock(&n->list_lock);
2037 }
2038
2039 do {
2040
2041 old.freelist = page->freelist;
2042 old.counters = page->counters;
2043 VM_BUG_ON(!old.frozen);
2044
2045 new.counters = old.counters;
2046 new.freelist = old.freelist;
2047
2048 new.frozen = 0;
2049
2050 } while (!__cmpxchg_double_slab(s, page,
2051 old.freelist, old.counters,
2052 new.freelist, new.counters,
2053 "unfreezing slab"));
2054
2055 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2056 page->next = discard_page;
2057 discard_page = page;
2058 } else {
2059 add_partial(n, page, DEACTIVATE_TO_TAIL);
2060 stat(s, FREE_ADD_PARTIAL);
2061 }
2062 }
2063
2064 if (n)
2065 spin_unlock(&n->list_lock);
2066
2067 while (discard_page) {
2068 page = discard_page;
2069 discard_page = discard_page->next;
2070
2071 stat(s, DEACTIVATE_EMPTY);
2072 discard_slab(s, page);
2073 stat(s, FREE_SLAB);
2074 }
2075 #endif
2076 }
2077
2078 /*
2079 * Put a page that was just frozen (in __slab_free) into a partial page
2080 * slot if available. This is done without interrupts disabled and without
2081 * preemption disabled. The cmpxchg is racy and may put the partial page
2082 * onto a random cpus partial slot.
2083 *
2084 * If we did not find a slot then simply move all the partials to the
2085 * per node partial list.
2086 */
2087 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2088 {
2089 #ifdef CONFIG_SLUB_CPU_PARTIAL
2090 struct page *oldpage;
2091 int pages;
2092 int pobjects;
2093
2094 preempt_disable();
2095 do {
2096 pages = 0;
2097 pobjects = 0;
2098 oldpage = this_cpu_read(s->cpu_slab->partial);
2099
2100 if (oldpage) {
2101 pobjects = oldpage->pobjects;
2102 pages = oldpage->pages;
2103 if (drain && pobjects > s->cpu_partial) {
2104 unsigned long flags;
2105 /*
2106 * partial array is full. Move the existing
2107 * set to the per node partial list.
2108 */
2109 local_irq_save(flags);
2110 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2111 local_irq_restore(flags);
2112 oldpage = NULL;
2113 pobjects = 0;
2114 pages = 0;
2115 stat(s, CPU_PARTIAL_DRAIN);
2116 }
2117 }
2118
2119 pages++;
2120 pobjects += page->objects - page->inuse;
2121
2122 page->pages = pages;
2123 page->pobjects = pobjects;
2124 page->next = oldpage;
2125
2126 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2127 != oldpage);
2128 if (unlikely(!s->cpu_partial)) {
2129 unsigned long flags;
2130
2131 local_irq_save(flags);
2132 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2133 local_irq_restore(flags);
2134 }
2135 preempt_enable();
2136 #endif
2137 }
2138
2139 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2140 {
2141 stat(s, CPUSLAB_FLUSH);
2142 deactivate_slab(s, c->page, c->freelist);
2143
2144 c->tid = next_tid(c->tid);
2145 c->page = NULL;
2146 c->freelist = NULL;
2147 }
2148
2149 /*
2150 * Flush cpu slab.
2151 *
2152 * Called from IPI handler with interrupts disabled.
2153 */
2154 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2155 {
2156 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2157
2158 if (likely(c)) {
2159 if (c->page)
2160 flush_slab(s, c);
2161
2162 unfreeze_partials(s, c);
2163 }
2164 }
2165
2166 static void flush_cpu_slab(void *d)
2167 {
2168 struct kmem_cache *s = d;
2169
2170 __flush_cpu_slab(s, smp_processor_id());
2171 }
2172
2173 static bool has_cpu_slab(int cpu, void *info)
2174 {
2175 struct kmem_cache *s = info;
2176 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2177
2178 return c->page || c->partial;
2179 }
2180
2181 static void flush_all(struct kmem_cache *s)
2182 {
2183 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2184 }
2185
2186 /*
2187 * Check if the objects in a per cpu structure fit numa
2188 * locality expectations.
2189 */
2190 static inline int node_match(struct page *page, int node)
2191 {
2192 #ifdef CONFIG_NUMA
2193 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2194 return 0;
2195 #endif
2196 return 1;
2197 }
2198
2199 #ifdef CONFIG_SLUB_DEBUG
2200 static int count_free(struct page *page)
2201 {
2202 return page->objects - page->inuse;
2203 }
2204
2205 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2206 {
2207 return atomic_long_read(&n->total_objects);
2208 }
2209 #endif /* CONFIG_SLUB_DEBUG */
2210
2211 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2212 static unsigned long count_partial(struct kmem_cache_node *n,
2213 int (*get_count)(struct page *))
2214 {
2215 unsigned long flags;
2216 unsigned long x = 0;
2217 struct page *page;
2218
2219 spin_lock_irqsave(&n->list_lock, flags);
2220 list_for_each_entry(page, &n->partial, lru)
2221 x += get_count(page);
2222 spin_unlock_irqrestore(&n->list_lock, flags);
2223 return x;
2224 }
2225 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2226
2227 static noinline void
2228 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2229 {
2230 #ifdef CONFIG_SLUB_DEBUG
2231 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2232 DEFAULT_RATELIMIT_BURST);
2233 int node;
2234 struct kmem_cache_node *n;
2235
2236 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2237 return;
2238
2239 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2240 nid, gfpflags, &gfpflags);
2241 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2242 s->name, s->object_size, s->size, oo_order(s->oo),
2243 oo_order(s->min));
2244
2245 if (oo_order(s->min) > get_order(s->object_size))
2246 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2247 s->name);
2248
2249 for_each_kmem_cache_node(s, node, n) {
2250 unsigned long nr_slabs;
2251 unsigned long nr_objs;
2252 unsigned long nr_free;
2253
2254 nr_free = count_partial(n, count_free);
2255 nr_slabs = node_nr_slabs(n);
2256 nr_objs = node_nr_objs(n);
2257
2258 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2259 node, nr_slabs, nr_objs, nr_free);
2260 }
2261 #endif
2262 }
2263
2264 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2265 int node, struct kmem_cache_cpu **pc)
2266 {
2267 void *freelist;
2268 struct kmem_cache_cpu *c = *pc;
2269 struct page *page;
2270
2271 freelist = get_partial(s, flags, node, c);
2272
2273 if (freelist)
2274 return freelist;
2275
2276 page = new_slab(s, flags, node);
2277 if (page) {
2278 c = raw_cpu_ptr(s->cpu_slab);
2279 if (c->page)
2280 flush_slab(s, c);
2281
2282 /*
2283 * No other reference to the page yet so we can
2284 * muck around with it freely without cmpxchg
2285 */
2286 freelist = page->freelist;
2287 page->freelist = NULL;
2288
2289 stat(s, ALLOC_SLAB);
2290 c->page = page;
2291 *pc = c;
2292 } else
2293 freelist = NULL;
2294
2295 return freelist;
2296 }
2297
2298 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2299 {
2300 if (unlikely(PageSlabPfmemalloc(page)))
2301 return gfp_pfmemalloc_allowed(gfpflags);
2302
2303 return true;
2304 }
2305
2306 /*
2307 * Check the page->freelist of a page and either transfer the freelist to the
2308 * per cpu freelist or deactivate the page.
2309 *
2310 * The page is still frozen if the return value is not NULL.
2311 *
2312 * If this function returns NULL then the page has been unfrozen.
2313 *
2314 * This function must be called with interrupt disabled.
2315 */
2316 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2317 {
2318 struct page new;
2319 unsigned long counters;
2320 void *freelist;
2321
2322 do {
2323 freelist = page->freelist;
2324 counters = page->counters;
2325
2326 new.counters = counters;
2327 VM_BUG_ON(!new.frozen);
2328
2329 new.inuse = page->objects;
2330 new.frozen = freelist != NULL;
2331
2332 } while (!__cmpxchg_double_slab(s, page,
2333 freelist, counters,
2334 NULL, new.counters,
2335 "get_freelist"));
2336
2337 return freelist;
2338 }
2339
2340 /*
2341 * Slow path. The lockless freelist is empty or we need to perform
2342 * debugging duties.
2343 *
2344 * Processing is still very fast if new objects have been freed to the
2345 * regular freelist. In that case we simply take over the regular freelist
2346 * as the lockless freelist and zap the regular freelist.
2347 *
2348 * If that is not working then we fall back to the partial lists. We take the
2349 * first element of the freelist as the object to allocate now and move the
2350 * rest of the freelist to the lockless freelist.
2351 *
2352 * And if we were unable to get a new slab from the partial slab lists then
2353 * we need to allocate a new slab. This is the slowest path since it involves
2354 * a call to the page allocator and the setup of a new slab.
2355 *
2356 * Version of __slab_alloc to use when we know that interrupts are
2357 * already disabled (which is the case for bulk allocation).
2358 */
2359 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2360 unsigned long addr, struct kmem_cache_cpu *c)
2361 {
2362 void *freelist;
2363 struct page *page;
2364
2365 page = c->page;
2366 if (!page)
2367 goto new_slab;
2368 redo:
2369
2370 if (unlikely(!node_match(page, node))) {
2371 int searchnode = node;
2372
2373 if (node != NUMA_NO_NODE && !node_present_pages(node))
2374 searchnode = node_to_mem_node(node);
2375
2376 if (unlikely(!node_match(page, searchnode))) {
2377 stat(s, ALLOC_NODE_MISMATCH);
2378 deactivate_slab(s, page, c->freelist);
2379 c->page = NULL;
2380 c->freelist = NULL;
2381 goto new_slab;
2382 }
2383 }
2384
2385 /*
2386 * By rights, we should be searching for a slab page that was
2387 * PFMEMALLOC but right now, we are losing the pfmemalloc
2388 * information when the page leaves the per-cpu allocator
2389 */
2390 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2391 deactivate_slab(s, page, c->freelist);
2392 c->page = NULL;
2393 c->freelist = NULL;
2394 goto new_slab;
2395 }
2396
2397 /* must check again c->freelist in case of cpu migration or IRQ */
2398 freelist = c->freelist;
2399 if (freelist)
2400 goto load_freelist;
2401
2402 freelist = get_freelist(s, page);
2403
2404 if (!freelist) {
2405 c->page = NULL;
2406 stat(s, DEACTIVATE_BYPASS);
2407 goto new_slab;
2408 }
2409
2410 stat(s, ALLOC_REFILL);
2411
2412 load_freelist:
2413 /*
2414 * freelist is pointing to the list of objects to be used.
2415 * page is pointing to the page from which the objects are obtained.
2416 * That page must be frozen for per cpu allocations to work.
2417 */
2418 VM_BUG_ON(!c->page->frozen);
2419 c->freelist = get_freepointer(s, freelist);
2420 c->tid = next_tid(c->tid);
2421 return freelist;
2422
2423 new_slab:
2424
2425 if (c->partial) {
2426 page = c->page = c->partial;
2427 c->partial = page->next;
2428 stat(s, CPU_PARTIAL_ALLOC);
2429 c->freelist = NULL;
2430 goto redo;
2431 }
2432
2433 freelist = new_slab_objects(s, gfpflags, node, &c);
2434
2435 if (unlikely(!freelist)) {
2436 slab_out_of_memory(s, gfpflags, node);
2437 return NULL;
2438 }
2439
2440 page = c->page;
2441 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2442 goto load_freelist;
2443
2444 /* Only entered in the debug case */
2445 if (kmem_cache_debug(s) &&
2446 !alloc_debug_processing(s, page, freelist, addr))
2447 goto new_slab; /* Slab failed checks. Next slab needed */
2448
2449 deactivate_slab(s, page, get_freepointer(s, freelist));
2450 c->page = NULL;
2451 c->freelist = NULL;
2452 return freelist;
2453 }
2454
2455 /*
2456 * Another one that disabled interrupt and compensates for possible
2457 * cpu changes by refetching the per cpu area pointer.
2458 */
2459 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2460 unsigned long addr, struct kmem_cache_cpu *c)
2461 {
2462 void *p;
2463 unsigned long flags;
2464
2465 local_irq_save(flags);
2466 #ifdef CONFIG_PREEMPT
2467 /*
2468 * We may have been preempted and rescheduled on a different
2469 * cpu before disabling interrupts. Need to reload cpu area
2470 * pointer.
2471 */
2472 c = this_cpu_ptr(s->cpu_slab);
2473 #endif
2474
2475 p = ___slab_alloc(s, gfpflags, node, addr, c);
2476 local_irq_restore(flags);
2477 return p;
2478 }
2479
2480 /*
2481 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2482 * have the fastpath folded into their functions. So no function call
2483 * overhead for requests that can be satisfied on the fastpath.
2484 *
2485 * The fastpath works by first checking if the lockless freelist can be used.
2486 * If not then __slab_alloc is called for slow processing.
2487 *
2488 * Otherwise we can simply pick the next object from the lockless free list.
2489 */
2490 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2491 gfp_t gfpflags, int node, unsigned long addr)
2492 {
2493 void *object;
2494 struct kmem_cache_cpu *c;
2495 struct page *page;
2496 unsigned long tid;
2497
2498 s = slab_pre_alloc_hook(s, gfpflags);
2499 if (!s)
2500 return NULL;
2501 redo:
2502 /*
2503 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2504 * enabled. We may switch back and forth between cpus while
2505 * reading from one cpu area. That does not matter as long
2506 * as we end up on the original cpu again when doing the cmpxchg.
2507 *
2508 * We should guarantee that tid and kmem_cache are retrieved on
2509 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2510 * to check if it is matched or not.
2511 */
2512 do {
2513 tid = this_cpu_read(s->cpu_slab->tid);
2514 c = raw_cpu_ptr(s->cpu_slab);
2515 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2516 unlikely(tid != READ_ONCE(c->tid)));
2517
2518 /*
2519 * Irqless object alloc/free algorithm used here depends on sequence
2520 * of fetching cpu_slab's data. tid should be fetched before anything
2521 * on c to guarantee that object and page associated with previous tid
2522 * won't be used with current tid. If we fetch tid first, object and
2523 * page could be one associated with next tid and our alloc/free
2524 * request will be failed. In this case, we will retry. So, no problem.
2525 */
2526 barrier();
2527
2528 /*
2529 * The transaction ids are globally unique per cpu and per operation on
2530 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2531 * occurs on the right processor and that there was no operation on the
2532 * linked list in between.
2533 */
2534
2535 object = c->freelist;
2536 page = c->page;
2537 if (unlikely(!object || !node_match(page, node))) {
2538 object = __slab_alloc(s, gfpflags, node, addr, c);
2539 stat(s, ALLOC_SLOWPATH);
2540 } else {
2541 void *next_object = get_freepointer_safe(s, object);
2542
2543 /*
2544 * The cmpxchg will only match if there was no additional
2545 * operation and if we are on the right processor.
2546 *
2547 * The cmpxchg does the following atomically (without lock
2548 * semantics!)
2549 * 1. Relocate first pointer to the current per cpu area.
2550 * 2. Verify that tid and freelist have not been changed
2551 * 3. If they were not changed replace tid and freelist
2552 *
2553 * Since this is without lock semantics the protection is only
2554 * against code executing on this cpu *not* from access by
2555 * other cpus.
2556 */
2557 if (unlikely(!this_cpu_cmpxchg_double(
2558 s->cpu_slab->freelist, s->cpu_slab->tid,
2559 object, tid,
2560 next_object, next_tid(tid)))) {
2561
2562 note_cmpxchg_failure("slab_alloc", s, tid);
2563 goto redo;
2564 }
2565 prefetch_freepointer(s, next_object);
2566 stat(s, ALLOC_FASTPATH);
2567 }
2568
2569 if (unlikely(gfpflags & __GFP_ZERO) && object)
2570 memset(object, 0, s->object_size);
2571
2572 slab_post_alloc_hook(s, gfpflags, 1, &object);
2573
2574 return object;
2575 }
2576
2577 static __always_inline void *slab_alloc(struct kmem_cache *s,
2578 gfp_t gfpflags, unsigned long addr)
2579 {
2580 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2581 }
2582
2583 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2584 {
2585 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2586
2587 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2588 s->size, gfpflags);
2589
2590 return ret;
2591 }
2592 EXPORT_SYMBOL(kmem_cache_alloc);
2593
2594 #ifdef CONFIG_TRACING
2595 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2596 {
2597 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2598 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2599 kasan_kmalloc(s, ret, size, gfpflags);
2600 return ret;
2601 }
2602 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2603 #endif
2604
2605 #ifdef CONFIG_NUMA
2606 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2607 {
2608 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2609
2610 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2611 s->object_size, s->size, gfpflags, node);
2612
2613 return ret;
2614 }
2615 EXPORT_SYMBOL(kmem_cache_alloc_node);
2616
2617 #ifdef CONFIG_TRACING
2618 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2619 gfp_t gfpflags,
2620 int node, size_t size)
2621 {
2622 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2623
2624 trace_kmalloc_node(_RET_IP_, ret,
2625 size, s->size, gfpflags, node);
2626
2627 kasan_kmalloc(s, ret, size, gfpflags);
2628 return ret;
2629 }
2630 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2631 #endif
2632 #endif
2633
2634 /*
2635 * Slow path handling. This may still be called frequently since objects
2636 * have a longer lifetime than the cpu slabs in most processing loads.
2637 *
2638 * So we still attempt to reduce cache line usage. Just take the slab
2639 * lock and free the item. If there is no additional partial page
2640 * handling required then we can return immediately.
2641 */
2642 static void __slab_free(struct kmem_cache *s, struct page *page,
2643 void *head, void *tail, int cnt,
2644 unsigned long addr)
2645
2646 {
2647 void *prior;
2648 int was_frozen;
2649 struct page new;
2650 unsigned long counters;
2651 struct kmem_cache_node *n = NULL;
2652 unsigned long uninitialized_var(flags);
2653
2654 stat(s, FREE_SLOWPATH);
2655
2656 if (kmem_cache_debug(s) &&
2657 !free_debug_processing(s, page, head, tail, cnt, addr))
2658 return;
2659
2660 do {
2661 if (unlikely(n)) {
2662 spin_unlock_irqrestore(&n->list_lock, flags);
2663 n = NULL;
2664 }
2665 prior = page->freelist;
2666 counters = page->counters;
2667 set_freepointer(s, tail, prior);
2668 new.counters = counters;
2669 was_frozen = new.frozen;
2670 new.inuse -= cnt;
2671 if ((!new.inuse || !prior) && !was_frozen) {
2672
2673 if (kmem_cache_has_cpu_partial(s) && !prior) {
2674
2675 /*
2676 * Slab was on no list before and will be
2677 * partially empty
2678 * We can defer the list move and instead
2679 * freeze it.
2680 */
2681 new.frozen = 1;
2682
2683 } else { /* Needs to be taken off a list */
2684
2685 n = get_node(s, page_to_nid(page));
2686 /*
2687 * Speculatively acquire the list_lock.
2688 * If the cmpxchg does not succeed then we may
2689 * drop the list_lock without any processing.
2690 *
2691 * Otherwise the list_lock will synchronize with
2692 * other processors updating the list of slabs.
2693 */
2694 spin_lock_irqsave(&n->list_lock, flags);
2695
2696 }
2697 }
2698
2699 } while (!cmpxchg_double_slab(s, page,
2700 prior, counters,
2701 head, new.counters,
2702 "__slab_free"));
2703
2704 if (likely(!n)) {
2705
2706 /*
2707 * If we just froze the page then put it onto the
2708 * per cpu partial list.
2709 */
2710 if (new.frozen && !was_frozen) {
2711 put_cpu_partial(s, page, 1);
2712 stat(s, CPU_PARTIAL_FREE);
2713 }
2714 /*
2715 * The list lock was not taken therefore no list
2716 * activity can be necessary.
2717 */
2718 if (was_frozen)
2719 stat(s, FREE_FROZEN);
2720 return;
2721 }
2722
2723 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2724 goto slab_empty;
2725
2726 /*
2727 * Objects left in the slab. If it was not on the partial list before
2728 * then add it.
2729 */
2730 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2731 if (kmem_cache_debug(s))
2732 remove_full(s, n, page);
2733 add_partial(n, page, DEACTIVATE_TO_TAIL);
2734 stat(s, FREE_ADD_PARTIAL);
2735 }
2736 spin_unlock_irqrestore(&n->list_lock, flags);
2737 return;
2738
2739 slab_empty:
2740 if (prior) {
2741 /*
2742 * Slab on the partial list.
2743 */
2744 remove_partial(n, page);
2745 stat(s, FREE_REMOVE_PARTIAL);
2746 } else {
2747 /* Slab must be on the full list */
2748 remove_full(s, n, page);
2749 }
2750
2751 spin_unlock_irqrestore(&n->list_lock, flags);
2752 stat(s, FREE_SLAB);
2753 discard_slab(s, page);
2754 }
2755
2756 /*
2757 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2758 * can perform fastpath freeing without additional function calls.
2759 *
2760 * The fastpath is only possible if we are freeing to the current cpu slab
2761 * of this processor. This typically the case if we have just allocated
2762 * the item before.
2763 *
2764 * If fastpath is not possible then fall back to __slab_free where we deal
2765 * with all sorts of special processing.
2766 *
2767 * Bulk free of a freelist with several objects (all pointing to the
2768 * same page) possible by specifying head and tail ptr, plus objects
2769 * count (cnt). Bulk free indicated by tail pointer being set.
2770 */
2771 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
2772 void *head, void *tail, int cnt,
2773 unsigned long addr)
2774 {
2775 void *tail_obj = tail ? : head;
2776 struct kmem_cache_cpu *c;
2777 unsigned long tid;
2778
2779 slab_free_freelist_hook(s, head, tail);
2780
2781 redo:
2782 /*
2783 * Determine the currently cpus per cpu slab.
2784 * The cpu may change afterward. However that does not matter since
2785 * data is retrieved via this pointer. If we are on the same cpu
2786 * during the cmpxchg then the free will succeed.
2787 */
2788 do {
2789 tid = this_cpu_read(s->cpu_slab->tid);
2790 c = raw_cpu_ptr(s->cpu_slab);
2791 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2792 unlikely(tid != READ_ONCE(c->tid)));
2793
2794 /* Same with comment on barrier() in slab_alloc_node() */
2795 barrier();
2796
2797 if (likely(page == c->page)) {
2798 set_freepointer(s, tail_obj, c->freelist);
2799
2800 if (unlikely(!this_cpu_cmpxchg_double(
2801 s->cpu_slab->freelist, s->cpu_slab->tid,
2802 c->freelist, tid,
2803 head, next_tid(tid)))) {
2804
2805 note_cmpxchg_failure("slab_free", s, tid);
2806 goto redo;
2807 }
2808 stat(s, FREE_FASTPATH);
2809 } else
2810 __slab_free(s, page, head, tail_obj, cnt, addr);
2811
2812 }
2813
2814 void kmem_cache_free(struct kmem_cache *s, void *x)
2815 {
2816 s = cache_from_obj(s, x);
2817 if (!s)
2818 return;
2819 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
2820 trace_kmem_cache_free(_RET_IP_, x);
2821 }
2822 EXPORT_SYMBOL(kmem_cache_free);
2823
2824 struct detached_freelist {
2825 struct page *page;
2826 void *tail;
2827 void *freelist;
2828 int cnt;
2829 struct kmem_cache *s;
2830 };
2831
2832 /*
2833 * This function progressively scans the array with free objects (with
2834 * a limited look ahead) and extract objects belonging to the same
2835 * page. It builds a detached freelist directly within the given
2836 * page/objects. This can happen without any need for
2837 * synchronization, because the objects are owned by running process.
2838 * The freelist is build up as a single linked list in the objects.
2839 * The idea is, that this detached freelist can then be bulk
2840 * transferred to the real freelist(s), but only requiring a single
2841 * synchronization primitive. Look ahead in the array is limited due
2842 * to performance reasons.
2843 */
2844 static inline
2845 int build_detached_freelist(struct kmem_cache *s, size_t size,
2846 void **p, struct detached_freelist *df)
2847 {
2848 size_t first_skipped_index = 0;
2849 int lookahead = 3;
2850 void *object;
2851 struct page *page;
2852
2853 /* Always re-init detached_freelist */
2854 df->page = NULL;
2855
2856 do {
2857 object = p[--size];
2858 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
2859 } while (!object && size);
2860
2861 if (!object)
2862 return 0;
2863
2864 page = virt_to_head_page(object);
2865 if (!s) {
2866 /* Handle kalloc'ed objects */
2867 if (unlikely(!PageSlab(page))) {
2868 BUG_ON(!PageCompound(page));
2869 kfree_hook(object);
2870 __free_kmem_pages(page, compound_order(page));
2871 p[size] = NULL; /* mark object processed */
2872 return size;
2873 }
2874 /* Derive kmem_cache from object */
2875 df->s = page->slab_cache;
2876 } else {
2877 df->s = cache_from_obj(s, object); /* Support for memcg */
2878 }
2879
2880 /* Start new detached freelist */
2881 df->page = page;
2882 set_freepointer(df->s, object, NULL);
2883 df->tail = object;
2884 df->freelist = object;
2885 p[size] = NULL; /* mark object processed */
2886 df->cnt = 1;
2887
2888 while (size) {
2889 object = p[--size];
2890 if (!object)
2891 continue; /* Skip processed objects */
2892
2893 /* df->page is always set at this point */
2894 if (df->page == virt_to_head_page(object)) {
2895 /* Opportunity build freelist */
2896 set_freepointer(df->s, object, df->freelist);
2897 df->freelist = object;
2898 df->cnt++;
2899 p[size] = NULL; /* mark object processed */
2900
2901 continue;
2902 }
2903
2904 /* Limit look ahead search */
2905 if (!--lookahead)
2906 break;
2907
2908 if (!first_skipped_index)
2909 first_skipped_index = size + 1;
2910 }
2911
2912 return first_skipped_index;
2913 }
2914
2915 /* Note that interrupts must be enabled when calling this function. */
2916 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
2917 {
2918 if (WARN_ON(!size))
2919 return;
2920
2921 do {
2922 struct detached_freelist df;
2923
2924 size = build_detached_freelist(s, size, p, &df);
2925 if (unlikely(!df.page))
2926 continue;
2927
2928 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
2929 } while (likely(size));
2930 }
2931 EXPORT_SYMBOL(kmem_cache_free_bulk);
2932
2933 /* Note that interrupts must be enabled when calling this function. */
2934 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
2935 void **p)
2936 {
2937 struct kmem_cache_cpu *c;
2938 int i;
2939
2940 /* memcg and kmem_cache debug support */
2941 s = slab_pre_alloc_hook(s, flags);
2942 if (unlikely(!s))
2943 return false;
2944 /*
2945 * Drain objects in the per cpu slab, while disabling local
2946 * IRQs, which protects against PREEMPT and interrupts
2947 * handlers invoking normal fastpath.
2948 */
2949 local_irq_disable();
2950 c = this_cpu_ptr(s->cpu_slab);
2951
2952 for (i = 0; i < size; i++) {
2953 void *object = c->freelist;
2954
2955 if (unlikely(!object)) {
2956 /*
2957 * Invoking slow path likely have side-effect
2958 * of re-populating per CPU c->freelist
2959 */
2960 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
2961 _RET_IP_, c);
2962 if (unlikely(!p[i]))
2963 goto error;
2964
2965 c = this_cpu_ptr(s->cpu_slab);
2966 continue; /* goto for-loop */
2967 }
2968 c->freelist = get_freepointer(s, object);
2969 p[i] = object;
2970 }
2971 c->tid = next_tid(c->tid);
2972 local_irq_enable();
2973
2974 /* Clear memory outside IRQ disabled fastpath loop */
2975 if (unlikely(flags & __GFP_ZERO)) {
2976 int j;
2977
2978 for (j = 0; j < i; j++)
2979 memset(p[j], 0, s->object_size);
2980 }
2981
2982 /* memcg and kmem_cache debug support */
2983 slab_post_alloc_hook(s, flags, size, p);
2984 return i;
2985 error:
2986 local_irq_enable();
2987 slab_post_alloc_hook(s, flags, i, p);
2988 __kmem_cache_free_bulk(s, i, p);
2989 return 0;
2990 }
2991 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
2992
2993
2994 /*
2995 * Object placement in a slab is made very easy because we always start at
2996 * offset 0. If we tune the size of the object to the alignment then we can
2997 * get the required alignment by putting one properly sized object after
2998 * another.
2999 *
3000 * Notice that the allocation order determines the sizes of the per cpu
3001 * caches. Each processor has always one slab available for allocations.
3002 * Increasing the allocation order reduces the number of times that slabs
3003 * must be moved on and off the partial lists and is therefore a factor in
3004 * locking overhead.
3005 */
3006
3007 /*
3008 * Mininum / Maximum order of slab pages. This influences locking overhead
3009 * and slab fragmentation. A higher order reduces the number of partial slabs
3010 * and increases the number of allocations possible without having to
3011 * take the list_lock.
3012 */
3013 static int slub_min_order;
3014 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3015 static int slub_min_objects;
3016
3017 /*
3018 * Calculate the order of allocation given an slab object size.
3019 *
3020 * The order of allocation has significant impact on performance and other
3021 * system components. Generally order 0 allocations should be preferred since
3022 * order 0 does not cause fragmentation in the page allocator. Larger objects
3023 * be problematic to put into order 0 slabs because there may be too much
3024 * unused space left. We go to a higher order if more than 1/16th of the slab
3025 * would be wasted.
3026 *
3027 * In order to reach satisfactory performance we must ensure that a minimum
3028 * number of objects is in one slab. Otherwise we may generate too much
3029 * activity on the partial lists which requires taking the list_lock. This is
3030 * less a concern for large slabs though which are rarely used.
3031 *
3032 * slub_max_order specifies the order where we begin to stop considering the
3033 * number of objects in a slab as critical. If we reach slub_max_order then
3034 * we try to keep the page order as low as possible. So we accept more waste
3035 * of space in favor of a small page order.
3036 *
3037 * Higher order allocations also allow the placement of more objects in a
3038 * slab and thereby reduce object handling overhead. If the user has
3039 * requested a higher mininum order then we start with that one instead of
3040 * the smallest order which will fit the object.
3041 */
3042 static inline int slab_order(int size, int min_objects,
3043 int max_order, int fract_leftover, int reserved)
3044 {
3045 int order;
3046 int rem;
3047 int min_order = slub_min_order;
3048
3049 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
3050 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3051
3052 for (order = max(min_order, get_order(min_objects * size + reserved));
3053 order <= max_order; order++) {
3054
3055 unsigned long slab_size = PAGE_SIZE << order;
3056
3057 rem = (slab_size - reserved) % size;
3058
3059 if (rem <= slab_size / fract_leftover)
3060 break;
3061 }
3062
3063 return order;
3064 }
3065
3066 static inline int calculate_order(int size, int reserved)
3067 {
3068 int order;
3069 int min_objects;
3070 int fraction;
3071 int max_objects;
3072
3073 /*
3074 * Attempt to find best configuration for a slab. This
3075 * works by first attempting to generate a layout with
3076 * the best configuration and backing off gradually.
3077 *
3078 * First we increase the acceptable waste in a slab. Then
3079 * we reduce the minimum objects required in a slab.
3080 */
3081 min_objects = slub_min_objects;
3082 if (!min_objects)
3083 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3084 max_objects = order_objects(slub_max_order, size, reserved);
3085 min_objects = min(min_objects, max_objects);
3086
3087 while (min_objects > 1) {
3088 fraction = 16;
3089 while (fraction >= 4) {
3090 order = slab_order(size, min_objects,
3091 slub_max_order, fraction, reserved);
3092 if (order <= slub_max_order)
3093 return order;
3094 fraction /= 2;
3095 }
3096 min_objects--;
3097 }
3098
3099 /*
3100 * We were unable to place multiple objects in a slab. Now
3101 * lets see if we can place a single object there.
3102 */
3103 order = slab_order(size, 1, slub_max_order, 1, reserved);
3104 if (order <= slub_max_order)
3105 return order;
3106
3107 /*
3108 * Doh this slab cannot be placed using slub_max_order.
3109 */
3110 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
3111 if (order < MAX_ORDER)
3112 return order;
3113 return -ENOSYS;
3114 }
3115
3116 static void
3117 init_kmem_cache_node(struct kmem_cache_node *n)
3118 {
3119 n->nr_partial = 0;
3120 spin_lock_init(&n->list_lock);
3121 INIT_LIST_HEAD(&n->partial);
3122 #ifdef CONFIG_SLUB_DEBUG
3123 atomic_long_set(&n->nr_slabs, 0);
3124 atomic_long_set(&n->total_objects, 0);
3125 INIT_LIST_HEAD(&n->full);
3126 #endif
3127 }
3128
3129 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3130 {
3131 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3132 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3133
3134 /*
3135 * Must align to double word boundary for the double cmpxchg
3136 * instructions to work; see __pcpu_double_call_return_bool().
3137 */
3138 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3139 2 * sizeof(void *));
3140
3141 if (!s->cpu_slab)
3142 return 0;
3143
3144 init_kmem_cache_cpus(s);
3145
3146 return 1;
3147 }
3148
3149 static struct kmem_cache *kmem_cache_node;
3150
3151 /*
3152 * No kmalloc_node yet so do it by hand. We know that this is the first
3153 * slab on the node for this slabcache. There are no concurrent accesses
3154 * possible.
3155 *
3156 * Note that this function only works on the kmem_cache_node
3157 * when allocating for the kmem_cache_node. This is used for bootstrapping
3158 * memory on a fresh node that has no slab structures yet.
3159 */
3160 static void early_kmem_cache_node_alloc(int node)
3161 {
3162 struct page *page;
3163 struct kmem_cache_node *n;
3164
3165 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3166
3167 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3168
3169 BUG_ON(!page);
3170 if (page_to_nid(page) != node) {
3171 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3172 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3173 }
3174
3175 n = page->freelist;
3176 BUG_ON(!n);
3177 page->freelist = get_freepointer(kmem_cache_node, n);
3178 page->inuse = 1;
3179 page->frozen = 0;
3180 kmem_cache_node->node[node] = n;
3181 #ifdef CONFIG_SLUB_DEBUG
3182 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3183 init_tracking(kmem_cache_node, n);
3184 #endif
3185 kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node),
3186 GFP_KERNEL);
3187 init_kmem_cache_node(n);
3188 inc_slabs_node(kmem_cache_node, node, page->objects);
3189
3190 /*
3191 * No locks need to be taken here as it has just been
3192 * initialized and there is no concurrent access.
3193 */
3194 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3195 }
3196
3197 static void free_kmem_cache_nodes(struct kmem_cache *s)
3198 {
3199 int node;
3200 struct kmem_cache_node *n;
3201
3202 for_each_kmem_cache_node(s, node, n) {
3203 kmem_cache_free(kmem_cache_node, n);
3204 s->node[node] = NULL;
3205 }
3206 }
3207
3208 void __kmem_cache_release(struct kmem_cache *s)
3209 {
3210 free_percpu(s->cpu_slab);
3211 free_kmem_cache_nodes(s);
3212 }
3213
3214 static int init_kmem_cache_nodes(struct kmem_cache *s)
3215 {
3216 int node;
3217
3218 for_each_node_state(node, N_NORMAL_MEMORY) {
3219 struct kmem_cache_node *n;
3220
3221 if (slab_state == DOWN) {
3222 early_kmem_cache_node_alloc(node);
3223 continue;
3224 }
3225 n = kmem_cache_alloc_node(kmem_cache_node,
3226 GFP_KERNEL, node);
3227
3228 if (!n) {
3229 free_kmem_cache_nodes(s);
3230 return 0;
3231 }
3232
3233 s->node[node] = n;
3234 init_kmem_cache_node(n);
3235 }
3236 return 1;
3237 }
3238
3239 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3240 {
3241 if (min < MIN_PARTIAL)
3242 min = MIN_PARTIAL;
3243 else if (min > MAX_PARTIAL)
3244 min = MAX_PARTIAL;
3245 s->min_partial = min;
3246 }
3247
3248 /*
3249 * calculate_sizes() determines the order and the distribution of data within
3250 * a slab object.
3251 */
3252 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3253 {
3254 unsigned long flags = s->flags;
3255 unsigned long size = s->object_size;
3256 int order;
3257
3258 /*
3259 * Round up object size to the next word boundary. We can only
3260 * place the free pointer at word boundaries and this determines
3261 * the possible location of the free pointer.
3262 */
3263 size = ALIGN(size, sizeof(void *));
3264
3265 #ifdef CONFIG_SLUB_DEBUG
3266 /*
3267 * Determine if we can poison the object itself. If the user of
3268 * the slab may touch the object after free or before allocation
3269 * then we should never poison the object itself.
3270 */
3271 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
3272 !s->ctor)
3273 s->flags |= __OBJECT_POISON;
3274 else
3275 s->flags &= ~__OBJECT_POISON;
3276
3277
3278 /*
3279 * If we are Redzoning then check if there is some space between the
3280 * end of the object and the free pointer. If not then add an
3281 * additional word to have some bytes to store Redzone information.
3282 */
3283 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3284 size += sizeof(void *);
3285 #endif
3286
3287 /*
3288 * With that we have determined the number of bytes in actual use
3289 * by the object. This is the potential offset to the free pointer.
3290 */
3291 s->inuse = size;
3292
3293 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
3294 s->ctor)) {
3295 /*
3296 * Relocate free pointer after the object if it is not
3297 * permitted to overwrite the first word of the object on
3298 * kmem_cache_free.
3299 *
3300 * This is the case if we do RCU, have a constructor or
3301 * destructor or are poisoning the objects.
3302 */
3303 s->offset = size;
3304 size += sizeof(void *);
3305 }
3306
3307 #ifdef CONFIG_SLUB_DEBUG
3308 if (flags & SLAB_STORE_USER)
3309 /*
3310 * Need to store information about allocs and frees after
3311 * the object.
3312 */
3313 size += 2 * sizeof(struct track);
3314
3315 if (flags & SLAB_RED_ZONE) {
3316 /*
3317 * Add some empty padding so that we can catch
3318 * overwrites from earlier objects rather than let
3319 * tracking information or the free pointer be
3320 * corrupted if a user writes before the start
3321 * of the object.
3322 */
3323 size += sizeof(void *);
3324
3325 s->red_left_pad = sizeof(void *);
3326 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3327 size += s->red_left_pad;
3328 }
3329 #endif
3330
3331 /*
3332 * SLUB stores one object immediately after another beginning from
3333 * offset 0. In order to align the objects we have to simply size
3334 * each object to conform to the alignment.
3335 */
3336 size = ALIGN(size, s->align);
3337 s->size = size;
3338 if (forced_order >= 0)
3339 order = forced_order;
3340 else
3341 order = calculate_order(size, s->reserved);
3342
3343 if (order < 0)
3344 return 0;
3345
3346 s->allocflags = 0;
3347 if (order)
3348 s->allocflags |= __GFP_COMP;
3349
3350 if (s->flags & SLAB_CACHE_DMA)
3351 s->allocflags |= GFP_DMA;
3352
3353 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3354 s->allocflags |= __GFP_RECLAIMABLE;
3355
3356 /*
3357 * Determine the number of objects per slab
3358 */
3359 s->oo = oo_make(order, size, s->reserved);
3360 s->min = oo_make(get_order(size), size, s->reserved);
3361 if (oo_objects(s->oo) > oo_objects(s->max))
3362 s->max = s->oo;
3363
3364 return !!oo_objects(s->oo);
3365 }
3366
3367 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3368 {
3369 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3370 s->reserved = 0;
3371
3372 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3373 s->reserved = sizeof(struct rcu_head);
3374
3375 if (!calculate_sizes(s, -1))
3376 goto error;
3377 if (disable_higher_order_debug) {
3378 /*
3379 * Disable debugging flags that store metadata if the min slab
3380 * order increased.
3381 */
3382 if (get_order(s->size) > get_order(s->object_size)) {
3383 s->flags &= ~DEBUG_METADATA_FLAGS;
3384 s->offset = 0;
3385 if (!calculate_sizes(s, -1))
3386 goto error;
3387 }
3388 }
3389
3390 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3391 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3392 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
3393 /* Enable fast mode */
3394 s->flags |= __CMPXCHG_DOUBLE;
3395 #endif
3396
3397 /*
3398 * The larger the object size is, the more pages we want on the partial
3399 * list to avoid pounding the page allocator excessively.
3400 */
3401 set_min_partial(s, ilog2(s->size) / 2);
3402
3403 /*
3404 * cpu_partial determined the maximum number of objects kept in the
3405 * per cpu partial lists of a processor.
3406 *
3407 * Per cpu partial lists mainly contain slabs that just have one
3408 * object freed. If they are used for allocation then they can be
3409 * filled up again with minimal effort. The slab will never hit the
3410 * per node partial lists and therefore no locking will be required.
3411 *
3412 * This setting also determines
3413 *
3414 * A) The number of objects from per cpu partial slabs dumped to the
3415 * per node list when we reach the limit.
3416 * B) The number of objects in cpu partial slabs to extract from the
3417 * per node list when we run out of per cpu objects. We only fetch
3418 * 50% to keep some capacity around for frees.
3419 */
3420 if (!kmem_cache_has_cpu_partial(s))
3421 s->cpu_partial = 0;
3422 else if (s->size >= PAGE_SIZE)
3423 s->cpu_partial = 2;
3424 else if (s->size >= 1024)
3425 s->cpu_partial = 6;
3426 else if (s->size >= 256)
3427 s->cpu_partial = 13;
3428 else
3429 s->cpu_partial = 30;
3430
3431 #ifdef CONFIG_NUMA
3432 s->remote_node_defrag_ratio = 1000;
3433 #endif
3434 if (!init_kmem_cache_nodes(s))
3435 goto error;
3436
3437 if (alloc_kmem_cache_cpus(s))
3438 return 0;
3439
3440 free_kmem_cache_nodes(s);
3441 error:
3442 if (flags & SLAB_PANIC)
3443 panic("Cannot create slab %s size=%lu realsize=%u order=%u offset=%u flags=%lx\n",
3444 s->name, (unsigned long)s->size, s->size,
3445 oo_order(s->oo), s->offset, flags);
3446 return -EINVAL;
3447 }
3448
3449 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3450 const char *text)
3451 {
3452 #ifdef CONFIG_SLUB_DEBUG
3453 void *addr = page_address(page);
3454 void *p;
3455 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3456 sizeof(long), GFP_ATOMIC);
3457 if (!map)
3458 return;
3459 slab_err(s, page, text, s->name);
3460 slab_lock(page);
3461
3462 get_map(s, page, map);
3463 for_each_object(p, s, addr, page->objects) {
3464
3465 if (!test_bit(slab_index(p, s, addr), map)) {
3466 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3467 print_tracking(s, p);
3468 }
3469 }
3470 slab_unlock(page);
3471 kfree(map);
3472 #endif
3473 }
3474
3475 /*
3476 * Attempt to free all partial slabs on a node.
3477 * This is called from __kmem_cache_shutdown(). We must take list_lock
3478 * because sysfs file might still access partial list after the shutdowning.
3479 */
3480 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3481 {
3482 struct page *page, *h;
3483
3484 BUG_ON(irqs_disabled());
3485 spin_lock_irq(&n->list_lock);
3486 list_for_each_entry_safe(page, h, &n->partial, lru) {
3487 if (!page->inuse) {
3488 remove_partial(n, page);
3489 discard_slab(s, page);
3490 } else {
3491 list_slab_objects(s, page,
3492 "Objects remaining in %s on __kmem_cache_shutdown()");
3493 }
3494 }
3495 spin_unlock_irq(&n->list_lock);
3496 }
3497
3498 /*
3499 * Release all resources used by a slab cache.
3500 */
3501 int __kmem_cache_shutdown(struct kmem_cache *s)
3502 {
3503 int node;
3504 struct kmem_cache_node *n;
3505
3506 flush_all(s);
3507 /* Attempt to free all objects */
3508 for_each_kmem_cache_node(s, node, n) {
3509 free_partial(s, n);
3510 if (n->nr_partial || slabs_node(s, node))
3511 return 1;
3512 }
3513 return 0;
3514 }
3515
3516 /********************************************************************
3517 * Kmalloc subsystem
3518 *******************************************************************/
3519
3520 static int __init setup_slub_min_order(char *str)
3521 {
3522 get_option(&str, &slub_min_order);
3523
3524 return 1;
3525 }
3526
3527 __setup("slub_min_order=", setup_slub_min_order);
3528
3529 static int __init setup_slub_max_order(char *str)
3530 {
3531 get_option(&str, &slub_max_order);
3532 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3533
3534 return 1;
3535 }
3536
3537 __setup("slub_max_order=", setup_slub_max_order);
3538
3539 static int __init setup_slub_min_objects(char *str)
3540 {
3541 get_option(&str, &slub_min_objects);
3542
3543 return 1;
3544 }
3545
3546 __setup("slub_min_objects=", setup_slub_min_objects);
3547
3548 void *__kmalloc(size_t size, gfp_t flags)
3549 {
3550 struct kmem_cache *s;
3551 void *ret;
3552
3553 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3554 return kmalloc_large(size, flags);
3555
3556 s = kmalloc_slab(size, flags);
3557
3558 if (unlikely(ZERO_OR_NULL_PTR(s)))
3559 return s;
3560
3561 ret = slab_alloc(s, flags, _RET_IP_);
3562
3563 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3564
3565 kasan_kmalloc(s, ret, size, flags);
3566
3567 return ret;
3568 }
3569 EXPORT_SYMBOL(__kmalloc);
3570
3571 #ifdef CONFIG_NUMA
3572 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3573 {
3574 struct page *page;
3575 void *ptr = NULL;
3576
3577 flags |= __GFP_COMP | __GFP_NOTRACK;
3578 page = alloc_kmem_pages_node(node, flags, get_order(size));
3579 if (page)
3580 ptr = page_address(page);
3581
3582 kmalloc_large_node_hook(ptr, size, flags);
3583 return ptr;
3584 }
3585
3586 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3587 {
3588 struct kmem_cache *s;
3589 void *ret;
3590
3591 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3592 ret = kmalloc_large_node(size, flags, node);
3593
3594 trace_kmalloc_node(_RET_IP_, ret,
3595 size, PAGE_SIZE << get_order(size),
3596 flags, node);
3597
3598 return ret;
3599 }
3600
3601 s = kmalloc_slab(size, flags);
3602
3603 if (unlikely(ZERO_OR_NULL_PTR(s)))
3604 return s;
3605
3606 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3607
3608 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3609
3610 kasan_kmalloc(s, ret, size, flags);
3611
3612 return ret;
3613 }
3614 EXPORT_SYMBOL(__kmalloc_node);
3615 #endif
3616
3617 static size_t __ksize(const void *object)
3618 {
3619 struct page *page;
3620
3621 if (unlikely(object == ZERO_SIZE_PTR))
3622 return 0;
3623
3624 page = virt_to_head_page(object);
3625
3626 if (unlikely(!PageSlab(page))) {
3627 WARN_ON(!PageCompound(page));
3628 return PAGE_SIZE << compound_order(page);
3629 }
3630
3631 return slab_ksize(page->slab_cache);
3632 }
3633
3634 size_t ksize(const void *object)
3635 {
3636 size_t size = __ksize(object);
3637 /* We assume that ksize callers could use whole allocated area,
3638 * so we need to unpoison this area.
3639 */
3640 kasan_unpoison_shadow(object, size);
3641 return size;
3642 }
3643 EXPORT_SYMBOL(ksize);
3644
3645 void kfree(const void *x)
3646 {
3647 struct page *page;
3648 void *object = (void *)x;
3649
3650 trace_kfree(_RET_IP_, x);
3651
3652 if (unlikely(ZERO_OR_NULL_PTR(x)))
3653 return;
3654
3655 page = virt_to_head_page(x);
3656 if (unlikely(!PageSlab(page))) {
3657 BUG_ON(!PageCompound(page));
3658 kfree_hook(x);
3659 __free_kmem_pages(page, compound_order(page));
3660 return;
3661 }
3662 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
3663 }
3664 EXPORT_SYMBOL(kfree);
3665
3666 #define SHRINK_PROMOTE_MAX 32
3667
3668 /*
3669 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3670 * up most to the head of the partial lists. New allocations will then
3671 * fill those up and thus they can be removed from the partial lists.
3672 *
3673 * The slabs with the least items are placed last. This results in them
3674 * being allocated from last increasing the chance that the last objects
3675 * are freed in them.
3676 */
3677 int __kmem_cache_shrink(struct kmem_cache *s, bool deactivate)
3678 {
3679 int node;
3680 int i;
3681 struct kmem_cache_node *n;
3682 struct page *page;
3683 struct page *t;
3684 struct list_head discard;
3685 struct list_head promote[SHRINK_PROMOTE_MAX];
3686 unsigned long flags;
3687 int ret = 0;
3688
3689 if (deactivate) {
3690 /*
3691 * Disable empty slabs caching. Used to avoid pinning offline
3692 * memory cgroups by kmem pages that can be freed.
3693 */
3694 s->cpu_partial = 0;
3695 s->min_partial = 0;
3696
3697 /*
3698 * s->cpu_partial is checked locklessly (see put_cpu_partial),
3699 * so we have to make sure the change is visible.
3700 */
3701 synchronize_sched();
3702 }
3703
3704 flush_all(s);
3705 for_each_kmem_cache_node(s, node, n) {
3706 INIT_LIST_HEAD(&discard);
3707 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3708 INIT_LIST_HEAD(promote + i);
3709
3710 spin_lock_irqsave(&n->list_lock, flags);
3711
3712 /*
3713 * Build lists of slabs to discard or promote.
3714 *
3715 * Note that concurrent frees may occur while we hold the
3716 * list_lock. page->inuse here is the upper limit.
3717 */
3718 list_for_each_entry_safe(page, t, &n->partial, lru) {
3719 int free = page->objects - page->inuse;
3720
3721 /* Do not reread page->inuse */
3722 barrier();
3723
3724 /* We do not keep full slabs on the list */
3725 BUG_ON(free <= 0);
3726
3727 if (free == page->objects) {
3728 list_move(&page->lru, &discard);
3729 n->nr_partial--;
3730 } else if (free <= SHRINK_PROMOTE_MAX)
3731 list_move(&page->lru, promote + free - 1);
3732 }
3733
3734 /*
3735 * Promote the slabs filled up most to the head of the
3736 * partial list.
3737 */
3738 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
3739 list_splice(promote + i, &n->partial);
3740
3741 spin_unlock_irqrestore(&n->list_lock, flags);
3742
3743 /* Release empty slabs */
3744 list_for_each_entry_safe(page, t, &discard, lru)
3745 discard_slab(s, page);
3746
3747 if (slabs_node(s, node))
3748 ret = 1;
3749 }
3750
3751 return ret;
3752 }
3753
3754 static int slab_mem_going_offline_callback(void *arg)
3755 {
3756 struct kmem_cache *s;
3757
3758 mutex_lock(&slab_mutex);
3759 list_for_each_entry(s, &slab_caches, list)
3760 __kmem_cache_shrink(s, false);
3761 mutex_unlock(&slab_mutex);
3762
3763 return 0;
3764 }
3765
3766 static void slab_mem_offline_callback(void *arg)
3767 {
3768 struct kmem_cache_node *n;
3769 struct kmem_cache *s;
3770 struct memory_notify *marg = arg;
3771 int offline_node;
3772
3773 offline_node = marg->status_change_nid_normal;
3774
3775 /*
3776 * If the node still has available memory. we need kmem_cache_node
3777 * for it yet.
3778 */
3779 if (offline_node < 0)
3780 return;
3781
3782 mutex_lock(&slab_mutex);
3783 list_for_each_entry(s, &slab_caches, list) {
3784 n = get_node(s, offline_node);
3785 if (n) {
3786 /*
3787 * if n->nr_slabs > 0, slabs still exist on the node
3788 * that is going down. We were unable to free them,
3789 * and offline_pages() function shouldn't call this
3790 * callback. So, we must fail.
3791 */
3792 BUG_ON(slabs_node(s, offline_node));
3793
3794 s->node[offline_node] = NULL;
3795 kmem_cache_free(kmem_cache_node, n);
3796 }
3797 }
3798 mutex_unlock(&slab_mutex);
3799 }
3800
3801 static int slab_mem_going_online_callback(void *arg)
3802 {
3803 struct kmem_cache_node *n;
3804 struct kmem_cache *s;
3805 struct memory_notify *marg = arg;
3806 int nid = marg->status_change_nid_normal;
3807 int ret = 0;
3808
3809 /*
3810 * If the node's memory is already available, then kmem_cache_node is
3811 * already created. Nothing to do.
3812 */
3813 if (nid < 0)
3814 return 0;
3815
3816 /*
3817 * We are bringing a node online. No memory is available yet. We must
3818 * allocate a kmem_cache_node structure in order to bring the node
3819 * online.
3820 */
3821 mutex_lock(&slab_mutex);
3822 list_for_each_entry(s, &slab_caches, list) {
3823 /*
3824 * XXX: kmem_cache_alloc_node will fallback to other nodes
3825 * since memory is not yet available from the node that
3826 * is brought up.
3827 */
3828 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3829 if (!n) {
3830 ret = -ENOMEM;
3831 goto out;
3832 }
3833 init_kmem_cache_node(n);
3834 s->node[nid] = n;
3835 }
3836 out:
3837 mutex_unlock(&slab_mutex);
3838 return ret;
3839 }
3840
3841 static int slab_memory_callback(struct notifier_block *self,
3842 unsigned long action, void *arg)
3843 {
3844 int ret = 0;
3845
3846 switch (action) {
3847 case MEM_GOING_ONLINE:
3848 ret = slab_mem_going_online_callback(arg);
3849 break;
3850 case MEM_GOING_OFFLINE:
3851 ret = slab_mem_going_offline_callback(arg);
3852 break;
3853 case MEM_OFFLINE:
3854 case MEM_CANCEL_ONLINE:
3855 slab_mem_offline_callback(arg);
3856 break;
3857 case MEM_ONLINE:
3858 case MEM_CANCEL_OFFLINE:
3859 break;
3860 }
3861 if (ret)
3862 ret = notifier_from_errno(ret);
3863 else
3864 ret = NOTIFY_OK;
3865 return ret;
3866 }
3867
3868 static struct notifier_block slab_memory_callback_nb = {
3869 .notifier_call = slab_memory_callback,
3870 .priority = SLAB_CALLBACK_PRI,
3871 };
3872
3873 /********************************************************************
3874 * Basic setup of slabs
3875 *******************************************************************/
3876
3877 /*
3878 * Used for early kmem_cache structures that were allocated using
3879 * the page allocator. Allocate them properly then fix up the pointers
3880 * that may be pointing to the wrong kmem_cache structure.
3881 */
3882
3883 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3884 {
3885 int node;
3886 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3887 struct kmem_cache_node *n;
3888
3889 memcpy(s, static_cache, kmem_cache->object_size);
3890
3891 /*
3892 * This runs very early, and only the boot processor is supposed to be
3893 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3894 * IPIs around.
3895 */
3896 __flush_cpu_slab(s, smp_processor_id());
3897 for_each_kmem_cache_node(s, node, n) {
3898 struct page *p;
3899
3900 list_for_each_entry(p, &n->partial, lru)
3901 p->slab_cache = s;
3902
3903 #ifdef CONFIG_SLUB_DEBUG
3904 list_for_each_entry(p, &n->full, lru)
3905 p->slab_cache = s;
3906 #endif
3907 }
3908 slab_init_memcg_params(s);
3909 list_add(&s->list, &slab_caches);
3910 return s;
3911 }
3912
3913 void __init kmem_cache_init(void)
3914 {
3915 static __initdata struct kmem_cache boot_kmem_cache,
3916 boot_kmem_cache_node;
3917
3918 if (debug_guardpage_minorder())
3919 slub_max_order = 0;
3920
3921 kmem_cache_node = &boot_kmem_cache_node;
3922 kmem_cache = &boot_kmem_cache;
3923
3924 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3925 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3926
3927 register_hotmemory_notifier(&slab_memory_callback_nb);
3928
3929 /* Able to allocate the per node structures */
3930 slab_state = PARTIAL;
3931
3932 create_boot_cache(kmem_cache, "kmem_cache",
3933 offsetof(struct kmem_cache, node) +
3934 nr_node_ids * sizeof(struct kmem_cache_node *),
3935 SLAB_HWCACHE_ALIGN);
3936
3937 kmem_cache = bootstrap(&boot_kmem_cache);
3938
3939 /*
3940 * Allocate kmem_cache_node properly from the kmem_cache slab.
3941 * kmem_cache_node is separately allocated so no need to
3942 * update any list pointers.
3943 */
3944 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3945
3946 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3947 setup_kmalloc_cache_index_table();
3948 create_kmalloc_caches(0);
3949
3950 #ifdef CONFIG_SMP
3951 register_cpu_notifier(&slab_notifier);
3952 #endif
3953
3954 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
3955 cache_line_size(),
3956 slub_min_order, slub_max_order, slub_min_objects,
3957 nr_cpu_ids, nr_node_ids);
3958 }
3959
3960 void __init kmem_cache_init_late(void)
3961 {
3962 }
3963
3964 struct kmem_cache *
3965 __kmem_cache_alias(const char *name, size_t size, size_t align,
3966 unsigned long flags, void (*ctor)(void *))
3967 {
3968 struct kmem_cache *s, *c;
3969
3970 s = find_mergeable(size, align, flags, name, ctor);
3971 if (s) {
3972 s->refcount++;
3973
3974 /*
3975 * Adjust the object sizes so that we clear
3976 * the complete object on kzalloc.
3977 */
3978 s->object_size = max(s->object_size, (int)size);
3979 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3980
3981 for_each_memcg_cache(c, s) {
3982 c->object_size = s->object_size;
3983 c->inuse = max_t(int, c->inuse,
3984 ALIGN(size, sizeof(void *)));
3985 }
3986
3987 if (sysfs_slab_alias(s, name)) {
3988 s->refcount--;
3989 s = NULL;
3990 }
3991 }
3992
3993 return s;
3994 }
3995
3996 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3997 {
3998 int err;
3999
4000 err = kmem_cache_open(s, flags);
4001 if (err)
4002 return err;
4003
4004 /* Mutex is not taken during early boot */
4005 if (slab_state <= UP)
4006 return 0;
4007
4008 memcg_propagate_slab_attrs(s);
4009 err = sysfs_slab_add(s);
4010 if (err)
4011 __kmem_cache_release(s);
4012
4013 return err;
4014 }
4015
4016 #ifdef CONFIG_SMP
4017 /*
4018 * Use the cpu notifier to insure that the cpu slabs are flushed when
4019 * necessary.
4020 */
4021 static int slab_cpuup_callback(struct notifier_block *nfb,
4022 unsigned long action, void *hcpu)
4023 {
4024 long cpu = (long)hcpu;
4025 struct kmem_cache *s;
4026 unsigned long flags;
4027
4028 switch (action) {
4029 case CPU_UP_CANCELED:
4030 case CPU_UP_CANCELED_FROZEN:
4031 case CPU_DEAD:
4032 case CPU_DEAD_FROZEN:
4033 mutex_lock(&slab_mutex);
4034 list_for_each_entry(s, &slab_caches, list) {
4035 local_irq_save(flags);
4036 __flush_cpu_slab(s, cpu);
4037 local_irq_restore(flags);
4038 }
4039 mutex_unlock(&slab_mutex);
4040 break;
4041 default:
4042 break;
4043 }
4044 return NOTIFY_OK;
4045 }
4046
4047 static struct notifier_block slab_notifier = {
4048 .notifier_call = slab_cpuup_callback
4049 };
4050
4051 #endif
4052
4053 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4054 {
4055 struct kmem_cache *s;
4056 void *ret;
4057
4058 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4059 return kmalloc_large(size, gfpflags);
4060
4061 s = kmalloc_slab(size, gfpflags);
4062
4063 if (unlikely(ZERO_OR_NULL_PTR(s)))
4064 return s;
4065
4066 ret = slab_alloc(s, gfpflags, caller);
4067
4068 /* Honor the call site pointer we received. */
4069 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4070
4071 return ret;
4072 }
4073
4074 #ifdef CONFIG_NUMA
4075 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4076 int node, unsigned long caller)
4077 {
4078 struct kmem_cache *s;
4079 void *ret;
4080
4081 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4082 ret = kmalloc_large_node(size, gfpflags, node);
4083
4084 trace_kmalloc_node(caller, ret,
4085 size, PAGE_SIZE << get_order(size),
4086 gfpflags, node);
4087
4088 return ret;
4089 }
4090
4091 s = kmalloc_slab(size, gfpflags);
4092
4093 if (unlikely(ZERO_OR_NULL_PTR(s)))
4094 return s;
4095
4096 ret = slab_alloc_node(s, gfpflags, node, caller);
4097
4098 /* Honor the call site pointer we received. */
4099 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4100
4101 return ret;
4102 }
4103 #endif
4104
4105 #ifdef CONFIG_SYSFS
4106 static int count_inuse(struct page *page)
4107 {
4108 return page->inuse;
4109 }
4110
4111 static int count_total(struct page *page)
4112 {
4113 return page->objects;
4114 }
4115 #endif
4116
4117 #ifdef CONFIG_SLUB_DEBUG
4118 static int validate_slab(struct kmem_cache *s, struct page *page,
4119 unsigned long *map)
4120 {
4121 void *p;
4122 void *addr = page_address(page);
4123
4124 if (!check_slab(s, page) ||
4125 !on_freelist(s, page, NULL))
4126 return 0;
4127
4128 /* Now we know that a valid freelist exists */
4129 bitmap_zero(map, page->objects);
4130
4131 get_map(s, page, map);
4132 for_each_object(p, s, addr, page->objects) {
4133 if (test_bit(slab_index(p, s, addr), map))
4134 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4135 return 0;
4136 }
4137
4138 for_each_object(p, s, addr, page->objects)
4139 if (!test_bit(slab_index(p, s, addr), map))
4140 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4141 return 0;
4142 return 1;
4143 }
4144
4145 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4146 unsigned long *map)
4147 {
4148 slab_lock(page);
4149 validate_slab(s, page, map);
4150 slab_unlock(page);
4151 }
4152
4153 static int validate_slab_node(struct kmem_cache *s,
4154 struct kmem_cache_node *n, unsigned long *map)
4155 {
4156 unsigned long count = 0;
4157 struct page *page;
4158 unsigned long flags;
4159
4160 spin_lock_irqsave(&n->list_lock, flags);
4161
4162 list_for_each_entry(page, &n->partial, lru) {
4163 validate_slab_slab(s, page, map);
4164 count++;
4165 }
4166 if (count != n->nr_partial)
4167 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4168 s->name, count, n->nr_partial);
4169
4170 if (!(s->flags & SLAB_STORE_USER))
4171 goto out;
4172
4173 list_for_each_entry(page, &n->full, lru) {
4174 validate_slab_slab(s, page, map);
4175 count++;
4176 }
4177 if (count != atomic_long_read(&n->nr_slabs))
4178 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4179 s->name, count, atomic_long_read(&n->nr_slabs));
4180
4181 out:
4182 spin_unlock_irqrestore(&n->list_lock, flags);
4183 return count;
4184 }
4185
4186 static long validate_slab_cache(struct kmem_cache *s)
4187 {
4188 int node;
4189 unsigned long count = 0;
4190 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4191 sizeof(unsigned long), GFP_KERNEL);
4192 struct kmem_cache_node *n;
4193
4194 if (!map)
4195 return -ENOMEM;
4196
4197 flush_all(s);
4198 for_each_kmem_cache_node(s, node, n)
4199 count += validate_slab_node(s, n, map);
4200 kfree(map);
4201 return count;
4202 }
4203 /*
4204 * Generate lists of code addresses where slabcache objects are allocated
4205 * and freed.
4206 */
4207
4208 struct location {
4209 unsigned long count;
4210 unsigned long addr;
4211 long long sum_time;
4212 long min_time;
4213 long max_time;
4214 long min_pid;
4215 long max_pid;
4216 DECLARE_BITMAP(cpus, NR_CPUS);
4217 nodemask_t nodes;
4218 };
4219
4220 struct loc_track {
4221 unsigned long max;
4222 unsigned long count;
4223 struct location *loc;
4224 };
4225
4226 static void free_loc_track(struct loc_track *t)
4227 {
4228 if (t->max)
4229 free_pages((unsigned long)t->loc,
4230 get_order(sizeof(struct location) * t->max));
4231 }
4232
4233 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4234 {
4235 struct location *l;
4236 int order;
4237
4238 order = get_order(sizeof(struct location) * max);
4239
4240 l = (void *)__get_free_pages(flags, order);
4241 if (!l)
4242 return 0;
4243
4244 if (t->count) {
4245 memcpy(l, t->loc, sizeof(struct location) * t->count);
4246 free_loc_track(t);
4247 }
4248 t->max = max;
4249 t->loc = l;
4250 return 1;
4251 }
4252
4253 static int add_location(struct loc_track *t, struct kmem_cache *s,
4254 const struct track *track)
4255 {
4256 long start, end, pos;
4257 struct location *l;
4258 unsigned long caddr;
4259 unsigned long age = jiffies - track->when;
4260
4261 start = -1;
4262 end = t->count;
4263
4264 for ( ; ; ) {
4265 pos = start + (end - start + 1) / 2;
4266
4267 /*
4268 * There is nothing at "end". If we end up there
4269 * we need to add something to before end.
4270 */
4271 if (pos == end)
4272 break;
4273
4274 caddr = t->loc[pos].addr;
4275 if (track->addr == caddr) {
4276
4277 l = &t->loc[pos];
4278 l->count++;
4279 if (track->when) {
4280 l->sum_time += age;
4281 if (age < l->min_time)
4282 l->min_time = age;
4283 if (age > l->max_time)
4284 l->max_time = age;
4285
4286 if (track->pid < l->min_pid)
4287 l->min_pid = track->pid;
4288 if (track->pid > l->max_pid)
4289 l->max_pid = track->pid;
4290
4291 cpumask_set_cpu(track->cpu,
4292 to_cpumask(l->cpus));
4293 }
4294 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4295 return 1;
4296 }
4297
4298 if (track->addr < caddr)
4299 end = pos;
4300 else
4301 start = pos;
4302 }
4303
4304 /*
4305 * Not found. Insert new tracking element.
4306 */
4307 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4308 return 0;
4309
4310 l = t->loc + pos;
4311 if (pos < t->count)
4312 memmove(l + 1, l,
4313 (t->count - pos) * sizeof(struct location));
4314 t->count++;
4315 l->count = 1;
4316 l->addr = track->addr;
4317 l->sum_time = age;
4318 l->min_time = age;
4319 l->max_time = age;
4320 l->min_pid = track->pid;
4321 l->max_pid = track->pid;
4322 cpumask_clear(to_cpumask(l->cpus));
4323 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4324 nodes_clear(l->nodes);
4325 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4326 return 1;
4327 }
4328
4329 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4330 struct page *page, enum track_item alloc,
4331 unsigned long *map)
4332 {
4333 void *addr = page_address(page);
4334 void *p;
4335
4336 bitmap_zero(map, page->objects);
4337 get_map(s, page, map);
4338
4339 for_each_object(p, s, addr, page->objects)
4340 if (!test_bit(slab_index(p, s, addr), map))
4341 add_location(t, s, get_track(s, p, alloc));
4342 }
4343
4344 static int list_locations(struct kmem_cache *s, char *buf,
4345 enum track_item alloc)
4346 {
4347 int len = 0;
4348 unsigned long i;
4349 struct loc_track t = { 0, 0, NULL };
4350 int node;
4351 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4352 sizeof(unsigned long), GFP_KERNEL);
4353 struct kmem_cache_node *n;
4354
4355 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4356 GFP_TEMPORARY)) {
4357 kfree(map);
4358 return sprintf(buf, "Out of memory\n");
4359 }
4360 /* Push back cpu slabs */
4361 flush_all(s);
4362
4363 for_each_kmem_cache_node(s, node, n) {
4364 unsigned long flags;
4365 struct page *page;
4366
4367 if (!atomic_long_read(&n->nr_slabs))
4368 continue;
4369
4370 spin_lock_irqsave(&n->list_lock, flags);
4371 list_for_each_entry(page, &n->partial, lru)
4372 process_slab(&t, s, page, alloc, map);
4373 list_for_each_entry(page, &n->full, lru)
4374 process_slab(&t, s, page, alloc, map);
4375 spin_unlock_irqrestore(&n->list_lock, flags);
4376 }
4377
4378 for (i = 0; i < t.count; i++) {
4379 struct location *l = &t.loc[i];
4380
4381 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4382 break;
4383 len += sprintf(buf + len, "%7ld ", l->count);
4384
4385 if (l->addr)
4386 len += sprintf(buf + len, "%pS", (void *)l->addr);
4387 else
4388 len += sprintf(buf + len, "<not-available>");
4389
4390 if (l->sum_time != l->min_time) {
4391 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4392 l->min_time,
4393 (long)div_u64(l->sum_time, l->count),
4394 l->max_time);
4395 } else
4396 len += sprintf(buf + len, " age=%ld",
4397 l->min_time);
4398
4399 if (l->min_pid != l->max_pid)
4400 len += sprintf(buf + len, " pid=%ld-%ld",
4401 l->min_pid, l->max_pid);
4402 else
4403 len += sprintf(buf + len, " pid=%ld",
4404 l->min_pid);
4405
4406 if (num_online_cpus() > 1 &&
4407 !cpumask_empty(to_cpumask(l->cpus)) &&
4408 len < PAGE_SIZE - 60)
4409 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4410 " cpus=%*pbl",
4411 cpumask_pr_args(to_cpumask(l->cpus)));
4412
4413 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4414 len < PAGE_SIZE - 60)
4415 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4416 " nodes=%*pbl",
4417 nodemask_pr_args(&l->nodes));
4418
4419 len += sprintf(buf + len, "\n");
4420 }
4421
4422 free_loc_track(&t);
4423 kfree(map);
4424 if (!t.count)
4425 len += sprintf(buf, "No data\n");
4426 return len;
4427 }
4428 #endif
4429
4430 #ifdef SLUB_RESILIENCY_TEST
4431 static void __init resiliency_test(void)
4432 {
4433 u8 *p;
4434
4435 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4436
4437 pr_err("SLUB resiliency testing\n");
4438 pr_err("-----------------------\n");
4439 pr_err("A. Corruption after allocation\n");
4440
4441 p = kzalloc(16, GFP_KERNEL);
4442 p[16] = 0x12;
4443 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4444 p + 16);
4445
4446 validate_slab_cache(kmalloc_caches[4]);
4447
4448 /* Hmmm... The next two are dangerous */
4449 p = kzalloc(32, GFP_KERNEL);
4450 p[32 + sizeof(void *)] = 0x34;
4451 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4452 p);
4453 pr_err("If allocated object is overwritten then not detectable\n\n");
4454
4455 validate_slab_cache(kmalloc_caches[5]);
4456 p = kzalloc(64, GFP_KERNEL);
4457 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4458 *p = 0x56;
4459 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4460 p);
4461 pr_err("If allocated object is overwritten then not detectable\n\n");
4462 validate_slab_cache(kmalloc_caches[6]);
4463
4464 pr_err("\nB. Corruption after free\n");
4465 p = kzalloc(128, GFP_KERNEL);
4466 kfree(p);
4467 *p = 0x78;
4468 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4469 validate_slab_cache(kmalloc_caches[7]);
4470
4471 p = kzalloc(256, GFP_KERNEL);
4472 kfree(p);
4473 p[50] = 0x9a;
4474 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4475 validate_slab_cache(kmalloc_caches[8]);
4476
4477 p = kzalloc(512, GFP_KERNEL);
4478 kfree(p);
4479 p[512] = 0xab;
4480 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4481 validate_slab_cache(kmalloc_caches[9]);
4482 }
4483 #else
4484 #ifdef CONFIG_SYSFS
4485 static void resiliency_test(void) {};
4486 #endif
4487 #endif
4488
4489 #ifdef CONFIG_SYSFS
4490 enum slab_stat_type {
4491 SL_ALL, /* All slabs */
4492 SL_PARTIAL, /* Only partially allocated slabs */
4493 SL_CPU, /* Only slabs used for cpu caches */
4494 SL_OBJECTS, /* Determine allocated objects not slabs */
4495 SL_TOTAL /* Determine object capacity not slabs */
4496 };
4497
4498 #define SO_ALL (1 << SL_ALL)
4499 #define SO_PARTIAL (1 << SL_PARTIAL)
4500 #define SO_CPU (1 << SL_CPU)
4501 #define SO_OBJECTS (1 << SL_OBJECTS)
4502 #define SO_TOTAL (1 << SL_TOTAL)
4503
4504 static ssize_t show_slab_objects(struct kmem_cache *s,
4505 char *buf, unsigned long flags)
4506 {
4507 unsigned long total = 0;
4508 int node;
4509 int x;
4510 unsigned long *nodes;
4511
4512 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4513 if (!nodes)
4514 return -ENOMEM;
4515
4516 if (flags & SO_CPU) {
4517 int cpu;
4518
4519 for_each_possible_cpu(cpu) {
4520 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4521 cpu);
4522 int node;
4523 struct page *page;
4524
4525 page = READ_ONCE(c->page);
4526 if (!page)
4527 continue;
4528
4529 node = page_to_nid(page);
4530 if (flags & SO_TOTAL)
4531 x = page->objects;
4532 else if (flags & SO_OBJECTS)
4533 x = page->inuse;
4534 else
4535 x = 1;
4536
4537 total += x;
4538 nodes[node] += x;
4539
4540 page = READ_ONCE(c->partial);
4541 if (page) {
4542 node = page_to_nid(page);
4543 if (flags & SO_TOTAL)
4544 WARN_ON_ONCE(1);
4545 else if (flags & SO_OBJECTS)
4546 WARN_ON_ONCE(1);
4547 else
4548 x = page->pages;
4549 total += x;
4550 nodes[node] += x;
4551 }
4552 }
4553 }
4554
4555 get_online_mems();
4556 #ifdef CONFIG_SLUB_DEBUG
4557 if (flags & SO_ALL) {
4558 struct kmem_cache_node *n;
4559
4560 for_each_kmem_cache_node(s, node, n) {
4561
4562 if (flags & SO_TOTAL)
4563 x = atomic_long_read(&n->total_objects);
4564 else if (flags & SO_OBJECTS)
4565 x = atomic_long_read(&n->total_objects) -
4566 count_partial(n, count_free);
4567 else
4568 x = atomic_long_read(&n->nr_slabs);
4569 total += x;
4570 nodes[node] += x;
4571 }
4572
4573 } else
4574 #endif
4575 if (flags & SO_PARTIAL) {
4576 struct kmem_cache_node *n;
4577
4578 for_each_kmem_cache_node(s, node, n) {
4579 if (flags & SO_TOTAL)
4580 x = count_partial(n, count_total);
4581 else if (flags & SO_OBJECTS)
4582 x = count_partial(n, count_inuse);
4583 else
4584 x = n->nr_partial;
4585 total += x;
4586 nodes[node] += x;
4587 }
4588 }
4589 x = sprintf(buf, "%lu", total);
4590 #ifdef CONFIG_NUMA
4591 for (node = 0; node < nr_node_ids; node++)
4592 if (nodes[node])
4593 x += sprintf(buf + x, " N%d=%lu",
4594 node, nodes[node]);
4595 #endif
4596 put_online_mems();
4597 kfree(nodes);
4598 return x + sprintf(buf + x, "\n");
4599 }
4600
4601 #ifdef CONFIG_SLUB_DEBUG
4602 static int any_slab_objects(struct kmem_cache *s)
4603 {
4604 int node;
4605 struct kmem_cache_node *n;
4606
4607 for_each_kmem_cache_node(s, node, n)
4608 if (atomic_long_read(&n->total_objects))
4609 return 1;
4610
4611 return 0;
4612 }
4613 #endif
4614
4615 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4616 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4617
4618 struct slab_attribute {
4619 struct attribute attr;
4620 ssize_t (*show)(struct kmem_cache *s, char *buf);
4621 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4622 };
4623
4624 #define SLAB_ATTR_RO(_name) \
4625 static struct slab_attribute _name##_attr = \
4626 __ATTR(_name, 0400, _name##_show, NULL)
4627
4628 #define SLAB_ATTR(_name) \
4629 static struct slab_attribute _name##_attr = \
4630 __ATTR(_name, 0600, _name##_show, _name##_store)
4631
4632 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4633 {
4634 return sprintf(buf, "%d\n", s->size);
4635 }
4636 SLAB_ATTR_RO(slab_size);
4637
4638 static ssize_t align_show(struct kmem_cache *s, char *buf)
4639 {
4640 return sprintf(buf, "%d\n", s->align);
4641 }
4642 SLAB_ATTR_RO(align);
4643
4644 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4645 {
4646 return sprintf(buf, "%d\n", s->object_size);
4647 }
4648 SLAB_ATTR_RO(object_size);
4649
4650 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4651 {
4652 return sprintf(buf, "%d\n", oo_objects(s->oo));
4653 }
4654 SLAB_ATTR_RO(objs_per_slab);
4655
4656 static ssize_t order_store(struct kmem_cache *s,
4657 const char *buf, size_t length)
4658 {
4659 unsigned long order;
4660 int err;
4661
4662 err = kstrtoul(buf, 10, &order);
4663 if (err)
4664 return err;
4665
4666 if (order > slub_max_order || order < slub_min_order)
4667 return -EINVAL;
4668
4669 calculate_sizes(s, order);
4670 return length;
4671 }
4672
4673 static ssize_t order_show(struct kmem_cache *s, char *buf)
4674 {
4675 return sprintf(buf, "%d\n", oo_order(s->oo));
4676 }
4677 SLAB_ATTR(order);
4678
4679 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4680 {
4681 return sprintf(buf, "%lu\n", s->min_partial);
4682 }
4683
4684 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4685 size_t length)
4686 {
4687 unsigned long min;
4688 int err;
4689
4690 err = kstrtoul(buf, 10, &min);
4691 if (err)
4692 return err;
4693
4694 set_min_partial(s, min);
4695 return length;
4696 }
4697 SLAB_ATTR(min_partial);
4698
4699 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4700 {
4701 return sprintf(buf, "%u\n", s->cpu_partial);
4702 }
4703
4704 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4705 size_t length)
4706 {
4707 unsigned long objects;
4708 int err;
4709
4710 err = kstrtoul(buf, 10, &objects);
4711 if (err)
4712 return err;
4713 if (objects && !kmem_cache_has_cpu_partial(s))
4714 return -EINVAL;
4715
4716 s->cpu_partial = objects;
4717 flush_all(s);
4718 return length;
4719 }
4720 SLAB_ATTR(cpu_partial);
4721
4722 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4723 {
4724 if (!s->ctor)
4725 return 0;
4726 return sprintf(buf, "%pS\n", s->ctor);
4727 }
4728 SLAB_ATTR_RO(ctor);
4729
4730 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4731 {
4732 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
4733 }
4734 SLAB_ATTR_RO(aliases);
4735
4736 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4737 {
4738 return show_slab_objects(s, buf, SO_PARTIAL);
4739 }
4740 SLAB_ATTR_RO(partial);
4741
4742 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4743 {
4744 return show_slab_objects(s, buf, SO_CPU);
4745 }
4746 SLAB_ATTR_RO(cpu_slabs);
4747
4748 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4749 {
4750 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4751 }
4752 SLAB_ATTR_RO(objects);
4753
4754 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4755 {
4756 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4757 }
4758 SLAB_ATTR_RO(objects_partial);
4759
4760 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4761 {
4762 int objects = 0;
4763 int pages = 0;
4764 int cpu;
4765 int len;
4766
4767 for_each_online_cpu(cpu) {
4768 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4769
4770 if (page) {
4771 pages += page->pages;
4772 objects += page->pobjects;
4773 }
4774 }
4775
4776 len = sprintf(buf, "%d(%d)", objects, pages);
4777
4778 #ifdef CONFIG_SMP
4779 for_each_online_cpu(cpu) {
4780 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4781
4782 if (page && len < PAGE_SIZE - 20)
4783 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4784 page->pobjects, page->pages);
4785 }
4786 #endif
4787 return len + sprintf(buf + len, "\n");
4788 }
4789 SLAB_ATTR_RO(slabs_cpu_partial);
4790
4791 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4792 {
4793 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4794 }
4795
4796 static ssize_t reclaim_account_store(struct kmem_cache *s,
4797 const char *buf, size_t length)
4798 {
4799 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4800 if (buf[0] == '1')
4801 s->flags |= SLAB_RECLAIM_ACCOUNT;
4802 return length;
4803 }
4804 SLAB_ATTR(reclaim_account);
4805
4806 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4807 {
4808 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4809 }
4810 SLAB_ATTR_RO(hwcache_align);
4811
4812 #ifdef CONFIG_ZONE_DMA
4813 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4814 {
4815 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4816 }
4817 SLAB_ATTR_RO(cache_dma);
4818 #endif
4819
4820 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4821 {
4822 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4823 }
4824 SLAB_ATTR_RO(destroy_by_rcu);
4825
4826 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4827 {
4828 return sprintf(buf, "%d\n", s->reserved);
4829 }
4830 SLAB_ATTR_RO(reserved);
4831
4832 #ifdef CONFIG_SLUB_DEBUG
4833 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4834 {
4835 return show_slab_objects(s, buf, SO_ALL);
4836 }
4837 SLAB_ATTR_RO(slabs);
4838
4839 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4840 {
4841 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4842 }
4843 SLAB_ATTR_RO(total_objects);
4844
4845 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4846 {
4847 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
4848 }
4849
4850 static ssize_t sanity_checks_store(struct kmem_cache *s,
4851 const char *buf, size_t length)
4852 {
4853 s->flags &= ~SLAB_CONSISTENCY_CHECKS;
4854 if (buf[0] == '1') {
4855 s->flags &= ~__CMPXCHG_DOUBLE;
4856 s->flags |= SLAB_CONSISTENCY_CHECKS;
4857 }
4858 return length;
4859 }
4860 SLAB_ATTR(sanity_checks);
4861
4862 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4863 {
4864 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4865 }
4866
4867 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4868 size_t length)
4869 {
4870 /*
4871 * Tracing a merged cache is going to give confusing results
4872 * as well as cause other issues like converting a mergeable
4873 * cache into an umergeable one.
4874 */
4875 if (s->refcount > 1)
4876 return -EINVAL;
4877
4878 s->flags &= ~SLAB_TRACE;
4879 if (buf[0] == '1') {
4880 s->flags &= ~__CMPXCHG_DOUBLE;
4881 s->flags |= SLAB_TRACE;
4882 }
4883 return length;
4884 }
4885 SLAB_ATTR(trace);
4886
4887 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4888 {
4889 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4890 }
4891
4892 static ssize_t red_zone_store(struct kmem_cache *s,
4893 const char *buf, size_t length)
4894 {
4895 if (any_slab_objects(s))
4896 return -EBUSY;
4897
4898 s->flags &= ~SLAB_RED_ZONE;
4899 if (buf[0] == '1') {
4900 s->flags |= SLAB_RED_ZONE;
4901 }
4902 calculate_sizes(s, -1);
4903 return length;
4904 }
4905 SLAB_ATTR(red_zone);
4906
4907 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4908 {
4909 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4910 }
4911
4912 static ssize_t poison_store(struct kmem_cache *s,
4913 const char *buf, size_t length)
4914 {
4915 if (any_slab_objects(s))
4916 return -EBUSY;
4917
4918 s->flags &= ~SLAB_POISON;
4919 if (buf[0] == '1') {
4920 s->flags |= SLAB_POISON;
4921 }
4922 calculate_sizes(s, -1);
4923 return length;
4924 }
4925 SLAB_ATTR(poison);
4926
4927 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4928 {
4929 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4930 }
4931
4932 static ssize_t store_user_store(struct kmem_cache *s,
4933 const char *buf, size_t length)
4934 {
4935 if (any_slab_objects(s))
4936 return -EBUSY;
4937
4938 s->flags &= ~SLAB_STORE_USER;
4939 if (buf[0] == '1') {
4940 s->flags &= ~__CMPXCHG_DOUBLE;
4941 s->flags |= SLAB_STORE_USER;
4942 }
4943 calculate_sizes(s, -1);
4944 return length;
4945 }
4946 SLAB_ATTR(store_user);
4947
4948 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4949 {
4950 return 0;
4951 }
4952
4953 static ssize_t validate_store(struct kmem_cache *s,
4954 const char *buf, size_t length)
4955 {
4956 int ret = -EINVAL;
4957
4958 if (buf[0] == '1') {
4959 ret = validate_slab_cache(s);
4960 if (ret >= 0)
4961 ret = length;
4962 }
4963 return ret;
4964 }
4965 SLAB_ATTR(validate);
4966
4967 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4968 {
4969 if (!(s->flags & SLAB_STORE_USER))
4970 return -ENOSYS;
4971 return list_locations(s, buf, TRACK_ALLOC);
4972 }
4973 SLAB_ATTR_RO(alloc_calls);
4974
4975 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4976 {
4977 if (!(s->flags & SLAB_STORE_USER))
4978 return -ENOSYS;
4979 return list_locations(s, buf, TRACK_FREE);
4980 }
4981 SLAB_ATTR_RO(free_calls);
4982 #endif /* CONFIG_SLUB_DEBUG */
4983
4984 #ifdef CONFIG_FAILSLAB
4985 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4986 {
4987 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4988 }
4989
4990 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4991 size_t length)
4992 {
4993 if (s->refcount > 1)
4994 return -EINVAL;
4995
4996 s->flags &= ~SLAB_FAILSLAB;
4997 if (buf[0] == '1')
4998 s->flags |= SLAB_FAILSLAB;
4999 return length;
5000 }
5001 SLAB_ATTR(failslab);
5002 #endif
5003
5004 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5005 {
5006 return 0;
5007 }
5008
5009 static ssize_t shrink_store(struct kmem_cache *s,
5010 const char *buf, size_t length)
5011 {
5012 if (buf[0] == '1')
5013 kmem_cache_shrink(s);
5014 else
5015 return -EINVAL;
5016 return length;
5017 }
5018 SLAB_ATTR(shrink);
5019
5020 #ifdef CONFIG_NUMA
5021 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5022 {
5023 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
5024 }
5025
5026 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5027 const char *buf, size_t length)
5028 {
5029 unsigned long ratio;
5030 int err;
5031
5032 err = kstrtoul(buf, 10, &ratio);
5033 if (err)
5034 return err;
5035
5036 if (ratio <= 100)
5037 s->remote_node_defrag_ratio = ratio * 10;
5038
5039 return length;
5040 }
5041 SLAB_ATTR(remote_node_defrag_ratio);
5042 #endif
5043
5044 #ifdef CONFIG_SLUB_STATS
5045 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5046 {
5047 unsigned long sum = 0;
5048 int cpu;
5049 int len;
5050 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
5051
5052 if (!data)
5053 return -ENOMEM;
5054
5055 for_each_online_cpu(cpu) {
5056 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5057
5058 data[cpu] = x;
5059 sum += x;
5060 }
5061
5062 len = sprintf(buf, "%lu", sum);
5063
5064 #ifdef CONFIG_SMP
5065 for_each_online_cpu(cpu) {
5066 if (data[cpu] && len < PAGE_SIZE - 20)
5067 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5068 }
5069 #endif
5070 kfree(data);
5071 return len + sprintf(buf + len, "\n");
5072 }
5073
5074 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5075 {
5076 int cpu;
5077
5078 for_each_online_cpu(cpu)
5079 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5080 }
5081
5082 #define STAT_ATTR(si, text) \
5083 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5084 { \
5085 return show_stat(s, buf, si); \
5086 } \
5087 static ssize_t text##_store(struct kmem_cache *s, \
5088 const char *buf, size_t length) \
5089 { \
5090 if (buf[0] != '0') \
5091 return -EINVAL; \
5092 clear_stat(s, si); \
5093 return length; \
5094 } \
5095 SLAB_ATTR(text); \
5096
5097 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5098 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5099 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5100 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5101 STAT_ATTR(FREE_FROZEN, free_frozen);
5102 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5103 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5104 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5105 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5106 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5107 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5108 STAT_ATTR(FREE_SLAB, free_slab);
5109 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5110 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5111 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5112 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5113 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5114 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5115 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5116 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5117 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5118 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5119 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5120 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5121 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5122 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5123 #endif
5124
5125 static struct attribute *slab_attrs[] = {
5126 &slab_size_attr.attr,
5127 &object_size_attr.attr,
5128 &objs_per_slab_attr.attr,
5129 &order_attr.attr,
5130 &min_partial_attr.attr,
5131 &cpu_partial_attr.attr,
5132 &objects_attr.attr,
5133 &objects_partial_attr.attr,
5134 &partial_attr.attr,
5135 &cpu_slabs_attr.attr,
5136 &ctor_attr.attr,
5137 &aliases_attr.attr,
5138 &align_attr.attr,
5139 &hwcache_align_attr.attr,
5140 &reclaim_account_attr.attr,
5141 &destroy_by_rcu_attr.attr,
5142 &shrink_attr.attr,
5143 &reserved_attr.attr,
5144 &slabs_cpu_partial_attr.attr,
5145 #ifdef CONFIG_SLUB_DEBUG
5146 &total_objects_attr.attr,
5147 &slabs_attr.attr,
5148 &sanity_checks_attr.attr,
5149 &trace_attr.attr,
5150 &red_zone_attr.attr,
5151 &poison_attr.attr,
5152 &store_user_attr.attr,
5153 &validate_attr.attr,
5154 &alloc_calls_attr.attr,
5155 &free_calls_attr.attr,
5156 #endif
5157 #ifdef CONFIG_ZONE_DMA
5158 &cache_dma_attr.attr,
5159 #endif
5160 #ifdef CONFIG_NUMA
5161 &remote_node_defrag_ratio_attr.attr,
5162 #endif
5163 #ifdef CONFIG_SLUB_STATS
5164 &alloc_fastpath_attr.attr,
5165 &alloc_slowpath_attr.attr,
5166 &free_fastpath_attr.attr,
5167 &free_slowpath_attr.attr,
5168 &free_frozen_attr.attr,
5169 &free_add_partial_attr.attr,
5170 &free_remove_partial_attr.attr,
5171 &alloc_from_partial_attr.attr,
5172 &alloc_slab_attr.attr,
5173 &alloc_refill_attr.attr,
5174 &alloc_node_mismatch_attr.attr,
5175 &free_slab_attr.attr,
5176 &cpuslab_flush_attr.attr,
5177 &deactivate_full_attr.attr,
5178 &deactivate_empty_attr.attr,
5179 &deactivate_to_head_attr.attr,
5180 &deactivate_to_tail_attr.attr,
5181 &deactivate_remote_frees_attr.attr,
5182 &deactivate_bypass_attr.attr,
5183 &order_fallback_attr.attr,
5184 &cmpxchg_double_fail_attr.attr,
5185 &cmpxchg_double_cpu_fail_attr.attr,
5186 &cpu_partial_alloc_attr.attr,
5187 &cpu_partial_free_attr.attr,
5188 &cpu_partial_node_attr.attr,
5189 &cpu_partial_drain_attr.attr,
5190 #endif
5191 #ifdef CONFIG_FAILSLAB
5192 &failslab_attr.attr,
5193 #endif
5194
5195 NULL
5196 };
5197
5198 static struct attribute_group slab_attr_group = {
5199 .attrs = slab_attrs,
5200 };
5201
5202 static ssize_t slab_attr_show(struct kobject *kobj,
5203 struct attribute *attr,
5204 char *buf)
5205 {
5206 struct slab_attribute *attribute;
5207 struct kmem_cache *s;
5208 int err;
5209
5210 attribute = to_slab_attr(attr);
5211 s = to_slab(kobj);
5212
5213 if (!attribute->show)
5214 return -EIO;
5215
5216 err = attribute->show(s, buf);
5217
5218 return err;
5219 }
5220
5221 static ssize_t slab_attr_store(struct kobject *kobj,
5222 struct attribute *attr,
5223 const char *buf, size_t len)
5224 {
5225 struct slab_attribute *attribute;
5226 struct kmem_cache *s;
5227 int err;
5228
5229 attribute = to_slab_attr(attr);
5230 s = to_slab(kobj);
5231
5232 if (!attribute->store)
5233 return -EIO;
5234
5235 err = attribute->store(s, buf, len);
5236 #ifdef CONFIG_MEMCG
5237 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5238 struct kmem_cache *c;
5239
5240 mutex_lock(&slab_mutex);
5241 if (s->max_attr_size < len)
5242 s->max_attr_size = len;
5243
5244 /*
5245 * This is a best effort propagation, so this function's return
5246 * value will be determined by the parent cache only. This is
5247 * basically because not all attributes will have a well
5248 * defined semantics for rollbacks - most of the actions will
5249 * have permanent effects.
5250 *
5251 * Returning the error value of any of the children that fail
5252 * is not 100 % defined, in the sense that users seeing the
5253 * error code won't be able to know anything about the state of
5254 * the cache.
5255 *
5256 * Only returning the error code for the parent cache at least
5257 * has well defined semantics. The cache being written to
5258 * directly either failed or succeeded, in which case we loop
5259 * through the descendants with best-effort propagation.
5260 */
5261 for_each_memcg_cache(c, s)
5262 attribute->store(c, buf, len);
5263 mutex_unlock(&slab_mutex);
5264 }
5265 #endif
5266 return err;
5267 }
5268
5269 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5270 {
5271 #ifdef CONFIG_MEMCG
5272 int i;
5273 char *buffer = NULL;
5274 struct kmem_cache *root_cache;
5275
5276 if (is_root_cache(s))
5277 return;
5278
5279 root_cache = s->memcg_params.root_cache;
5280
5281 /*
5282 * This mean this cache had no attribute written. Therefore, no point
5283 * in copying default values around
5284 */
5285 if (!root_cache->max_attr_size)
5286 return;
5287
5288 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5289 char mbuf[64];
5290 char *buf;
5291 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5292
5293 if (!attr || !attr->store || !attr->show)
5294 continue;
5295
5296 /*
5297 * It is really bad that we have to allocate here, so we will
5298 * do it only as a fallback. If we actually allocate, though,
5299 * we can just use the allocated buffer until the end.
5300 *
5301 * Most of the slub attributes will tend to be very small in
5302 * size, but sysfs allows buffers up to a page, so they can
5303 * theoretically happen.
5304 */
5305 if (buffer)
5306 buf = buffer;
5307 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5308 buf = mbuf;
5309 else {
5310 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5311 if (WARN_ON(!buffer))
5312 continue;
5313 buf = buffer;
5314 }
5315
5316 attr->show(root_cache, buf);
5317 attr->store(s, buf, strlen(buf));
5318 }
5319
5320 if (buffer)
5321 free_page((unsigned long)buffer);
5322 #endif
5323 }
5324
5325 static void kmem_cache_release(struct kobject *k)
5326 {
5327 slab_kmem_cache_release(to_slab(k));
5328 }
5329
5330 static const struct sysfs_ops slab_sysfs_ops = {
5331 .show = slab_attr_show,
5332 .store = slab_attr_store,
5333 };
5334
5335 static struct kobj_type slab_ktype = {
5336 .sysfs_ops = &slab_sysfs_ops,
5337 .release = kmem_cache_release,
5338 };
5339
5340 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5341 {
5342 struct kobj_type *ktype = get_ktype(kobj);
5343
5344 if (ktype == &slab_ktype)
5345 return 1;
5346 return 0;
5347 }
5348
5349 static const struct kset_uevent_ops slab_uevent_ops = {
5350 .filter = uevent_filter,
5351 };
5352
5353 static struct kset *slab_kset;
5354
5355 static inline struct kset *cache_kset(struct kmem_cache *s)
5356 {
5357 #ifdef CONFIG_MEMCG
5358 if (!is_root_cache(s))
5359 return s->memcg_params.root_cache->memcg_kset;
5360 #endif
5361 return slab_kset;
5362 }
5363
5364 #define ID_STR_LENGTH 64
5365
5366 /* Create a unique string id for a slab cache:
5367 *
5368 * Format :[flags-]size
5369 */
5370 static char *create_unique_id(struct kmem_cache *s)
5371 {
5372 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5373 char *p = name;
5374
5375 BUG_ON(!name);
5376
5377 *p++ = ':';
5378 /*
5379 * First flags affecting slabcache operations. We will only
5380 * get here for aliasable slabs so we do not need to support
5381 * too many flags. The flags here must cover all flags that
5382 * are matched during merging to guarantee that the id is
5383 * unique.
5384 */
5385 if (s->flags & SLAB_CACHE_DMA)
5386 *p++ = 'd';
5387 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5388 *p++ = 'a';
5389 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5390 *p++ = 'F';
5391 if (!(s->flags & SLAB_NOTRACK))
5392 *p++ = 't';
5393 if (s->flags & SLAB_ACCOUNT)
5394 *p++ = 'A';
5395 if (p != name + 1)
5396 *p++ = '-';
5397 p += sprintf(p, "%07d", s->size);
5398
5399 BUG_ON(p > name + ID_STR_LENGTH - 1);
5400 return name;
5401 }
5402
5403 static int sysfs_slab_add(struct kmem_cache *s)
5404 {
5405 int err;
5406 const char *name;
5407 int unmergeable = slab_unmergeable(s);
5408
5409 if (unmergeable) {
5410 /*
5411 * Slabcache can never be merged so we can use the name proper.
5412 * This is typically the case for debug situations. In that
5413 * case we can catch duplicate names easily.
5414 */
5415 sysfs_remove_link(&slab_kset->kobj, s->name);
5416 name = s->name;
5417 } else {
5418 /*
5419 * Create a unique name for the slab as a target
5420 * for the symlinks.
5421 */
5422 name = create_unique_id(s);
5423 }
5424
5425 s->kobj.kset = cache_kset(s);
5426 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5427 if (err)
5428 goto out;
5429
5430 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5431 if (err)
5432 goto out_del_kobj;
5433
5434 #ifdef CONFIG_MEMCG
5435 if (is_root_cache(s)) {
5436 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5437 if (!s->memcg_kset) {
5438 err = -ENOMEM;
5439 goto out_del_kobj;
5440 }
5441 }
5442 #endif
5443
5444 kobject_uevent(&s->kobj, KOBJ_ADD);
5445 if (!unmergeable) {
5446 /* Setup first alias */
5447 sysfs_slab_alias(s, s->name);
5448 }
5449 out:
5450 if (!unmergeable)
5451 kfree(name);
5452 return err;
5453 out_del_kobj:
5454 kobject_del(&s->kobj);
5455 goto out;
5456 }
5457
5458 void sysfs_slab_remove(struct kmem_cache *s)
5459 {
5460 if (slab_state < FULL)
5461 /*
5462 * Sysfs has not been setup yet so no need to remove the
5463 * cache from sysfs.
5464 */
5465 return;
5466
5467 #ifdef CONFIG_MEMCG
5468 kset_unregister(s->memcg_kset);
5469 #endif
5470 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5471 kobject_del(&s->kobj);
5472 kobject_put(&s->kobj);
5473 }
5474
5475 /*
5476 * Need to buffer aliases during bootup until sysfs becomes
5477 * available lest we lose that information.
5478 */
5479 struct saved_alias {
5480 struct kmem_cache *s;
5481 const char *name;
5482 struct saved_alias *next;
5483 };
5484
5485 static struct saved_alias *alias_list;
5486
5487 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5488 {
5489 struct saved_alias *al;
5490
5491 if (slab_state == FULL) {
5492 /*
5493 * If we have a leftover link then remove it.
5494 */
5495 sysfs_remove_link(&slab_kset->kobj, name);
5496 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5497 }
5498
5499 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5500 if (!al)
5501 return -ENOMEM;
5502
5503 al->s = s;
5504 al->name = name;
5505 al->next = alias_list;
5506 alias_list = al;
5507 return 0;
5508 }
5509
5510 static int __init slab_sysfs_init(void)
5511 {
5512 struct kmem_cache *s;
5513 int err;
5514
5515 mutex_lock(&slab_mutex);
5516
5517 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5518 if (!slab_kset) {
5519 mutex_unlock(&slab_mutex);
5520 pr_err("Cannot register slab subsystem.\n");
5521 return -ENOSYS;
5522 }
5523
5524 slab_state = FULL;
5525
5526 list_for_each_entry(s, &slab_caches, list) {
5527 err = sysfs_slab_add(s);
5528 if (err)
5529 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5530 s->name);
5531 }
5532
5533 while (alias_list) {
5534 struct saved_alias *al = alias_list;
5535
5536 alias_list = alias_list->next;
5537 err = sysfs_slab_alias(al->s, al->name);
5538 if (err)
5539 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5540 al->name);
5541 kfree(al);
5542 }
5543
5544 mutex_unlock(&slab_mutex);
5545 resiliency_test();
5546 return 0;
5547 }
5548
5549 __initcall(slab_sysfs_init);
5550 #endif /* CONFIG_SYSFS */
5551
5552 /*
5553 * The /proc/slabinfo ABI
5554 */
5555 #ifdef CONFIG_SLABINFO
5556 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5557 {
5558 unsigned long nr_slabs = 0;
5559 unsigned long nr_objs = 0;
5560 unsigned long nr_free = 0;
5561 int node;
5562 struct kmem_cache_node *n;
5563
5564 for_each_kmem_cache_node(s, node, n) {
5565 nr_slabs += node_nr_slabs(n);
5566 nr_objs += node_nr_objs(n);
5567 nr_free += count_partial(n, count_free);
5568 }
5569
5570 sinfo->active_objs = nr_objs - nr_free;
5571 sinfo->num_objs = nr_objs;
5572 sinfo->active_slabs = nr_slabs;
5573 sinfo->num_slabs = nr_slabs;
5574 sinfo->objects_per_slab = oo_objects(s->oo);
5575 sinfo->cache_order = oo_order(s->oo);
5576 }
5577
5578 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5579 {
5580 }
5581
5582 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5583 size_t count, loff_t *ppos)
5584 {
5585 return -EIO;
5586 }
5587 #endif /* CONFIG_SLABINFO */