]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blob - mm/slub.c
Merge tag 'xtensa-20150830' of git://github.com/czankel/xtensa-linux
[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 bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
128 {
129 #ifdef CONFIG_SLUB_CPU_PARTIAL
130 return !kmem_cache_debug(s);
131 #else
132 return false;
133 #endif
134 }
135
136 /*
137 * Issues still to be resolved:
138 *
139 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
140 *
141 * - Variable sizing of the per node arrays
142 */
143
144 /* Enable to test recovery from slab corruption on boot */
145 #undef SLUB_RESILIENCY_TEST
146
147 /* Enable to log cmpxchg failures */
148 #undef SLUB_DEBUG_CMPXCHG
149
150 /*
151 * Mininum number of partial slabs. These will be left on the partial
152 * lists even if they are empty. kmem_cache_shrink may reclaim them.
153 */
154 #define MIN_PARTIAL 5
155
156 /*
157 * Maximum number of desirable partial slabs.
158 * The existence of more partial slabs makes kmem_cache_shrink
159 * sort the partial list by the number of objects in use.
160 */
161 #define MAX_PARTIAL 10
162
163 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
164 SLAB_POISON | SLAB_STORE_USER)
165
166 /*
167 * Debugging flags that require metadata to be stored in the slab. These get
168 * disabled when slub_debug=O is used and a cache's min order increases with
169 * metadata.
170 */
171 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
172
173 #define OO_SHIFT 16
174 #define OO_MASK ((1 << OO_SHIFT) - 1)
175 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
176
177 /* Internal SLUB flags */
178 #define __OBJECT_POISON 0x80000000UL /* Poison object */
179 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
180
181 #ifdef CONFIG_SMP
182 static struct notifier_block slab_notifier;
183 #endif
184
185 /*
186 * Tracking user of a slab.
187 */
188 #define TRACK_ADDRS_COUNT 16
189 struct track {
190 unsigned long addr; /* Called from address */
191 #ifdef CONFIG_STACKTRACE
192 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
193 #endif
194 int cpu; /* Was running on cpu */
195 int pid; /* Pid context */
196 unsigned long when; /* When did the operation occur */
197 };
198
199 enum track_item { TRACK_ALLOC, TRACK_FREE };
200
201 #ifdef CONFIG_SYSFS
202 static int sysfs_slab_add(struct kmem_cache *);
203 static int sysfs_slab_alias(struct kmem_cache *, const char *);
204 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
205 #else
206 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
207 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
208 { return 0; }
209 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
210 #endif
211
212 static inline void stat(const struct kmem_cache *s, enum stat_item si)
213 {
214 #ifdef CONFIG_SLUB_STATS
215 /*
216 * The rmw is racy on a preemptible kernel but this is acceptable, so
217 * avoid this_cpu_add()'s irq-disable overhead.
218 */
219 raw_cpu_inc(s->cpu_slab->stat[si]);
220 #endif
221 }
222
223 /********************************************************************
224 * Core slab cache functions
225 *******************************************************************/
226
227 /* Verify that a pointer has an address that is valid within a slab page */
228 static inline int check_valid_pointer(struct kmem_cache *s,
229 struct page *page, const void *object)
230 {
231 void *base;
232
233 if (!object)
234 return 1;
235
236 base = page_address(page);
237 if (object < base || object >= base + page->objects * s->size ||
238 (object - base) % s->size) {
239 return 0;
240 }
241
242 return 1;
243 }
244
245 static inline void *get_freepointer(struct kmem_cache *s, void *object)
246 {
247 return *(void **)(object + s->offset);
248 }
249
250 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
251 {
252 prefetch(object + s->offset);
253 }
254
255 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
256 {
257 void *p;
258
259 #ifdef CONFIG_DEBUG_PAGEALLOC
260 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
261 #else
262 p = get_freepointer(s, object);
263 #endif
264 return p;
265 }
266
267 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
268 {
269 *(void **)(object + s->offset) = fp;
270 }
271
272 /* Loop over all objects in a slab */
273 #define for_each_object(__p, __s, __addr, __objects) \
274 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
275 __p += (__s)->size)
276
277 #define for_each_object_idx(__p, __idx, __s, __addr, __objects) \
278 for (__p = (__addr), __idx = 1; __idx <= __objects;\
279 __p += (__s)->size, __idx++)
280
281 /* Determine object index from a given position */
282 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
283 {
284 return (p - addr) / s->size;
285 }
286
287 static inline size_t slab_ksize(const struct kmem_cache *s)
288 {
289 #ifdef CONFIG_SLUB_DEBUG
290 /*
291 * Debugging requires use of the padding between object
292 * and whatever may come after it.
293 */
294 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
295 return s->object_size;
296
297 #endif
298 /*
299 * If we have the need to store the freelist pointer
300 * back there or track user information then we can
301 * only use the space before that information.
302 */
303 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
304 return s->inuse;
305 /*
306 * Else we can use all the padding etc for the allocation
307 */
308 return s->size;
309 }
310
311 static inline int order_objects(int order, unsigned long size, int reserved)
312 {
313 return ((PAGE_SIZE << order) - reserved) / size;
314 }
315
316 static inline struct kmem_cache_order_objects oo_make(int order,
317 unsigned long size, int reserved)
318 {
319 struct kmem_cache_order_objects x = {
320 (order << OO_SHIFT) + order_objects(order, size, reserved)
321 };
322
323 return x;
324 }
325
326 static inline int oo_order(struct kmem_cache_order_objects x)
327 {
328 return x.x >> OO_SHIFT;
329 }
330
331 static inline int oo_objects(struct kmem_cache_order_objects x)
332 {
333 return x.x & OO_MASK;
334 }
335
336 /*
337 * Per slab locking using the pagelock
338 */
339 static __always_inline void slab_lock(struct page *page)
340 {
341 bit_spin_lock(PG_locked, &page->flags);
342 }
343
344 static __always_inline void slab_unlock(struct page *page)
345 {
346 __bit_spin_unlock(PG_locked, &page->flags);
347 }
348
349 static inline void set_page_slub_counters(struct page *page, unsigned long counters_new)
350 {
351 struct page tmp;
352 tmp.counters = counters_new;
353 /*
354 * page->counters can cover frozen/inuse/objects as well
355 * as page->_count. If we assign to ->counters directly
356 * we run the risk of losing updates to page->_count, so
357 * be careful and only assign to the fields we need.
358 */
359 page->frozen = tmp.frozen;
360 page->inuse = tmp.inuse;
361 page->objects = tmp.objects;
362 }
363
364 /* Interrupts must be disabled (for the fallback code to work right) */
365 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
366 void *freelist_old, unsigned long counters_old,
367 void *freelist_new, unsigned long counters_new,
368 const char *n)
369 {
370 VM_BUG_ON(!irqs_disabled());
371 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
372 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
373 if (s->flags & __CMPXCHG_DOUBLE) {
374 if (cmpxchg_double(&page->freelist, &page->counters,
375 freelist_old, counters_old,
376 freelist_new, counters_new))
377 return true;
378 } else
379 #endif
380 {
381 slab_lock(page);
382 if (page->freelist == freelist_old &&
383 page->counters == counters_old) {
384 page->freelist = freelist_new;
385 set_page_slub_counters(page, counters_new);
386 slab_unlock(page);
387 return true;
388 }
389 slab_unlock(page);
390 }
391
392 cpu_relax();
393 stat(s, CMPXCHG_DOUBLE_FAIL);
394
395 #ifdef SLUB_DEBUG_CMPXCHG
396 pr_info("%s %s: cmpxchg double redo ", n, s->name);
397 #endif
398
399 return false;
400 }
401
402 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
403 void *freelist_old, unsigned long counters_old,
404 void *freelist_new, unsigned long counters_new,
405 const char *n)
406 {
407 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
408 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
409 if (s->flags & __CMPXCHG_DOUBLE) {
410 if (cmpxchg_double(&page->freelist, &page->counters,
411 freelist_old, counters_old,
412 freelist_new, counters_new))
413 return true;
414 } else
415 #endif
416 {
417 unsigned long flags;
418
419 local_irq_save(flags);
420 slab_lock(page);
421 if (page->freelist == freelist_old &&
422 page->counters == counters_old) {
423 page->freelist = freelist_new;
424 set_page_slub_counters(page, counters_new);
425 slab_unlock(page);
426 local_irq_restore(flags);
427 return true;
428 }
429 slab_unlock(page);
430 local_irq_restore(flags);
431 }
432
433 cpu_relax();
434 stat(s, CMPXCHG_DOUBLE_FAIL);
435
436 #ifdef SLUB_DEBUG_CMPXCHG
437 pr_info("%s %s: cmpxchg double redo ", n, s->name);
438 #endif
439
440 return false;
441 }
442
443 #ifdef CONFIG_SLUB_DEBUG
444 /*
445 * Determine a map of object in use on a page.
446 *
447 * Node listlock must be held to guarantee that the page does
448 * not vanish from under us.
449 */
450 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
451 {
452 void *p;
453 void *addr = page_address(page);
454
455 for (p = page->freelist; p; p = get_freepointer(s, p))
456 set_bit(slab_index(p, s, addr), map);
457 }
458
459 /*
460 * Debug settings:
461 */
462 #ifdef CONFIG_SLUB_DEBUG_ON
463 static int slub_debug = DEBUG_DEFAULT_FLAGS;
464 #else
465 static int slub_debug;
466 #endif
467
468 static char *slub_debug_slabs;
469 static int disable_higher_order_debug;
470
471 /*
472 * slub is about to manipulate internal object metadata. This memory lies
473 * outside the range of the allocated object, so accessing it would normally
474 * be reported by kasan as a bounds error. metadata_access_enable() is used
475 * to tell kasan that these accesses are OK.
476 */
477 static inline void metadata_access_enable(void)
478 {
479 kasan_disable_current();
480 }
481
482 static inline void metadata_access_disable(void)
483 {
484 kasan_enable_current();
485 }
486
487 /*
488 * Object debugging
489 */
490 static void print_section(char *text, u8 *addr, unsigned int length)
491 {
492 metadata_access_enable();
493 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
494 length, 1);
495 metadata_access_disable();
496 }
497
498 static struct track *get_track(struct kmem_cache *s, void *object,
499 enum track_item alloc)
500 {
501 struct track *p;
502
503 if (s->offset)
504 p = object + s->offset + sizeof(void *);
505 else
506 p = object + s->inuse;
507
508 return p + alloc;
509 }
510
511 static void set_track(struct kmem_cache *s, void *object,
512 enum track_item alloc, unsigned long addr)
513 {
514 struct track *p = get_track(s, object, alloc);
515
516 if (addr) {
517 #ifdef CONFIG_STACKTRACE
518 struct stack_trace trace;
519 int i;
520
521 trace.nr_entries = 0;
522 trace.max_entries = TRACK_ADDRS_COUNT;
523 trace.entries = p->addrs;
524 trace.skip = 3;
525 metadata_access_enable();
526 save_stack_trace(&trace);
527 metadata_access_disable();
528
529 /* See rant in lockdep.c */
530 if (trace.nr_entries != 0 &&
531 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
532 trace.nr_entries--;
533
534 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
535 p->addrs[i] = 0;
536 #endif
537 p->addr = addr;
538 p->cpu = smp_processor_id();
539 p->pid = current->pid;
540 p->when = jiffies;
541 } else
542 memset(p, 0, sizeof(struct track));
543 }
544
545 static void init_tracking(struct kmem_cache *s, void *object)
546 {
547 if (!(s->flags & SLAB_STORE_USER))
548 return;
549
550 set_track(s, object, TRACK_FREE, 0UL);
551 set_track(s, object, TRACK_ALLOC, 0UL);
552 }
553
554 static void print_track(const char *s, struct track *t)
555 {
556 if (!t->addr)
557 return;
558
559 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
560 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
561 #ifdef CONFIG_STACKTRACE
562 {
563 int i;
564 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
565 if (t->addrs[i])
566 pr_err("\t%pS\n", (void *)t->addrs[i]);
567 else
568 break;
569 }
570 #endif
571 }
572
573 static void print_tracking(struct kmem_cache *s, void *object)
574 {
575 if (!(s->flags & SLAB_STORE_USER))
576 return;
577
578 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
579 print_track("Freed", get_track(s, object, TRACK_FREE));
580 }
581
582 static void print_page_info(struct page *page)
583 {
584 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
585 page, page->objects, page->inuse, page->freelist, page->flags);
586
587 }
588
589 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
590 {
591 struct va_format vaf;
592 va_list args;
593
594 va_start(args, fmt);
595 vaf.fmt = fmt;
596 vaf.va = &args;
597 pr_err("=============================================================================\n");
598 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
599 pr_err("-----------------------------------------------------------------------------\n\n");
600
601 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
602 va_end(args);
603 }
604
605 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
606 {
607 struct va_format vaf;
608 va_list args;
609
610 va_start(args, fmt);
611 vaf.fmt = fmt;
612 vaf.va = &args;
613 pr_err("FIX %s: %pV\n", s->name, &vaf);
614 va_end(args);
615 }
616
617 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
618 {
619 unsigned int off; /* Offset of last byte */
620 u8 *addr = page_address(page);
621
622 print_tracking(s, p);
623
624 print_page_info(page);
625
626 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
627 p, p - addr, get_freepointer(s, p));
628
629 if (p > addr + 16)
630 print_section("Bytes b4 ", p - 16, 16);
631
632 print_section("Object ", p, min_t(unsigned long, s->object_size,
633 PAGE_SIZE));
634 if (s->flags & SLAB_RED_ZONE)
635 print_section("Redzone ", p + s->object_size,
636 s->inuse - s->object_size);
637
638 if (s->offset)
639 off = s->offset + sizeof(void *);
640 else
641 off = s->inuse;
642
643 if (s->flags & SLAB_STORE_USER)
644 off += 2 * sizeof(struct track);
645
646 if (off != s->size)
647 /* Beginning of the filler is the free pointer */
648 print_section("Padding ", p + off, s->size - off);
649
650 dump_stack();
651 }
652
653 void object_err(struct kmem_cache *s, struct page *page,
654 u8 *object, char *reason)
655 {
656 slab_bug(s, "%s", reason);
657 print_trailer(s, page, object);
658 }
659
660 static void slab_err(struct kmem_cache *s, struct page *page,
661 const char *fmt, ...)
662 {
663 va_list args;
664 char buf[100];
665
666 va_start(args, fmt);
667 vsnprintf(buf, sizeof(buf), fmt, args);
668 va_end(args);
669 slab_bug(s, "%s", buf);
670 print_page_info(page);
671 dump_stack();
672 }
673
674 static void init_object(struct kmem_cache *s, void *object, u8 val)
675 {
676 u8 *p = object;
677
678 if (s->flags & __OBJECT_POISON) {
679 memset(p, POISON_FREE, s->object_size - 1);
680 p[s->object_size - 1] = POISON_END;
681 }
682
683 if (s->flags & SLAB_RED_ZONE)
684 memset(p + s->object_size, val, s->inuse - s->object_size);
685 }
686
687 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
688 void *from, void *to)
689 {
690 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
691 memset(from, data, to - from);
692 }
693
694 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
695 u8 *object, char *what,
696 u8 *start, unsigned int value, unsigned int bytes)
697 {
698 u8 *fault;
699 u8 *end;
700
701 metadata_access_enable();
702 fault = memchr_inv(start, value, bytes);
703 metadata_access_disable();
704 if (!fault)
705 return 1;
706
707 end = start + bytes;
708 while (end > fault && end[-1] == value)
709 end--;
710
711 slab_bug(s, "%s overwritten", what);
712 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
713 fault, end - 1, fault[0], value);
714 print_trailer(s, page, object);
715
716 restore_bytes(s, what, value, fault, end);
717 return 0;
718 }
719
720 /*
721 * Object layout:
722 *
723 * object address
724 * Bytes of the object to be managed.
725 * If the freepointer may overlay the object then the free
726 * pointer is the first word of the object.
727 *
728 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
729 * 0xa5 (POISON_END)
730 *
731 * object + s->object_size
732 * Padding to reach word boundary. This is also used for Redzoning.
733 * Padding is extended by another word if Redzoning is enabled and
734 * object_size == inuse.
735 *
736 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
737 * 0xcc (RED_ACTIVE) for objects in use.
738 *
739 * object + s->inuse
740 * Meta data starts here.
741 *
742 * A. Free pointer (if we cannot overwrite object on free)
743 * B. Tracking data for SLAB_STORE_USER
744 * C. Padding to reach required alignment boundary or at mininum
745 * one word if debugging is on to be able to detect writes
746 * before the word boundary.
747 *
748 * Padding is done using 0x5a (POISON_INUSE)
749 *
750 * object + s->size
751 * Nothing is used beyond s->size.
752 *
753 * If slabcaches are merged then the object_size and inuse boundaries are mostly
754 * ignored. And therefore no slab options that rely on these boundaries
755 * may be used with merged slabcaches.
756 */
757
758 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
759 {
760 unsigned long off = s->inuse; /* The end of info */
761
762 if (s->offset)
763 /* Freepointer is placed after the object. */
764 off += sizeof(void *);
765
766 if (s->flags & SLAB_STORE_USER)
767 /* We also have user information there */
768 off += 2 * sizeof(struct track);
769
770 if (s->size == off)
771 return 1;
772
773 return check_bytes_and_report(s, page, p, "Object padding",
774 p + off, POISON_INUSE, s->size - off);
775 }
776
777 /* Check the pad bytes at the end of a slab page */
778 static int slab_pad_check(struct kmem_cache *s, struct page *page)
779 {
780 u8 *start;
781 u8 *fault;
782 u8 *end;
783 int length;
784 int remainder;
785
786 if (!(s->flags & SLAB_POISON))
787 return 1;
788
789 start = page_address(page);
790 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
791 end = start + length;
792 remainder = length % s->size;
793 if (!remainder)
794 return 1;
795
796 metadata_access_enable();
797 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
798 metadata_access_disable();
799 if (!fault)
800 return 1;
801 while (end > fault && end[-1] == POISON_INUSE)
802 end--;
803
804 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
805 print_section("Padding ", end - remainder, remainder);
806
807 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
808 return 0;
809 }
810
811 static int check_object(struct kmem_cache *s, struct page *page,
812 void *object, u8 val)
813 {
814 u8 *p = object;
815 u8 *endobject = object + s->object_size;
816
817 if (s->flags & SLAB_RED_ZONE) {
818 if (!check_bytes_and_report(s, page, object, "Redzone",
819 endobject, val, s->inuse - s->object_size))
820 return 0;
821 } else {
822 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
823 check_bytes_and_report(s, page, p, "Alignment padding",
824 endobject, POISON_INUSE,
825 s->inuse - s->object_size);
826 }
827 }
828
829 if (s->flags & SLAB_POISON) {
830 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
831 (!check_bytes_and_report(s, page, p, "Poison", p,
832 POISON_FREE, s->object_size - 1) ||
833 !check_bytes_and_report(s, page, p, "Poison",
834 p + s->object_size - 1, POISON_END, 1)))
835 return 0;
836 /*
837 * check_pad_bytes cleans up on its own.
838 */
839 check_pad_bytes(s, page, p);
840 }
841
842 if (!s->offset && val == SLUB_RED_ACTIVE)
843 /*
844 * Object and freepointer overlap. Cannot check
845 * freepointer while object is allocated.
846 */
847 return 1;
848
849 /* Check free pointer validity */
850 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
851 object_err(s, page, p, "Freepointer corrupt");
852 /*
853 * No choice but to zap it and thus lose the remainder
854 * of the free objects in this slab. May cause
855 * another error because the object count is now wrong.
856 */
857 set_freepointer(s, p, NULL);
858 return 0;
859 }
860 return 1;
861 }
862
863 static int check_slab(struct kmem_cache *s, struct page *page)
864 {
865 int maxobj;
866
867 VM_BUG_ON(!irqs_disabled());
868
869 if (!PageSlab(page)) {
870 slab_err(s, page, "Not a valid slab page");
871 return 0;
872 }
873
874 maxobj = order_objects(compound_order(page), s->size, s->reserved);
875 if (page->objects > maxobj) {
876 slab_err(s, page, "objects %u > max %u",
877 page->objects, maxobj);
878 return 0;
879 }
880 if (page->inuse > page->objects) {
881 slab_err(s, page, "inuse %u > max %u",
882 page->inuse, page->objects);
883 return 0;
884 }
885 /* Slab_pad_check fixes things up after itself */
886 slab_pad_check(s, page);
887 return 1;
888 }
889
890 /*
891 * Determine if a certain object on a page is on the freelist. Must hold the
892 * slab lock to guarantee that the chains are in a consistent state.
893 */
894 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
895 {
896 int nr = 0;
897 void *fp;
898 void *object = NULL;
899 int max_objects;
900
901 fp = page->freelist;
902 while (fp && nr <= page->objects) {
903 if (fp == search)
904 return 1;
905 if (!check_valid_pointer(s, page, fp)) {
906 if (object) {
907 object_err(s, page, object,
908 "Freechain corrupt");
909 set_freepointer(s, object, NULL);
910 } else {
911 slab_err(s, page, "Freepointer corrupt");
912 page->freelist = NULL;
913 page->inuse = page->objects;
914 slab_fix(s, "Freelist cleared");
915 return 0;
916 }
917 break;
918 }
919 object = fp;
920 fp = get_freepointer(s, object);
921 nr++;
922 }
923
924 max_objects = order_objects(compound_order(page), s->size, s->reserved);
925 if (max_objects > MAX_OBJS_PER_PAGE)
926 max_objects = MAX_OBJS_PER_PAGE;
927
928 if (page->objects != max_objects) {
929 slab_err(s, page, "Wrong number of objects. Found %d but "
930 "should be %d", page->objects, max_objects);
931 page->objects = max_objects;
932 slab_fix(s, "Number of objects adjusted.");
933 }
934 if (page->inuse != page->objects - nr) {
935 slab_err(s, page, "Wrong object count. Counter is %d but "
936 "counted were %d", page->inuse, page->objects - nr);
937 page->inuse = page->objects - nr;
938 slab_fix(s, "Object count adjusted.");
939 }
940 return search == NULL;
941 }
942
943 static void trace(struct kmem_cache *s, struct page *page, void *object,
944 int alloc)
945 {
946 if (s->flags & SLAB_TRACE) {
947 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
948 s->name,
949 alloc ? "alloc" : "free",
950 object, page->inuse,
951 page->freelist);
952
953 if (!alloc)
954 print_section("Object ", (void *)object,
955 s->object_size);
956
957 dump_stack();
958 }
959 }
960
961 /*
962 * Tracking of fully allocated slabs for debugging purposes.
963 */
964 static void add_full(struct kmem_cache *s,
965 struct kmem_cache_node *n, struct page *page)
966 {
967 if (!(s->flags & SLAB_STORE_USER))
968 return;
969
970 lockdep_assert_held(&n->list_lock);
971 list_add(&page->lru, &n->full);
972 }
973
974 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
975 {
976 if (!(s->flags & SLAB_STORE_USER))
977 return;
978
979 lockdep_assert_held(&n->list_lock);
980 list_del(&page->lru);
981 }
982
983 /* Tracking of the number of slabs for debugging purposes */
984 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
985 {
986 struct kmem_cache_node *n = get_node(s, node);
987
988 return atomic_long_read(&n->nr_slabs);
989 }
990
991 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
992 {
993 return atomic_long_read(&n->nr_slabs);
994 }
995
996 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
997 {
998 struct kmem_cache_node *n = get_node(s, node);
999
1000 /*
1001 * May be called early in order to allocate a slab for the
1002 * kmem_cache_node structure. Solve the chicken-egg
1003 * dilemma by deferring the increment of the count during
1004 * bootstrap (see early_kmem_cache_node_alloc).
1005 */
1006 if (likely(n)) {
1007 atomic_long_inc(&n->nr_slabs);
1008 atomic_long_add(objects, &n->total_objects);
1009 }
1010 }
1011 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1012 {
1013 struct kmem_cache_node *n = get_node(s, node);
1014
1015 atomic_long_dec(&n->nr_slabs);
1016 atomic_long_sub(objects, &n->total_objects);
1017 }
1018
1019 /* Object debug checks for alloc/free paths */
1020 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1021 void *object)
1022 {
1023 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1024 return;
1025
1026 init_object(s, object, SLUB_RED_INACTIVE);
1027 init_tracking(s, object);
1028 }
1029
1030 static noinline int alloc_debug_processing(struct kmem_cache *s,
1031 struct page *page,
1032 void *object, unsigned long addr)
1033 {
1034 if (!check_slab(s, page))
1035 goto bad;
1036
1037 if (!check_valid_pointer(s, page, object)) {
1038 object_err(s, page, object, "Freelist Pointer check fails");
1039 goto bad;
1040 }
1041
1042 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1043 goto bad;
1044
1045 /* Success perform special debug activities for allocs */
1046 if (s->flags & SLAB_STORE_USER)
1047 set_track(s, object, TRACK_ALLOC, addr);
1048 trace(s, page, object, 1);
1049 init_object(s, object, SLUB_RED_ACTIVE);
1050 return 1;
1051
1052 bad:
1053 if (PageSlab(page)) {
1054 /*
1055 * If this is a slab page then lets do the best we can
1056 * to avoid issues in the future. Marking all objects
1057 * as used avoids touching the remaining objects.
1058 */
1059 slab_fix(s, "Marking all objects used");
1060 page->inuse = page->objects;
1061 page->freelist = NULL;
1062 }
1063 return 0;
1064 }
1065
1066 static noinline struct kmem_cache_node *free_debug_processing(
1067 struct kmem_cache *s, struct page *page, void *object,
1068 unsigned long addr, unsigned long *flags)
1069 {
1070 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1071
1072 spin_lock_irqsave(&n->list_lock, *flags);
1073 slab_lock(page);
1074
1075 if (!check_slab(s, page))
1076 goto fail;
1077
1078 if (!check_valid_pointer(s, page, object)) {
1079 slab_err(s, page, "Invalid object pointer 0x%p", object);
1080 goto fail;
1081 }
1082
1083 if (on_freelist(s, page, object)) {
1084 object_err(s, page, object, "Object already free");
1085 goto fail;
1086 }
1087
1088 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1089 goto out;
1090
1091 if (unlikely(s != page->slab_cache)) {
1092 if (!PageSlab(page)) {
1093 slab_err(s, page, "Attempt to free object(0x%p) "
1094 "outside of slab", object);
1095 } else if (!page->slab_cache) {
1096 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1097 object);
1098 dump_stack();
1099 } else
1100 object_err(s, page, object,
1101 "page slab pointer corrupt.");
1102 goto fail;
1103 }
1104
1105 if (s->flags & SLAB_STORE_USER)
1106 set_track(s, object, TRACK_FREE, addr);
1107 trace(s, page, object, 0);
1108 init_object(s, object, SLUB_RED_INACTIVE);
1109 out:
1110 slab_unlock(page);
1111 /*
1112 * Keep node_lock to preserve integrity
1113 * until the object is actually freed
1114 */
1115 return n;
1116
1117 fail:
1118 slab_unlock(page);
1119 spin_unlock_irqrestore(&n->list_lock, *flags);
1120 slab_fix(s, "Object at 0x%p not freed", object);
1121 return NULL;
1122 }
1123
1124 static int __init setup_slub_debug(char *str)
1125 {
1126 slub_debug = DEBUG_DEFAULT_FLAGS;
1127 if (*str++ != '=' || !*str)
1128 /*
1129 * No options specified. Switch on full debugging.
1130 */
1131 goto out;
1132
1133 if (*str == ',')
1134 /*
1135 * No options but restriction on slabs. This means full
1136 * debugging for slabs matching a pattern.
1137 */
1138 goto check_slabs;
1139
1140 slub_debug = 0;
1141 if (*str == '-')
1142 /*
1143 * Switch off all debugging measures.
1144 */
1145 goto out;
1146
1147 /*
1148 * Determine which debug features should be switched on
1149 */
1150 for (; *str && *str != ','; str++) {
1151 switch (tolower(*str)) {
1152 case 'f':
1153 slub_debug |= SLAB_DEBUG_FREE;
1154 break;
1155 case 'z':
1156 slub_debug |= SLAB_RED_ZONE;
1157 break;
1158 case 'p':
1159 slub_debug |= SLAB_POISON;
1160 break;
1161 case 'u':
1162 slub_debug |= SLAB_STORE_USER;
1163 break;
1164 case 't':
1165 slub_debug |= SLAB_TRACE;
1166 break;
1167 case 'a':
1168 slub_debug |= SLAB_FAILSLAB;
1169 break;
1170 case 'o':
1171 /*
1172 * Avoid enabling debugging on caches if its minimum
1173 * order would increase as a result.
1174 */
1175 disable_higher_order_debug = 1;
1176 break;
1177 default:
1178 pr_err("slub_debug option '%c' unknown. skipped\n",
1179 *str);
1180 }
1181 }
1182
1183 check_slabs:
1184 if (*str == ',')
1185 slub_debug_slabs = str + 1;
1186 out:
1187 return 1;
1188 }
1189
1190 __setup("slub_debug", setup_slub_debug);
1191
1192 unsigned long kmem_cache_flags(unsigned long object_size,
1193 unsigned long flags, const char *name,
1194 void (*ctor)(void *))
1195 {
1196 /*
1197 * Enable debugging if selected on the kernel commandline.
1198 */
1199 if (slub_debug && (!slub_debug_slabs || (name &&
1200 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1201 flags |= slub_debug;
1202
1203 return flags;
1204 }
1205 #else
1206 static inline void setup_object_debug(struct kmem_cache *s,
1207 struct page *page, void *object) {}
1208
1209 static inline int alloc_debug_processing(struct kmem_cache *s,
1210 struct page *page, void *object, unsigned long addr) { return 0; }
1211
1212 static inline struct kmem_cache_node *free_debug_processing(
1213 struct kmem_cache *s, struct page *page, void *object,
1214 unsigned long addr, unsigned long *flags) { return NULL; }
1215
1216 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1217 { return 1; }
1218 static inline int check_object(struct kmem_cache *s, struct page *page,
1219 void *object, u8 val) { return 1; }
1220 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1221 struct page *page) {}
1222 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1223 struct page *page) {}
1224 unsigned long kmem_cache_flags(unsigned long object_size,
1225 unsigned long flags, const char *name,
1226 void (*ctor)(void *))
1227 {
1228 return flags;
1229 }
1230 #define slub_debug 0
1231
1232 #define disable_higher_order_debug 0
1233
1234 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1235 { return 0; }
1236 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1237 { return 0; }
1238 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1239 int objects) {}
1240 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1241 int objects) {}
1242
1243 #endif /* CONFIG_SLUB_DEBUG */
1244
1245 /*
1246 * Hooks for other subsystems that check memory allocations. In a typical
1247 * production configuration these hooks all should produce no code at all.
1248 */
1249 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1250 {
1251 kmemleak_alloc(ptr, size, 1, flags);
1252 kasan_kmalloc_large(ptr, size);
1253 }
1254
1255 static inline void kfree_hook(const void *x)
1256 {
1257 kmemleak_free(x);
1258 kasan_kfree_large(x);
1259 }
1260
1261 static inline struct kmem_cache *slab_pre_alloc_hook(struct kmem_cache *s,
1262 gfp_t flags)
1263 {
1264 flags &= gfp_allowed_mask;
1265 lockdep_trace_alloc(flags);
1266 might_sleep_if(flags & __GFP_WAIT);
1267
1268 if (should_failslab(s->object_size, flags, s->flags))
1269 return NULL;
1270
1271 return memcg_kmem_get_cache(s, flags);
1272 }
1273
1274 static inline void slab_post_alloc_hook(struct kmem_cache *s,
1275 gfp_t flags, void *object)
1276 {
1277 flags &= gfp_allowed_mask;
1278 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
1279 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
1280 memcg_kmem_put_cache(s);
1281 kasan_slab_alloc(s, object);
1282 }
1283
1284 static inline void slab_free_hook(struct kmem_cache *s, void *x)
1285 {
1286 kmemleak_free_recursive(x, s->flags);
1287
1288 /*
1289 * Trouble is that we may no longer disable interrupts in the fast path
1290 * So in order to make the debug calls that expect irqs to be
1291 * disabled we need to disable interrupts temporarily.
1292 */
1293 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1294 {
1295 unsigned long flags;
1296
1297 local_irq_save(flags);
1298 kmemcheck_slab_free(s, x, s->object_size);
1299 debug_check_no_locks_freed(x, s->object_size);
1300 local_irq_restore(flags);
1301 }
1302 #endif
1303 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1304 debug_check_no_obj_freed(x, s->object_size);
1305
1306 kasan_slab_free(s, x);
1307 }
1308
1309 /*
1310 * Slab allocation and freeing
1311 */
1312 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1313 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1314 {
1315 struct page *page;
1316 int order = oo_order(oo);
1317
1318 flags |= __GFP_NOTRACK;
1319
1320 if (memcg_charge_slab(s, flags, order))
1321 return NULL;
1322
1323 if (node == NUMA_NO_NODE)
1324 page = alloc_pages(flags, order);
1325 else
1326 page = alloc_pages_exact_node(node, flags, order);
1327
1328 if (!page)
1329 memcg_uncharge_slab(s, order);
1330
1331 return page;
1332 }
1333
1334 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1335 {
1336 struct page *page;
1337 struct kmem_cache_order_objects oo = s->oo;
1338 gfp_t alloc_gfp;
1339
1340 flags &= gfp_allowed_mask;
1341
1342 if (flags & __GFP_WAIT)
1343 local_irq_enable();
1344
1345 flags |= s->allocflags;
1346
1347 /*
1348 * Let the initial higher-order allocation fail under memory pressure
1349 * so we fall-back to the minimum order allocation.
1350 */
1351 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1352
1353 page = alloc_slab_page(s, alloc_gfp, node, oo);
1354 if (unlikely(!page)) {
1355 oo = s->min;
1356 alloc_gfp = flags;
1357 /*
1358 * Allocation may have failed due to fragmentation.
1359 * Try a lower order alloc if possible
1360 */
1361 page = alloc_slab_page(s, alloc_gfp, node, oo);
1362
1363 if (page)
1364 stat(s, ORDER_FALLBACK);
1365 }
1366
1367 if (kmemcheck_enabled && page
1368 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1369 int pages = 1 << oo_order(oo);
1370
1371 kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node);
1372
1373 /*
1374 * Objects from caches that have a constructor don't get
1375 * cleared when they're allocated, so we need to do it here.
1376 */
1377 if (s->ctor)
1378 kmemcheck_mark_uninitialized_pages(page, pages);
1379 else
1380 kmemcheck_mark_unallocated_pages(page, pages);
1381 }
1382
1383 if (flags & __GFP_WAIT)
1384 local_irq_disable();
1385 if (!page)
1386 return NULL;
1387
1388 page->objects = oo_objects(oo);
1389 mod_zone_page_state(page_zone(page),
1390 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1391 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1392 1 << oo_order(oo));
1393
1394 return page;
1395 }
1396
1397 static void setup_object(struct kmem_cache *s, struct page *page,
1398 void *object)
1399 {
1400 setup_object_debug(s, page, object);
1401 if (unlikely(s->ctor)) {
1402 kasan_unpoison_object_data(s, object);
1403 s->ctor(object);
1404 kasan_poison_object_data(s, object);
1405 }
1406 }
1407
1408 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1409 {
1410 struct page *page;
1411 void *start;
1412 void *p;
1413 int order;
1414 int idx;
1415
1416 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1417 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
1418 BUG();
1419 }
1420
1421 page = allocate_slab(s,
1422 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1423 if (!page)
1424 goto out;
1425
1426 order = compound_order(page);
1427 inc_slabs_node(s, page_to_nid(page), page->objects);
1428 page->slab_cache = s;
1429 __SetPageSlab(page);
1430 if (page_is_pfmemalloc(page))
1431 SetPageSlabPfmemalloc(page);
1432
1433 start = page_address(page);
1434
1435 if (unlikely(s->flags & SLAB_POISON))
1436 memset(start, POISON_INUSE, PAGE_SIZE << order);
1437
1438 kasan_poison_slab(page);
1439
1440 for_each_object_idx(p, idx, s, start, page->objects) {
1441 setup_object(s, page, p);
1442 if (likely(idx < page->objects))
1443 set_freepointer(s, p, p + s->size);
1444 else
1445 set_freepointer(s, p, NULL);
1446 }
1447
1448 page->freelist = start;
1449 page->inuse = page->objects;
1450 page->frozen = 1;
1451 out:
1452 return page;
1453 }
1454
1455 static void __free_slab(struct kmem_cache *s, struct page *page)
1456 {
1457 int order = compound_order(page);
1458 int pages = 1 << order;
1459
1460 if (kmem_cache_debug(s)) {
1461 void *p;
1462
1463 slab_pad_check(s, page);
1464 for_each_object(p, s, page_address(page),
1465 page->objects)
1466 check_object(s, page, p, SLUB_RED_INACTIVE);
1467 }
1468
1469 kmemcheck_free_shadow(page, compound_order(page));
1470
1471 mod_zone_page_state(page_zone(page),
1472 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1473 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1474 -pages);
1475
1476 __ClearPageSlabPfmemalloc(page);
1477 __ClearPageSlab(page);
1478
1479 page_mapcount_reset(page);
1480 if (current->reclaim_state)
1481 current->reclaim_state->reclaimed_slab += pages;
1482 __free_pages(page, order);
1483 memcg_uncharge_slab(s, order);
1484 }
1485
1486 #define need_reserve_slab_rcu \
1487 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1488
1489 static void rcu_free_slab(struct rcu_head *h)
1490 {
1491 struct page *page;
1492
1493 if (need_reserve_slab_rcu)
1494 page = virt_to_head_page(h);
1495 else
1496 page = container_of((struct list_head *)h, struct page, lru);
1497
1498 __free_slab(page->slab_cache, page);
1499 }
1500
1501 static void free_slab(struct kmem_cache *s, struct page *page)
1502 {
1503 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1504 struct rcu_head *head;
1505
1506 if (need_reserve_slab_rcu) {
1507 int order = compound_order(page);
1508 int offset = (PAGE_SIZE << order) - s->reserved;
1509
1510 VM_BUG_ON(s->reserved != sizeof(*head));
1511 head = page_address(page) + offset;
1512 } else {
1513 /*
1514 * RCU free overloads the RCU head over the LRU
1515 */
1516 head = (void *)&page->lru;
1517 }
1518
1519 call_rcu(head, rcu_free_slab);
1520 } else
1521 __free_slab(s, page);
1522 }
1523
1524 static void discard_slab(struct kmem_cache *s, struct page *page)
1525 {
1526 dec_slabs_node(s, page_to_nid(page), page->objects);
1527 free_slab(s, page);
1528 }
1529
1530 /*
1531 * Management of partially allocated slabs.
1532 */
1533 static inline void
1534 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1535 {
1536 n->nr_partial++;
1537 if (tail == DEACTIVATE_TO_TAIL)
1538 list_add_tail(&page->lru, &n->partial);
1539 else
1540 list_add(&page->lru, &n->partial);
1541 }
1542
1543 static inline void add_partial(struct kmem_cache_node *n,
1544 struct page *page, int tail)
1545 {
1546 lockdep_assert_held(&n->list_lock);
1547 __add_partial(n, page, tail);
1548 }
1549
1550 static inline void
1551 __remove_partial(struct kmem_cache_node *n, struct page *page)
1552 {
1553 list_del(&page->lru);
1554 n->nr_partial--;
1555 }
1556
1557 static inline void remove_partial(struct kmem_cache_node *n,
1558 struct page *page)
1559 {
1560 lockdep_assert_held(&n->list_lock);
1561 __remove_partial(n, page);
1562 }
1563
1564 /*
1565 * Remove slab from the partial list, freeze it and
1566 * return the pointer to the freelist.
1567 *
1568 * Returns a list of objects or NULL if it fails.
1569 */
1570 static inline void *acquire_slab(struct kmem_cache *s,
1571 struct kmem_cache_node *n, struct page *page,
1572 int mode, int *objects)
1573 {
1574 void *freelist;
1575 unsigned long counters;
1576 struct page new;
1577
1578 lockdep_assert_held(&n->list_lock);
1579
1580 /*
1581 * Zap the freelist and set the frozen bit.
1582 * The old freelist is the list of objects for the
1583 * per cpu allocation list.
1584 */
1585 freelist = page->freelist;
1586 counters = page->counters;
1587 new.counters = counters;
1588 *objects = new.objects - new.inuse;
1589 if (mode) {
1590 new.inuse = page->objects;
1591 new.freelist = NULL;
1592 } else {
1593 new.freelist = freelist;
1594 }
1595
1596 VM_BUG_ON(new.frozen);
1597 new.frozen = 1;
1598
1599 if (!__cmpxchg_double_slab(s, page,
1600 freelist, counters,
1601 new.freelist, new.counters,
1602 "acquire_slab"))
1603 return NULL;
1604
1605 remove_partial(n, page);
1606 WARN_ON(!freelist);
1607 return freelist;
1608 }
1609
1610 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1611 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1612
1613 /*
1614 * Try to allocate a partial slab from a specific node.
1615 */
1616 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1617 struct kmem_cache_cpu *c, gfp_t flags)
1618 {
1619 struct page *page, *page2;
1620 void *object = NULL;
1621 int available = 0;
1622 int objects;
1623
1624 /*
1625 * Racy check. If we mistakenly see no partial slabs then we
1626 * just allocate an empty slab. If we mistakenly try to get a
1627 * partial slab and there is none available then get_partials()
1628 * will return NULL.
1629 */
1630 if (!n || !n->nr_partial)
1631 return NULL;
1632
1633 spin_lock(&n->list_lock);
1634 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1635 void *t;
1636
1637 if (!pfmemalloc_match(page, flags))
1638 continue;
1639
1640 t = acquire_slab(s, n, page, object == NULL, &objects);
1641 if (!t)
1642 break;
1643
1644 available += objects;
1645 if (!object) {
1646 c->page = page;
1647 stat(s, ALLOC_FROM_PARTIAL);
1648 object = t;
1649 } else {
1650 put_cpu_partial(s, page, 0);
1651 stat(s, CPU_PARTIAL_NODE);
1652 }
1653 if (!kmem_cache_has_cpu_partial(s)
1654 || available > s->cpu_partial / 2)
1655 break;
1656
1657 }
1658 spin_unlock(&n->list_lock);
1659 return object;
1660 }
1661
1662 /*
1663 * Get a page from somewhere. Search in increasing NUMA distances.
1664 */
1665 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1666 struct kmem_cache_cpu *c)
1667 {
1668 #ifdef CONFIG_NUMA
1669 struct zonelist *zonelist;
1670 struct zoneref *z;
1671 struct zone *zone;
1672 enum zone_type high_zoneidx = gfp_zone(flags);
1673 void *object;
1674 unsigned int cpuset_mems_cookie;
1675
1676 /*
1677 * The defrag ratio allows a configuration of the tradeoffs between
1678 * inter node defragmentation and node local allocations. A lower
1679 * defrag_ratio increases the tendency to do local allocations
1680 * instead of attempting to obtain partial slabs from other nodes.
1681 *
1682 * If the defrag_ratio is set to 0 then kmalloc() always
1683 * returns node local objects. If the ratio is higher then kmalloc()
1684 * may return off node objects because partial slabs are obtained
1685 * from other nodes and filled up.
1686 *
1687 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1688 * defrag_ratio = 1000) then every (well almost) allocation will
1689 * first attempt to defrag slab caches on other nodes. This means
1690 * scanning over all nodes to look for partial slabs which may be
1691 * expensive if we do it every time we are trying to find a slab
1692 * with available objects.
1693 */
1694 if (!s->remote_node_defrag_ratio ||
1695 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1696 return NULL;
1697
1698 do {
1699 cpuset_mems_cookie = read_mems_allowed_begin();
1700 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1701 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1702 struct kmem_cache_node *n;
1703
1704 n = get_node(s, zone_to_nid(zone));
1705
1706 if (n && cpuset_zone_allowed(zone, flags) &&
1707 n->nr_partial > s->min_partial) {
1708 object = get_partial_node(s, n, c, flags);
1709 if (object) {
1710 /*
1711 * Don't check read_mems_allowed_retry()
1712 * here - if mems_allowed was updated in
1713 * parallel, that was a harmless race
1714 * between allocation and the cpuset
1715 * update
1716 */
1717 return object;
1718 }
1719 }
1720 }
1721 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1722 #endif
1723 return NULL;
1724 }
1725
1726 /*
1727 * Get a partial page, lock it and return it.
1728 */
1729 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1730 struct kmem_cache_cpu *c)
1731 {
1732 void *object;
1733 int searchnode = node;
1734
1735 if (node == NUMA_NO_NODE)
1736 searchnode = numa_mem_id();
1737 else if (!node_present_pages(node))
1738 searchnode = node_to_mem_node(node);
1739
1740 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1741 if (object || node != NUMA_NO_NODE)
1742 return object;
1743
1744 return get_any_partial(s, flags, c);
1745 }
1746
1747 #ifdef CONFIG_PREEMPT
1748 /*
1749 * Calculate the next globally unique transaction for disambiguiation
1750 * during cmpxchg. The transactions start with the cpu number and are then
1751 * incremented by CONFIG_NR_CPUS.
1752 */
1753 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1754 #else
1755 /*
1756 * No preemption supported therefore also no need to check for
1757 * different cpus.
1758 */
1759 #define TID_STEP 1
1760 #endif
1761
1762 static inline unsigned long next_tid(unsigned long tid)
1763 {
1764 return tid + TID_STEP;
1765 }
1766
1767 static inline unsigned int tid_to_cpu(unsigned long tid)
1768 {
1769 return tid % TID_STEP;
1770 }
1771
1772 static inline unsigned long tid_to_event(unsigned long tid)
1773 {
1774 return tid / TID_STEP;
1775 }
1776
1777 static inline unsigned int init_tid(int cpu)
1778 {
1779 return cpu;
1780 }
1781
1782 static inline void note_cmpxchg_failure(const char *n,
1783 const struct kmem_cache *s, unsigned long tid)
1784 {
1785 #ifdef SLUB_DEBUG_CMPXCHG
1786 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1787
1788 pr_info("%s %s: cmpxchg redo ", n, s->name);
1789
1790 #ifdef CONFIG_PREEMPT
1791 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1792 pr_warn("due to cpu change %d -> %d\n",
1793 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1794 else
1795 #endif
1796 if (tid_to_event(tid) != tid_to_event(actual_tid))
1797 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1798 tid_to_event(tid), tid_to_event(actual_tid));
1799 else
1800 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1801 actual_tid, tid, next_tid(tid));
1802 #endif
1803 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1804 }
1805
1806 static void init_kmem_cache_cpus(struct kmem_cache *s)
1807 {
1808 int cpu;
1809
1810 for_each_possible_cpu(cpu)
1811 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1812 }
1813
1814 /*
1815 * Remove the cpu slab
1816 */
1817 static void deactivate_slab(struct kmem_cache *s, struct page *page,
1818 void *freelist)
1819 {
1820 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1821 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1822 int lock = 0;
1823 enum slab_modes l = M_NONE, m = M_NONE;
1824 void *nextfree;
1825 int tail = DEACTIVATE_TO_HEAD;
1826 struct page new;
1827 struct page old;
1828
1829 if (page->freelist) {
1830 stat(s, DEACTIVATE_REMOTE_FREES);
1831 tail = DEACTIVATE_TO_TAIL;
1832 }
1833
1834 /*
1835 * Stage one: Free all available per cpu objects back
1836 * to the page freelist while it is still frozen. Leave the
1837 * last one.
1838 *
1839 * There is no need to take the list->lock because the page
1840 * is still frozen.
1841 */
1842 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1843 void *prior;
1844 unsigned long counters;
1845
1846 do {
1847 prior = page->freelist;
1848 counters = page->counters;
1849 set_freepointer(s, freelist, prior);
1850 new.counters = counters;
1851 new.inuse--;
1852 VM_BUG_ON(!new.frozen);
1853
1854 } while (!__cmpxchg_double_slab(s, page,
1855 prior, counters,
1856 freelist, new.counters,
1857 "drain percpu freelist"));
1858
1859 freelist = nextfree;
1860 }
1861
1862 /*
1863 * Stage two: Ensure that the page is unfrozen while the
1864 * list presence reflects the actual number of objects
1865 * during unfreeze.
1866 *
1867 * We setup the list membership and then perform a cmpxchg
1868 * with the count. If there is a mismatch then the page
1869 * is not unfrozen but the page is on the wrong list.
1870 *
1871 * Then we restart the process which may have to remove
1872 * the page from the list that we just put it on again
1873 * because the number of objects in the slab may have
1874 * changed.
1875 */
1876 redo:
1877
1878 old.freelist = page->freelist;
1879 old.counters = page->counters;
1880 VM_BUG_ON(!old.frozen);
1881
1882 /* Determine target state of the slab */
1883 new.counters = old.counters;
1884 if (freelist) {
1885 new.inuse--;
1886 set_freepointer(s, freelist, old.freelist);
1887 new.freelist = freelist;
1888 } else
1889 new.freelist = old.freelist;
1890
1891 new.frozen = 0;
1892
1893 if (!new.inuse && n->nr_partial >= s->min_partial)
1894 m = M_FREE;
1895 else if (new.freelist) {
1896 m = M_PARTIAL;
1897 if (!lock) {
1898 lock = 1;
1899 /*
1900 * Taking the spinlock removes the possiblity
1901 * that acquire_slab() will see a slab page that
1902 * is frozen
1903 */
1904 spin_lock(&n->list_lock);
1905 }
1906 } else {
1907 m = M_FULL;
1908 if (kmem_cache_debug(s) && !lock) {
1909 lock = 1;
1910 /*
1911 * This also ensures that the scanning of full
1912 * slabs from diagnostic functions will not see
1913 * any frozen slabs.
1914 */
1915 spin_lock(&n->list_lock);
1916 }
1917 }
1918
1919 if (l != m) {
1920
1921 if (l == M_PARTIAL)
1922
1923 remove_partial(n, page);
1924
1925 else if (l == M_FULL)
1926
1927 remove_full(s, n, page);
1928
1929 if (m == M_PARTIAL) {
1930
1931 add_partial(n, page, tail);
1932 stat(s, tail);
1933
1934 } else if (m == M_FULL) {
1935
1936 stat(s, DEACTIVATE_FULL);
1937 add_full(s, n, page);
1938
1939 }
1940 }
1941
1942 l = m;
1943 if (!__cmpxchg_double_slab(s, page,
1944 old.freelist, old.counters,
1945 new.freelist, new.counters,
1946 "unfreezing slab"))
1947 goto redo;
1948
1949 if (lock)
1950 spin_unlock(&n->list_lock);
1951
1952 if (m == M_FREE) {
1953 stat(s, DEACTIVATE_EMPTY);
1954 discard_slab(s, page);
1955 stat(s, FREE_SLAB);
1956 }
1957 }
1958
1959 /*
1960 * Unfreeze all the cpu partial slabs.
1961 *
1962 * This function must be called with interrupts disabled
1963 * for the cpu using c (or some other guarantee must be there
1964 * to guarantee no concurrent accesses).
1965 */
1966 static void unfreeze_partials(struct kmem_cache *s,
1967 struct kmem_cache_cpu *c)
1968 {
1969 #ifdef CONFIG_SLUB_CPU_PARTIAL
1970 struct kmem_cache_node *n = NULL, *n2 = NULL;
1971 struct page *page, *discard_page = NULL;
1972
1973 while ((page = c->partial)) {
1974 struct page new;
1975 struct page old;
1976
1977 c->partial = page->next;
1978
1979 n2 = get_node(s, page_to_nid(page));
1980 if (n != n2) {
1981 if (n)
1982 spin_unlock(&n->list_lock);
1983
1984 n = n2;
1985 spin_lock(&n->list_lock);
1986 }
1987
1988 do {
1989
1990 old.freelist = page->freelist;
1991 old.counters = page->counters;
1992 VM_BUG_ON(!old.frozen);
1993
1994 new.counters = old.counters;
1995 new.freelist = old.freelist;
1996
1997 new.frozen = 0;
1998
1999 } while (!__cmpxchg_double_slab(s, page,
2000 old.freelist, old.counters,
2001 new.freelist, new.counters,
2002 "unfreezing slab"));
2003
2004 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2005 page->next = discard_page;
2006 discard_page = page;
2007 } else {
2008 add_partial(n, page, DEACTIVATE_TO_TAIL);
2009 stat(s, FREE_ADD_PARTIAL);
2010 }
2011 }
2012
2013 if (n)
2014 spin_unlock(&n->list_lock);
2015
2016 while (discard_page) {
2017 page = discard_page;
2018 discard_page = discard_page->next;
2019
2020 stat(s, DEACTIVATE_EMPTY);
2021 discard_slab(s, page);
2022 stat(s, FREE_SLAB);
2023 }
2024 #endif
2025 }
2026
2027 /*
2028 * Put a page that was just frozen (in __slab_free) into a partial page
2029 * slot if available. This is done without interrupts disabled and without
2030 * preemption disabled. The cmpxchg is racy and may put the partial page
2031 * onto a random cpus partial slot.
2032 *
2033 * If we did not find a slot then simply move all the partials to the
2034 * per node partial list.
2035 */
2036 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2037 {
2038 #ifdef CONFIG_SLUB_CPU_PARTIAL
2039 struct page *oldpage;
2040 int pages;
2041 int pobjects;
2042
2043 preempt_disable();
2044 do {
2045 pages = 0;
2046 pobjects = 0;
2047 oldpage = this_cpu_read(s->cpu_slab->partial);
2048
2049 if (oldpage) {
2050 pobjects = oldpage->pobjects;
2051 pages = oldpage->pages;
2052 if (drain && pobjects > s->cpu_partial) {
2053 unsigned long flags;
2054 /*
2055 * partial array is full. Move the existing
2056 * set to the per node partial list.
2057 */
2058 local_irq_save(flags);
2059 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2060 local_irq_restore(flags);
2061 oldpage = NULL;
2062 pobjects = 0;
2063 pages = 0;
2064 stat(s, CPU_PARTIAL_DRAIN);
2065 }
2066 }
2067
2068 pages++;
2069 pobjects += page->objects - page->inuse;
2070
2071 page->pages = pages;
2072 page->pobjects = pobjects;
2073 page->next = oldpage;
2074
2075 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2076 != oldpage);
2077 if (unlikely(!s->cpu_partial)) {
2078 unsigned long flags;
2079
2080 local_irq_save(flags);
2081 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2082 local_irq_restore(flags);
2083 }
2084 preempt_enable();
2085 #endif
2086 }
2087
2088 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2089 {
2090 stat(s, CPUSLAB_FLUSH);
2091 deactivate_slab(s, c->page, c->freelist);
2092
2093 c->tid = next_tid(c->tid);
2094 c->page = NULL;
2095 c->freelist = NULL;
2096 }
2097
2098 /*
2099 * Flush cpu slab.
2100 *
2101 * Called from IPI handler with interrupts disabled.
2102 */
2103 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2104 {
2105 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2106
2107 if (likely(c)) {
2108 if (c->page)
2109 flush_slab(s, c);
2110
2111 unfreeze_partials(s, c);
2112 }
2113 }
2114
2115 static void flush_cpu_slab(void *d)
2116 {
2117 struct kmem_cache *s = d;
2118
2119 __flush_cpu_slab(s, smp_processor_id());
2120 }
2121
2122 static bool has_cpu_slab(int cpu, void *info)
2123 {
2124 struct kmem_cache *s = info;
2125 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2126
2127 return c->page || c->partial;
2128 }
2129
2130 static void flush_all(struct kmem_cache *s)
2131 {
2132 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2133 }
2134
2135 /*
2136 * Check if the objects in a per cpu structure fit numa
2137 * locality expectations.
2138 */
2139 static inline int node_match(struct page *page, int node)
2140 {
2141 #ifdef CONFIG_NUMA
2142 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2143 return 0;
2144 #endif
2145 return 1;
2146 }
2147
2148 #ifdef CONFIG_SLUB_DEBUG
2149 static int count_free(struct page *page)
2150 {
2151 return page->objects - page->inuse;
2152 }
2153
2154 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2155 {
2156 return atomic_long_read(&n->total_objects);
2157 }
2158 #endif /* CONFIG_SLUB_DEBUG */
2159
2160 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2161 static unsigned long count_partial(struct kmem_cache_node *n,
2162 int (*get_count)(struct page *))
2163 {
2164 unsigned long flags;
2165 unsigned long x = 0;
2166 struct page *page;
2167
2168 spin_lock_irqsave(&n->list_lock, flags);
2169 list_for_each_entry(page, &n->partial, lru)
2170 x += get_count(page);
2171 spin_unlock_irqrestore(&n->list_lock, flags);
2172 return x;
2173 }
2174 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2175
2176 static noinline void
2177 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2178 {
2179 #ifdef CONFIG_SLUB_DEBUG
2180 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2181 DEFAULT_RATELIMIT_BURST);
2182 int node;
2183 struct kmem_cache_node *n;
2184
2185 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2186 return;
2187
2188 pr_warn("SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2189 nid, gfpflags);
2190 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2191 s->name, s->object_size, s->size, oo_order(s->oo),
2192 oo_order(s->min));
2193
2194 if (oo_order(s->min) > get_order(s->object_size))
2195 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2196 s->name);
2197
2198 for_each_kmem_cache_node(s, node, n) {
2199 unsigned long nr_slabs;
2200 unsigned long nr_objs;
2201 unsigned long nr_free;
2202
2203 nr_free = count_partial(n, count_free);
2204 nr_slabs = node_nr_slabs(n);
2205 nr_objs = node_nr_objs(n);
2206
2207 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2208 node, nr_slabs, nr_objs, nr_free);
2209 }
2210 #endif
2211 }
2212
2213 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2214 int node, struct kmem_cache_cpu **pc)
2215 {
2216 void *freelist;
2217 struct kmem_cache_cpu *c = *pc;
2218 struct page *page;
2219
2220 freelist = get_partial(s, flags, node, c);
2221
2222 if (freelist)
2223 return freelist;
2224
2225 page = new_slab(s, flags, node);
2226 if (page) {
2227 c = raw_cpu_ptr(s->cpu_slab);
2228 if (c->page)
2229 flush_slab(s, c);
2230
2231 /*
2232 * No other reference to the page yet so we can
2233 * muck around with it freely without cmpxchg
2234 */
2235 freelist = page->freelist;
2236 page->freelist = NULL;
2237
2238 stat(s, ALLOC_SLAB);
2239 c->page = page;
2240 *pc = c;
2241 } else
2242 freelist = NULL;
2243
2244 return freelist;
2245 }
2246
2247 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2248 {
2249 if (unlikely(PageSlabPfmemalloc(page)))
2250 return gfp_pfmemalloc_allowed(gfpflags);
2251
2252 return true;
2253 }
2254
2255 /*
2256 * Check the page->freelist of a page and either transfer the freelist to the
2257 * per cpu freelist or deactivate the page.
2258 *
2259 * The page is still frozen if the return value is not NULL.
2260 *
2261 * If this function returns NULL then the page has been unfrozen.
2262 *
2263 * This function must be called with interrupt disabled.
2264 */
2265 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2266 {
2267 struct page new;
2268 unsigned long counters;
2269 void *freelist;
2270
2271 do {
2272 freelist = page->freelist;
2273 counters = page->counters;
2274
2275 new.counters = counters;
2276 VM_BUG_ON(!new.frozen);
2277
2278 new.inuse = page->objects;
2279 new.frozen = freelist != NULL;
2280
2281 } while (!__cmpxchg_double_slab(s, page,
2282 freelist, counters,
2283 NULL, new.counters,
2284 "get_freelist"));
2285
2286 return freelist;
2287 }
2288
2289 /*
2290 * Slow path. The lockless freelist is empty or we need to perform
2291 * debugging duties.
2292 *
2293 * Processing is still very fast if new objects have been freed to the
2294 * regular freelist. In that case we simply take over the regular freelist
2295 * as the lockless freelist and zap the regular freelist.
2296 *
2297 * If that is not working then we fall back to the partial lists. We take the
2298 * first element of the freelist as the object to allocate now and move the
2299 * rest of the freelist to the lockless freelist.
2300 *
2301 * And if we were unable to get a new slab from the partial slab lists then
2302 * we need to allocate a new slab. This is the slowest path since it involves
2303 * a call to the page allocator and the setup of a new slab.
2304 */
2305 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2306 unsigned long addr, struct kmem_cache_cpu *c)
2307 {
2308 void *freelist;
2309 struct page *page;
2310 unsigned long flags;
2311
2312 local_irq_save(flags);
2313 #ifdef CONFIG_PREEMPT
2314 /*
2315 * We may have been preempted and rescheduled on a different
2316 * cpu before disabling interrupts. Need to reload cpu area
2317 * pointer.
2318 */
2319 c = this_cpu_ptr(s->cpu_slab);
2320 #endif
2321
2322 page = c->page;
2323 if (!page)
2324 goto new_slab;
2325 redo:
2326
2327 if (unlikely(!node_match(page, node))) {
2328 int searchnode = node;
2329
2330 if (node != NUMA_NO_NODE && !node_present_pages(node))
2331 searchnode = node_to_mem_node(node);
2332
2333 if (unlikely(!node_match(page, searchnode))) {
2334 stat(s, ALLOC_NODE_MISMATCH);
2335 deactivate_slab(s, page, c->freelist);
2336 c->page = NULL;
2337 c->freelist = NULL;
2338 goto new_slab;
2339 }
2340 }
2341
2342 /*
2343 * By rights, we should be searching for a slab page that was
2344 * PFMEMALLOC but right now, we are losing the pfmemalloc
2345 * information when the page leaves the per-cpu allocator
2346 */
2347 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2348 deactivate_slab(s, page, c->freelist);
2349 c->page = NULL;
2350 c->freelist = NULL;
2351 goto new_slab;
2352 }
2353
2354 /* must check again c->freelist in case of cpu migration or IRQ */
2355 freelist = c->freelist;
2356 if (freelist)
2357 goto load_freelist;
2358
2359 freelist = get_freelist(s, page);
2360
2361 if (!freelist) {
2362 c->page = NULL;
2363 stat(s, DEACTIVATE_BYPASS);
2364 goto new_slab;
2365 }
2366
2367 stat(s, ALLOC_REFILL);
2368
2369 load_freelist:
2370 /*
2371 * freelist is pointing to the list of objects to be used.
2372 * page is pointing to the page from which the objects are obtained.
2373 * That page must be frozen for per cpu allocations to work.
2374 */
2375 VM_BUG_ON(!c->page->frozen);
2376 c->freelist = get_freepointer(s, freelist);
2377 c->tid = next_tid(c->tid);
2378 local_irq_restore(flags);
2379 return freelist;
2380
2381 new_slab:
2382
2383 if (c->partial) {
2384 page = c->page = c->partial;
2385 c->partial = page->next;
2386 stat(s, CPU_PARTIAL_ALLOC);
2387 c->freelist = NULL;
2388 goto redo;
2389 }
2390
2391 freelist = new_slab_objects(s, gfpflags, node, &c);
2392
2393 if (unlikely(!freelist)) {
2394 slab_out_of_memory(s, gfpflags, node);
2395 local_irq_restore(flags);
2396 return NULL;
2397 }
2398
2399 page = c->page;
2400 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2401 goto load_freelist;
2402
2403 /* Only entered in the debug case */
2404 if (kmem_cache_debug(s) &&
2405 !alloc_debug_processing(s, page, freelist, addr))
2406 goto new_slab; /* Slab failed checks. Next slab needed */
2407
2408 deactivate_slab(s, page, get_freepointer(s, freelist));
2409 c->page = NULL;
2410 c->freelist = NULL;
2411 local_irq_restore(flags);
2412 return freelist;
2413 }
2414
2415 /*
2416 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2417 * have the fastpath folded into their functions. So no function call
2418 * overhead for requests that can be satisfied on the fastpath.
2419 *
2420 * The fastpath works by first checking if the lockless freelist can be used.
2421 * If not then __slab_alloc is called for slow processing.
2422 *
2423 * Otherwise we can simply pick the next object from the lockless free list.
2424 */
2425 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2426 gfp_t gfpflags, int node, unsigned long addr)
2427 {
2428 void **object;
2429 struct kmem_cache_cpu *c;
2430 struct page *page;
2431 unsigned long tid;
2432
2433 s = slab_pre_alloc_hook(s, gfpflags);
2434 if (!s)
2435 return NULL;
2436 redo:
2437 /*
2438 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2439 * enabled. We may switch back and forth between cpus while
2440 * reading from one cpu area. That does not matter as long
2441 * as we end up on the original cpu again when doing the cmpxchg.
2442 *
2443 * We should guarantee that tid and kmem_cache are retrieved on
2444 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2445 * to check if it is matched or not.
2446 */
2447 do {
2448 tid = this_cpu_read(s->cpu_slab->tid);
2449 c = raw_cpu_ptr(s->cpu_slab);
2450 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2451 unlikely(tid != READ_ONCE(c->tid)));
2452
2453 /*
2454 * Irqless object alloc/free algorithm used here depends on sequence
2455 * of fetching cpu_slab's data. tid should be fetched before anything
2456 * on c to guarantee that object and page associated with previous tid
2457 * won't be used with current tid. If we fetch tid first, object and
2458 * page could be one associated with next tid and our alloc/free
2459 * request will be failed. In this case, we will retry. So, no problem.
2460 */
2461 barrier();
2462
2463 /*
2464 * The transaction ids are globally unique per cpu and per operation on
2465 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2466 * occurs on the right processor and that there was no operation on the
2467 * linked list in between.
2468 */
2469
2470 object = c->freelist;
2471 page = c->page;
2472 if (unlikely(!object || !node_match(page, node))) {
2473 object = __slab_alloc(s, gfpflags, node, addr, c);
2474 stat(s, ALLOC_SLOWPATH);
2475 } else {
2476 void *next_object = get_freepointer_safe(s, object);
2477
2478 /*
2479 * The cmpxchg will only match if there was no additional
2480 * operation and if we are on the right processor.
2481 *
2482 * The cmpxchg does the following atomically (without lock
2483 * semantics!)
2484 * 1. Relocate first pointer to the current per cpu area.
2485 * 2. Verify that tid and freelist have not been changed
2486 * 3. If they were not changed replace tid and freelist
2487 *
2488 * Since this is without lock semantics the protection is only
2489 * against code executing on this cpu *not* from access by
2490 * other cpus.
2491 */
2492 if (unlikely(!this_cpu_cmpxchg_double(
2493 s->cpu_slab->freelist, s->cpu_slab->tid,
2494 object, tid,
2495 next_object, next_tid(tid)))) {
2496
2497 note_cmpxchg_failure("slab_alloc", s, tid);
2498 goto redo;
2499 }
2500 prefetch_freepointer(s, next_object);
2501 stat(s, ALLOC_FASTPATH);
2502 }
2503
2504 if (unlikely(gfpflags & __GFP_ZERO) && object)
2505 memset(object, 0, s->object_size);
2506
2507 slab_post_alloc_hook(s, gfpflags, object);
2508
2509 return object;
2510 }
2511
2512 static __always_inline void *slab_alloc(struct kmem_cache *s,
2513 gfp_t gfpflags, unsigned long addr)
2514 {
2515 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2516 }
2517
2518 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2519 {
2520 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2521
2522 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2523 s->size, gfpflags);
2524
2525 return ret;
2526 }
2527 EXPORT_SYMBOL(kmem_cache_alloc);
2528
2529 #ifdef CONFIG_TRACING
2530 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2531 {
2532 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2533 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2534 kasan_kmalloc(s, ret, size);
2535 return ret;
2536 }
2537 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2538 #endif
2539
2540 #ifdef CONFIG_NUMA
2541 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2542 {
2543 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2544
2545 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2546 s->object_size, s->size, gfpflags, node);
2547
2548 return ret;
2549 }
2550 EXPORT_SYMBOL(kmem_cache_alloc_node);
2551
2552 #ifdef CONFIG_TRACING
2553 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2554 gfp_t gfpflags,
2555 int node, size_t size)
2556 {
2557 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2558
2559 trace_kmalloc_node(_RET_IP_, ret,
2560 size, s->size, gfpflags, node);
2561
2562 kasan_kmalloc(s, ret, size);
2563 return ret;
2564 }
2565 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2566 #endif
2567 #endif
2568
2569 /*
2570 * Slow path handling. This may still be called frequently since objects
2571 * have a longer lifetime than the cpu slabs in most processing loads.
2572 *
2573 * So we still attempt to reduce cache line usage. Just take the slab
2574 * lock and free the item. If there is no additional partial page
2575 * handling required then we can return immediately.
2576 */
2577 static void __slab_free(struct kmem_cache *s, struct page *page,
2578 void *x, unsigned long addr)
2579 {
2580 void *prior;
2581 void **object = (void *)x;
2582 int was_frozen;
2583 struct page new;
2584 unsigned long counters;
2585 struct kmem_cache_node *n = NULL;
2586 unsigned long uninitialized_var(flags);
2587
2588 stat(s, FREE_SLOWPATH);
2589
2590 if (kmem_cache_debug(s) &&
2591 !(n = free_debug_processing(s, page, x, addr, &flags)))
2592 return;
2593
2594 do {
2595 if (unlikely(n)) {
2596 spin_unlock_irqrestore(&n->list_lock, flags);
2597 n = NULL;
2598 }
2599 prior = page->freelist;
2600 counters = page->counters;
2601 set_freepointer(s, object, prior);
2602 new.counters = counters;
2603 was_frozen = new.frozen;
2604 new.inuse--;
2605 if ((!new.inuse || !prior) && !was_frozen) {
2606
2607 if (kmem_cache_has_cpu_partial(s) && !prior) {
2608
2609 /*
2610 * Slab was on no list before and will be
2611 * partially empty
2612 * We can defer the list move and instead
2613 * freeze it.
2614 */
2615 new.frozen = 1;
2616
2617 } else { /* Needs to be taken off a list */
2618
2619 n = get_node(s, page_to_nid(page));
2620 /*
2621 * Speculatively acquire the list_lock.
2622 * If the cmpxchg does not succeed then we may
2623 * drop the list_lock without any processing.
2624 *
2625 * Otherwise the list_lock will synchronize with
2626 * other processors updating the list of slabs.
2627 */
2628 spin_lock_irqsave(&n->list_lock, flags);
2629
2630 }
2631 }
2632
2633 } while (!cmpxchg_double_slab(s, page,
2634 prior, counters,
2635 object, new.counters,
2636 "__slab_free"));
2637
2638 if (likely(!n)) {
2639
2640 /*
2641 * If we just froze the page then put it onto the
2642 * per cpu partial list.
2643 */
2644 if (new.frozen && !was_frozen) {
2645 put_cpu_partial(s, page, 1);
2646 stat(s, CPU_PARTIAL_FREE);
2647 }
2648 /*
2649 * The list lock was not taken therefore no list
2650 * activity can be necessary.
2651 */
2652 if (was_frozen)
2653 stat(s, FREE_FROZEN);
2654 return;
2655 }
2656
2657 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2658 goto slab_empty;
2659
2660 /*
2661 * Objects left in the slab. If it was not on the partial list before
2662 * then add it.
2663 */
2664 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2665 if (kmem_cache_debug(s))
2666 remove_full(s, n, page);
2667 add_partial(n, page, DEACTIVATE_TO_TAIL);
2668 stat(s, FREE_ADD_PARTIAL);
2669 }
2670 spin_unlock_irqrestore(&n->list_lock, flags);
2671 return;
2672
2673 slab_empty:
2674 if (prior) {
2675 /*
2676 * Slab on the partial list.
2677 */
2678 remove_partial(n, page);
2679 stat(s, FREE_REMOVE_PARTIAL);
2680 } else {
2681 /* Slab must be on the full list */
2682 remove_full(s, n, page);
2683 }
2684
2685 spin_unlock_irqrestore(&n->list_lock, flags);
2686 stat(s, FREE_SLAB);
2687 discard_slab(s, page);
2688 }
2689
2690 /*
2691 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2692 * can perform fastpath freeing without additional function calls.
2693 *
2694 * The fastpath is only possible if we are freeing to the current cpu slab
2695 * of this processor. This typically the case if we have just allocated
2696 * the item before.
2697 *
2698 * If fastpath is not possible then fall back to __slab_free where we deal
2699 * with all sorts of special processing.
2700 */
2701 static __always_inline void slab_free(struct kmem_cache *s,
2702 struct page *page, void *x, unsigned long addr)
2703 {
2704 void **object = (void *)x;
2705 struct kmem_cache_cpu *c;
2706 unsigned long tid;
2707
2708 slab_free_hook(s, x);
2709
2710 redo:
2711 /*
2712 * Determine the currently cpus per cpu slab.
2713 * The cpu may change afterward. However that does not matter since
2714 * data is retrieved via this pointer. If we are on the same cpu
2715 * during the cmpxchg then the free will succedd.
2716 */
2717 do {
2718 tid = this_cpu_read(s->cpu_slab->tid);
2719 c = raw_cpu_ptr(s->cpu_slab);
2720 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2721 unlikely(tid != READ_ONCE(c->tid)));
2722
2723 /* Same with comment on barrier() in slab_alloc_node() */
2724 barrier();
2725
2726 if (likely(page == c->page)) {
2727 set_freepointer(s, object, c->freelist);
2728
2729 if (unlikely(!this_cpu_cmpxchg_double(
2730 s->cpu_slab->freelist, s->cpu_slab->tid,
2731 c->freelist, tid,
2732 object, next_tid(tid)))) {
2733
2734 note_cmpxchg_failure("slab_free", s, tid);
2735 goto redo;
2736 }
2737 stat(s, FREE_FASTPATH);
2738 } else
2739 __slab_free(s, page, x, addr);
2740
2741 }
2742
2743 void kmem_cache_free(struct kmem_cache *s, void *x)
2744 {
2745 s = cache_from_obj(s, x);
2746 if (!s)
2747 return;
2748 slab_free(s, virt_to_head_page(x), x, _RET_IP_);
2749 trace_kmem_cache_free(_RET_IP_, x);
2750 }
2751 EXPORT_SYMBOL(kmem_cache_free);
2752
2753 /*
2754 * Object placement in a slab is made very easy because we always start at
2755 * offset 0. If we tune the size of the object to the alignment then we can
2756 * get the required alignment by putting one properly sized object after
2757 * another.
2758 *
2759 * Notice that the allocation order determines the sizes of the per cpu
2760 * caches. Each processor has always one slab available for allocations.
2761 * Increasing the allocation order reduces the number of times that slabs
2762 * must be moved on and off the partial lists and is therefore a factor in
2763 * locking overhead.
2764 */
2765
2766 /*
2767 * Mininum / Maximum order of slab pages. This influences locking overhead
2768 * and slab fragmentation. A higher order reduces the number of partial slabs
2769 * and increases the number of allocations possible without having to
2770 * take the list_lock.
2771 */
2772 static int slub_min_order;
2773 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2774 static int slub_min_objects;
2775
2776 /*
2777 * Calculate the order of allocation given an slab object size.
2778 *
2779 * The order of allocation has significant impact on performance and other
2780 * system components. Generally order 0 allocations should be preferred since
2781 * order 0 does not cause fragmentation in the page allocator. Larger objects
2782 * be problematic to put into order 0 slabs because there may be too much
2783 * unused space left. We go to a higher order if more than 1/16th of the slab
2784 * would be wasted.
2785 *
2786 * In order to reach satisfactory performance we must ensure that a minimum
2787 * number of objects is in one slab. Otherwise we may generate too much
2788 * activity on the partial lists which requires taking the list_lock. This is
2789 * less a concern for large slabs though which are rarely used.
2790 *
2791 * slub_max_order specifies the order where we begin to stop considering the
2792 * number of objects in a slab as critical. If we reach slub_max_order then
2793 * we try to keep the page order as low as possible. So we accept more waste
2794 * of space in favor of a small page order.
2795 *
2796 * Higher order allocations also allow the placement of more objects in a
2797 * slab and thereby reduce object handling overhead. If the user has
2798 * requested a higher mininum order then we start with that one instead of
2799 * the smallest order which will fit the object.
2800 */
2801 static inline int slab_order(int size, int min_objects,
2802 int max_order, int fract_leftover, int reserved)
2803 {
2804 int order;
2805 int rem;
2806 int min_order = slub_min_order;
2807
2808 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2809 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2810
2811 for (order = max(min_order,
2812 fls(min_objects * size - 1) - PAGE_SHIFT);
2813 order <= max_order; order++) {
2814
2815 unsigned long slab_size = PAGE_SIZE << order;
2816
2817 if (slab_size < min_objects * size + reserved)
2818 continue;
2819
2820 rem = (slab_size - reserved) % size;
2821
2822 if (rem <= slab_size / fract_leftover)
2823 break;
2824
2825 }
2826
2827 return order;
2828 }
2829
2830 static inline int calculate_order(int size, int reserved)
2831 {
2832 int order;
2833 int min_objects;
2834 int fraction;
2835 int max_objects;
2836
2837 /*
2838 * Attempt to find best configuration for a slab. This
2839 * works by first attempting to generate a layout with
2840 * the best configuration and backing off gradually.
2841 *
2842 * First we reduce the acceptable waste in a slab. Then
2843 * we reduce the minimum objects required in a slab.
2844 */
2845 min_objects = slub_min_objects;
2846 if (!min_objects)
2847 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2848 max_objects = order_objects(slub_max_order, size, reserved);
2849 min_objects = min(min_objects, max_objects);
2850
2851 while (min_objects > 1) {
2852 fraction = 16;
2853 while (fraction >= 4) {
2854 order = slab_order(size, min_objects,
2855 slub_max_order, fraction, reserved);
2856 if (order <= slub_max_order)
2857 return order;
2858 fraction /= 2;
2859 }
2860 min_objects--;
2861 }
2862
2863 /*
2864 * We were unable to place multiple objects in a slab. Now
2865 * lets see if we can place a single object there.
2866 */
2867 order = slab_order(size, 1, slub_max_order, 1, reserved);
2868 if (order <= slub_max_order)
2869 return order;
2870
2871 /*
2872 * Doh this slab cannot be placed using slub_max_order.
2873 */
2874 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2875 if (order < MAX_ORDER)
2876 return order;
2877 return -ENOSYS;
2878 }
2879
2880 static void
2881 init_kmem_cache_node(struct kmem_cache_node *n)
2882 {
2883 n->nr_partial = 0;
2884 spin_lock_init(&n->list_lock);
2885 INIT_LIST_HEAD(&n->partial);
2886 #ifdef CONFIG_SLUB_DEBUG
2887 atomic_long_set(&n->nr_slabs, 0);
2888 atomic_long_set(&n->total_objects, 0);
2889 INIT_LIST_HEAD(&n->full);
2890 #endif
2891 }
2892
2893 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2894 {
2895 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2896 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
2897
2898 /*
2899 * Must align to double word boundary for the double cmpxchg
2900 * instructions to work; see __pcpu_double_call_return_bool().
2901 */
2902 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2903 2 * sizeof(void *));
2904
2905 if (!s->cpu_slab)
2906 return 0;
2907
2908 init_kmem_cache_cpus(s);
2909
2910 return 1;
2911 }
2912
2913 static struct kmem_cache *kmem_cache_node;
2914
2915 /*
2916 * No kmalloc_node yet so do it by hand. We know that this is the first
2917 * slab on the node for this slabcache. There are no concurrent accesses
2918 * possible.
2919 *
2920 * Note that this function only works on the kmem_cache_node
2921 * when allocating for the kmem_cache_node. This is used for bootstrapping
2922 * memory on a fresh node that has no slab structures yet.
2923 */
2924 static void early_kmem_cache_node_alloc(int node)
2925 {
2926 struct page *page;
2927 struct kmem_cache_node *n;
2928
2929 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2930
2931 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2932
2933 BUG_ON(!page);
2934 if (page_to_nid(page) != node) {
2935 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
2936 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
2937 }
2938
2939 n = page->freelist;
2940 BUG_ON(!n);
2941 page->freelist = get_freepointer(kmem_cache_node, n);
2942 page->inuse = 1;
2943 page->frozen = 0;
2944 kmem_cache_node->node[node] = n;
2945 #ifdef CONFIG_SLUB_DEBUG
2946 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2947 init_tracking(kmem_cache_node, n);
2948 #endif
2949 kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node));
2950 init_kmem_cache_node(n);
2951 inc_slabs_node(kmem_cache_node, node, page->objects);
2952
2953 /*
2954 * No locks need to be taken here as it has just been
2955 * initialized and there is no concurrent access.
2956 */
2957 __add_partial(n, page, DEACTIVATE_TO_HEAD);
2958 }
2959
2960 static void free_kmem_cache_nodes(struct kmem_cache *s)
2961 {
2962 int node;
2963 struct kmem_cache_node *n;
2964
2965 for_each_kmem_cache_node(s, node, n) {
2966 kmem_cache_free(kmem_cache_node, n);
2967 s->node[node] = NULL;
2968 }
2969 }
2970
2971 static int init_kmem_cache_nodes(struct kmem_cache *s)
2972 {
2973 int node;
2974
2975 for_each_node_state(node, N_NORMAL_MEMORY) {
2976 struct kmem_cache_node *n;
2977
2978 if (slab_state == DOWN) {
2979 early_kmem_cache_node_alloc(node);
2980 continue;
2981 }
2982 n = kmem_cache_alloc_node(kmem_cache_node,
2983 GFP_KERNEL, node);
2984
2985 if (!n) {
2986 free_kmem_cache_nodes(s);
2987 return 0;
2988 }
2989
2990 s->node[node] = n;
2991 init_kmem_cache_node(n);
2992 }
2993 return 1;
2994 }
2995
2996 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2997 {
2998 if (min < MIN_PARTIAL)
2999 min = MIN_PARTIAL;
3000 else if (min > MAX_PARTIAL)
3001 min = MAX_PARTIAL;
3002 s->min_partial = min;
3003 }
3004
3005 /*
3006 * calculate_sizes() determines the order and the distribution of data within
3007 * a slab object.
3008 */
3009 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3010 {
3011 unsigned long flags = s->flags;
3012 unsigned long size = s->object_size;
3013 int order;
3014
3015 /*
3016 * Round up object size to the next word boundary. We can only
3017 * place the free pointer at word boundaries and this determines
3018 * the possible location of the free pointer.
3019 */
3020 size = ALIGN(size, sizeof(void *));
3021
3022 #ifdef CONFIG_SLUB_DEBUG
3023 /*
3024 * Determine if we can poison the object itself. If the user of
3025 * the slab may touch the object after free or before allocation
3026 * then we should never poison the object itself.
3027 */
3028 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
3029 !s->ctor)
3030 s->flags |= __OBJECT_POISON;
3031 else
3032 s->flags &= ~__OBJECT_POISON;
3033
3034
3035 /*
3036 * If we are Redzoning then check if there is some space between the
3037 * end of the object and the free pointer. If not then add an
3038 * additional word to have some bytes to store Redzone information.
3039 */
3040 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3041 size += sizeof(void *);
3042 #endif
3043
3044 /*
3045 * With that we have determined the number of bytes in actual use
3046 * by the object. This is the potential offset to the free pointer.
3047 */
3048 s->inuse = size;
3049
3050 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
3051 s->ctor)) {
3052 /*
3053 * Relocate free pointer after the object if it is not
3054 * permitted to overwrite the first word of the object on
3055 * kmem_cache_free.
3056 *
3057 * This is the case if we do RCU, have a constructor or
3058 * destructor or are poisoning the objects.
3059 */
3060 s->offset = size;
3061 size += sizeof(void *);
3062 }
3063
3064 #ifdef CONFIG_SLUB_DEBUG
3065 if (flags & SLAB_STORE_USER)
3066 /*
3067 * Need to store information about allocs and frees after
3068 * the object.
3069 */
3070 size += 2 * sizeof(struct track);
3071
3072 if (flags & SLAB_RED_ZONE)
3073 /*
3074 * Add some empty padding so that we can catch
3075 * overwrites from earlier objects rather than let
3076 * tracking information or the free pointer be
3077 * corrupted if a user writes before the start
3078 * of the object.
3079 */
3080 size += sizeof(void *);
3081 #endif
3082
3083 /*
3084 * SLUB stores one object immediately after another beginning from
3085 * offset 0. In order to align the objects we have to simply size
3086 * each object to conform to the alignment.
3087 */
3088 size = ALIGN(size, s->align);
3089 s->size = size;
3090 if (forced_order >= 0)
3091 order = forced_order;
3092 else
3093 order = calculate_order(size, s->reserved);
3094
3095 if (order < 0)
3096 return 0;
3097
3098 s->allocflags = 0;
3099 if (order)
3100 s->allocflags |= __GFP_COMP;
3101
3102 if (s->flags & SLAB_CACHE_DMA)
3103 s->allocflags |= GFP_DMA;
3104
3105 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3106 s->allocflags |= __GFP_RECLAIMABLE;
3107
3108 /*
3109 * Determine the number of objects per slab
3110 */
3111 s->oo = oo_make(order, size, s->reserved);
3112 s->min = oo_make(get_order(size), size, s->reserved);
3113 if (oo_objects(s->oo) > oo_objects(s->max))
3114 s->max = s->oo;
3115
3116 return !!oo_objects(s->oo);
3117 }
3118
3119 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3120 {
3121 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3122 s->reserved = 0;
3123
3124 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3125 s->reserved = sizeof(struct rcu_head);
3126
3127 if (!calculate_sizes(s, -1))
3128 goto error;
3129 if (disable_higher_order_debug) {
3130 /*
3131 * Disable debugging flags that store metadata if the min slab
3132 * order increased.
3133 */
3134 if (get_order(s->size) > get_order(s->object_size)) {
3135 s->flags &= ~DEBUG_METADATA_FLAGS;
3136 s->offset = 0;
3137 if (!calculate_sizes(s, -1))
3138 goto error;
3139 }
3140 }
3141
3142 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3143 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3144 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3145 /* Enable fast mode */
3146 s->flags |= __CMPXCHG_DOUBLE;
3147 #endif
3148
3149 /*
3150 * The larger the object size is, the more pages we want on the partial
3151 * list to avoid pounding the page allocator excessively.
3152 */
3153 set_min_partial(s, ilog2(s->size) / 2);
3154
3155 /*
3156 * cpu_partial determined the maximum number of objects kept in the
3157 * per cpu partial lists of a processor.
3158 *
3159 * Per cpu partial lists mainly contain slabs that just have one
3160 * object freed. If they are used for allocation then they can be
3161 * filled up again with minimal effort. The slab will never hit the
3162 * per node partial lists and therefore no locking will be required.
3163 *
3164 * This setting also determines
3165 *
3166 * A) The number of objects from per cpu partial slabs dumped to the
3167 * per node list when we reach the limit.
3168 * B) The number of objects in cpu partial slabs to extract from the
3169 * per node list when we run out of per cpu objects. We only fetch
3170 * 50% to keep some capacity around for frees.
3171 */
3172 if (!kmem_cache_has_cpu_partial(s))
3173 s->cpu_partial = 0;
3174 else if (s->size >= PAGE_SIZE)
3175 s->cpu_partial = 2;
3176 else if (s->size >= 1024)
3177 s->cpu_partial = 6;
3178 else if (s->size >= 256)
3179 s->cpu_partial = 13;
3180 else
3181 s->cpu_partial = 30;
3182
3183 #ifdef CONFIG_NUMA
3184 s->remote_node_defrag_ratio = 1000;
3185 #endif
3186 if (!init_kmem_cache_nodes(s))
3187 goto error;
3188
3189 if (alloc_kmem_cache_cpus(s))
3190 return 0;
3191
3192 free_kmem_cache_nodes(s);
3193 error:
3194 if (flags & SLAB_PANIC)
3195 panic("Cannot create slab %s size=%lu realsize=%u "
3196 "order=%u offset=%u flags=%lx\n",
3197 s->name, (unsigned long)s->size, s->size,
3198 oo_order(s->oo), s->offset, flags);
3199 return -EINVAL;
3200 }
3201
3202 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3203 const char *text)
3204 {
3205 #ifdef CONFIG_SLUB_DEBUG
3206 void *addr = page_address(page);
3207 void *p;
3208 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3209 sizeof(long), GFP_ATOMIC);
3210 if (!map)
3211 return;
3212 slab_err(s, page, text, s->name);
3213 slab_lock(page);
3214
3215 get_map(s, page, map);
3216 for_each_object(p, s, addr, page->objects) {
3217
3218 if (!test_bit(slab_index(p, s, addr), map)) {
3219 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3220 print_tracking(s, p);
3221 }
3222 }
3223 slab_unlock(page);
3224 kfree(map);
3225 #endif
3226 }
3227
3228 /*
3229 * Attempt to free all partial slabs on a node.
3230 * This is called from kmem_cache_close(). We must be the last thread
3231 * using the cache and therefore we do not need to lock anymore.
3232 */
3233 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3234 {
3235 struct page *page, *h;
3236
3237 list_for_each_entry_safe(page, h, &n->partial, lru) {
3238 if (!page->inuse) {
3239 __remove_partial(n, page);
3240 discard_slab(s, page);
3241 } else {
3242 list_slab_objects(s, page,
3243 "Objects remaining in %s on kmem_cache_close()");
3244 }
3245 }
3246 }
3247
3248 /*
3249 * Release all resources used by a slab cache.
3250 */
3251 static inline int kmem_cache_close(struct kmem_cache *s)
3252 {
3253 int node;
3254 struct kmem_cache_node *n;
3255
3256 flush_all(s);
3257 /* Attempt to free all objects */
3258 for_each_kmem_cache_node(s, node, n) {
3259 free_partial(s, n);
3260 if (n->nr_partial || slabs_node(s, node))
3261 return 1;
3262 }
3263 free_percpu(s->cpu_slab);
3264 free_kmem_cache_nodes(s);
3265 return 0;
3266 }
3267
3268 int __kmem_cache_shutdown(struct kmem_cache *s)
3269 {
3270 return kmem_cache_close(s);
3271 }
3272
3273 /********************************************************************
3274 * Kmalloc subsystem
3275 *******************************************************************/
3276
3277 static int __init setup_slub_min_order(char *str)
3278 {
3279 get_option(&str, &slub_min_order);
3280
3281 return 1;
3282 }
3283
3284 __setup("slub_min_order=", setup_slub_min_order);
3285
3286 static int __init setup_slub_max_order(char *str)
3287 {
3288 get_option(&str, &slub_max_order);
3289 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3290
3291 return 1;
3292 }
3293
3294 __setup("slub_max_order=", setup_slub_max_order);
3295
3296 static int __init setup_slub_min_objects(char *str)
3297 {
3298 get_option(&str, &slub_min_objects);
3299
3300 return 1;
3301 }
3302
3303 __setup("slub_min_objects=", setup_slub_min_objects);
3304
3305 void *__kmalloc(size_t size, gfp_t flags)
3306 {
3307 struct kmem_cache *s;
3308 void *ret;
3309
3310 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3311 return kmalloc_large(size, flags);
3312
3313 s = kmalloc_slab(size, flags);
3314
3315 if (unlikely(ZERO_OR_NULL_PTR(s)))
3316 return s;
3317
3318 ret = slab_alloc(s, flags, _RET_IP_);
3319
3320 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3321
3322 kasan_kmalloc(s, ret, size);
3323
3324 return ret;
3325 }
3326 EXPORT_SYMBOL(__kmalloc);
3327
3328 #ifdef CONFIG_NUMA
3329 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3330 {
3331 struct page *page;
3332 void *ptr = NULL;
3333
3334 flags |= __GFP_COMP | __GFP_NOTRACK;
3335 page = alloc_kmem_pages_node(node, flags, get_order(size));
3336 if (page)
3337 ptr = page_address(page);
3338
3339 kmalloc_large_node_hook(ptr, size, flags);
3340 return ptr;
3341 }
3342
3343 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3344 {
3345 struct kmem_cache *s;
3346 void *ret;
3347
3348 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3349 ret = kmalloc_large_node(size, flags, node);
3350
3351 trace_kmalloc_node(_RET_IP_, ret,
3352 size, PAGE_SIZE << get_order(size),
3353 flags, node);
3354
3355 return ret;
3356 }
3357
3358 s = kmalloc_slab(size, flags);
3359
3360 if (unlikely(ZERO_OR_NULL_PTR(s)))
3361 return s;
3362
3363 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3364
3365 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3366
3367 kasan_kmalloc(s, ret, size);
3368
3369 return ret;
3370 }
3371 EXPORT_SYMBOL(__kmalloc_node);
3372 #endif
3373
3374 static size_t __ksize(const void *object)
3375 {
3376 struct page *page;
3377
3378 if (unlikely(object == ZERO_SIZE_PTR))
3379 return 0;
3380
3381 page = virt_to_head_page(object);
3382
3383 if (unlikely(!PageSlab(page))) {
3384 WARN_ON(!PageCompound(page));
3385 return PAGE_SIZE << compound_order(page);
3386 }
3387
3388 return slab_ksize(page->slab_cache);
3389 }
3390
3391 size_t ksize(const void *object)
3392 {
3393 size_t size = __ksize(object);
3394 /* We assume that ksize callers could use whole allocated area,
3395 so we need unpoison this area. */
3396 kasan_krealloc(object, size);
3397 return size;
3398 }
3399 EXPORT_SYMBOL(ksize);
3400
3401 void kfree(const void *x)
3402 {
3403 struct page *page;
3404 void *object = (void *)x;
3405
3406 trace_kfree(_RET_IP_, x);
3407
3408 if (unlikely(ZERO_OR_NULL_PTR(x)))
3409 return;
3410
3411 page = virt_to_head_page(x);
3412 if (unlikely(!PageSlab(page))) {
3413 BUG_ON(!PageCompound(page));
3414 kfree_hook(x);
3415 __free_kmem_pages(page, compound_order(page));
3416 return;
3417 }
3418 slab_free(page->slab_cache, page, object, _RET_IP_);
3419 }
3420 EXPORT_SYMBOL(kfree);
3421
3422 #define SHRINK_PROMOTE_MAX 32
3423
3424 /*
3425 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3426 * up most to the head of the partial lists. New allocations will then
3427 * fill those up and thus they can be removed from the partial lists.
3428 *
3429 * The slabs with the least items are placed last. This results in them
3430 * being allocated from last increasing the chance that the last objects
3431 * are freed in them.
3432 */
3433 int __kmem_cache_shrink(struct kmem_cache *s, bool deactivate)
3434 {
3435 int node;
3436 int i;
3437 struct kmem_cache_node *n;
3438 struct page *page;
3439 struct page *t;
3440 struct list_head discard;
3441 struct list_head promote[SHRINK_PROMOTE_MAX];
3442 unsigned long flags;
3443 int ret = 0;
3444
3445 if (deactivate) {
3446 /*
3447 * Disable empty slabs caching. Used to avoid pinning offline
3448 * memory cgroups by kmem pages that can be freed.
3449 */
3450 s->cpu_partial = 0;
3451 s->min_partial = 0;
3452
3453 /*
3454 * s->cpu_partial is checked locklessly (see put_cpu_partial),
3455 * so we have to make sure the change is visible.
3456 */
3457 kick_all_cpus_sync();
3458 }
3459
3460 flush_all(s);
3461 for_each_kmem_cache_node(s, node, n) {
3462 INIT_LIST_HEAD(&discard);
3463 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3464 INIT_LIST_HEAD(promote + i);
3465
3466 spin_lock_irqsave(&n->list_lock, flags);
3467
3468 /*
3469 * Build lists of slabs to discard or promote.
3470 *
3471 * Note that concurrent frees may occur while we hold the
3472 * list_lock. page->inuse here is the upper limit.
3473 */
3474 list_for_each_entry_safe(page, t, &n->partial, lru) {
3475 int free = page->objects - page->inuse;
3476
3477 /* Do not reread page->inuse */
3478 barrier();
3479
3480 /* We do not keep full slabs on the list */
3481 BUG_ON(free <= 0);
3482
3483 if (free == page->objects) {
3484 list_move(&page->lru, &discard);
3485 n->nr_partial--;
3486 } else if (free <= SHRINK_PROMOTE_MAX)
3487 list_move(&page->lru, promote + free - 1);
3488 }
3489
3490 /*
3491 * Promote the slabs filled up most to the head of the
3492 * partial list.
3493 */
3494 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
3495 list_splice(promote + i, &n->partial);
3496
3497 spin_unlock_irqrestore(&n->list_lock, flags);
3498
3499 /* Release empty slabs */
3500 list_for_each_entry_safe(page, t, &discard, lru)
3501 discard_slab(s, page);
3502
3503 if (slabs_node(s, node))
3504 ret = 1;
3505 }
3506
3507 return ret;
3508 }
3509
3510 static int slab_mem_going_offline_callback(void *arg)
3511 {
3512 struct kmem_cache *s;
3513
3514 mutex_lock(&slab_mutex);
3515 list_for_each_entry(s, &slab_caches, list)
3516 __kmem_cache_shrink(s, false);
3517 mutex_unlock(&slab_mutex);
3518
3519 return 0;
3520 }
3521
3522 static void slab_mem_offline_callback(void *arg)
3523 {
3524 struct kmem_cache_node *n;
3525 struct kmem_cache *s;
3526 struct memory_notify *marg = arg;
3527 int offline_node;
3528
3529 offline_node = marg->status_change_nid_normal;
3530
3531 /*
3532 * If the node still has available memory. we need kmem_cache_node
3533 * for it yet.
3534 */
3535 if (offline_node < 0)
3536 return;
3537
3538 mutex_lock(&slab_mutex);
3539 list_for_each_entry(s, &slab_caches, list) {
3540 n = get_node(s, offline_node);
3541 if (n) {
3542 /*
3543 * if n->nr_slabs > 0, slabs still exist on the node
3544 * that is going down. We were unable to free them,
3545 * and offline_pages() function shouldn't call this
3546 * callback. So, we must fail.
3547 */
3548 BUG_ON(slabs_node(s, offline_node));
3549
3550 s->node[offline_node] = NULL;
3551 kmem_cache_free(kmem_cache_node, n);
3552 }
3553 }
3554 mutex_unlock(&slab_mutex);
3555 }
3556
3557 static int slab_mem_going_online_callback(void *arg)
3558 {
3559 struct kmem_cache_node *n;
3560 struct kmem_cache *s;
3561 struct memory_notify *marg = arg;
3562 int nid = marg->status_change_nid_normal;
3563 int ret = 0;
3564
3565 /*
3566 * If the node's memory is already available, then kmem_cache_node is
3567 * already created. Nothing to do.
3568 */
3569 if (nid < 0)
3570 return 0;
3571
3572 /*
3573 * We are bringing a node online. No memory is available yet. We must
3574 * allocate a kmem_cache_node structure in order to bring the node
3575 * online.
3576 */
3577 mutex_lock(&slab_mutex);
3578 list_for_each_entry(s, &slab_caches, list) {
3579 /*
3580 * XXX: kmem_cache_alloc_node will fallback to other nodes
3581 * since memory is not yet available from the node that
3582 * is brought up.
3583 */
3584 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3585 if (!n) {
3586 ret = -ENOMEM;
3587 goto out;
3588 }
3589 init_kmem_cache_node(n);
3590 s->node[nid] = n;
3591 }
3592 out:
3593 mutex_unlock(&slab_mutex);
3594 return ret;
3595 }
3596
3597 static int slab_memory_callback(struct notifier_block *self,
3598 unsigned long action, void *arg)
3599 {
3600 int ret = 0;
3601
3602 switch (action) {
3603 case MEM_GOING_ONLINE:
3604 ret = slab_mem_going_online_callback(arg);
3605 break;
3606 case MEM_GOING_OFFLINE:
3607 ret = slab_mem_going_offline_callback(arg);
3608 break;
3609 case MEM_OFFLINE:
3610 case MEM_CANCEL_ONLINE:
3611 slab_mem_offline_callback(arg);
3612 break;
3613 case MEM_ONLINE:
3614 case MEM_CANCEL_OFFLINE:
3615 break;
3616 }
3617 if (ret)
3618 ret = notifier_from_errno(ret);
3619 else
3620 ret = NOTIFY_OK;
3621 return ret;
3622 }
3623
3624 static struct notifier_block slab_memory_callback_nb = {
3625 .notifier_call = slab_memory_callback,
3626 .priority = SLAB_CALLBACK_PRI,
3627 };
3628
3629 /********************************************************************
3630 * Basic setup of slabs
3631 *******************************************************************/
3632
3633 /*
3634 * Used for early kmem_cache structures that were allocated using
3635 * the page allocator. Allocate them properly then fix up the pointers
3636 * that may be pointing to the wrong kmem_cache structure.
3637 */
3638
3639 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3640 {
3641 int node;
3642 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3643 struct kmem_cache_node *n;
3644
3645 memcpy(s, static_cache, kmem_cache->object_size);
3646
3647 /*
3648 * This runs very early, and only the boot processor is supposed to be
3649 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3650 * IPIs around.
3651 */
3652 __flush_cpu_slab(s, smp_processor_id());
3653 for_each_kmem_cache_node(s, node, n) {
3654 struct page *p;
3655
3656 list_for_each_entry(p, &n->partial, lru)
3657 p->slab_cache = s;
3658
3659 #ifdef CONFIG_SLUB_DEBUG
3660 list_for_each_entry(p, &n->full, lru)
3661 p->slab_cache = s;
3662 #endif
3663 }
3664 slab_init_memcg_params(s);
3665 list_add(&s->list, &slab_caches);
3666 return s;
3667 }
3668
3669 void __init kmem_cache_init(void)
3670 {
3671 static __initdata struct kmem_cache boot_kmem_cache,
3672 boot_kmem_cache_node;
3673
3674 if (debug_guardpage_minorder())
3675 slub_max_order = 0;
3676
3677 kmem_cache_node = &boot_kmem_cache_node;
3678 kmem_cache = &boot_kmem_cache;
3679
3680 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3681 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3682
3683 register_hotmemory_notifier(&slab_memory_callback_nb);
3684
3685 /* Able to allocate the per node structures */
3686 slab_state = PARTIAL;
3687
3688 create_boot_cache(kmem_cache, "kmem_cache",
3689 offsetof(struct kmem_cache, node) +
3690 nr_node_ids * sizeof(struct kmem_cache_node *),
3691 SLAB_HWCACHE_ALIGN);
3692
3693 kmem_cache = bootstrap(&boot_kmem_cache);
3694
3695 /*
3696 * Allocate kmem_cache_node properly from the kmem_cache slab.
3697 * kmem_cache_node is separately allocated so no need to
3698 * update any list pointers.
3699 */
3700 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3701
3702 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3703 setup_kmalloc_cache_index_table();
3704 create_kmalloc_caches(0);
3705
3706 #ifdef CONFIG_SMP
3707 register_cpu_notifier(&slab_notifier);
3708 #endif
3709
3710 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
3711 cache_line_size(),
3712 slub_min_order, slub_max_order, slub_min_objects,
3713 nr_cpu_ids, nr_node_ids);
3714 }
3715
3716 void __init kmem_cache_init_late(void)
3717 {
3718 }
3719
3720 struct kmem_cache *
3721 __kmem_cache_alias(const char *name, size_t size, size_t align,
3722 unsigned long flags, void (*ctor)(void *))
3723 {
3724 struct kmem_cache *s, *c;
3725
3726 s = find_mergeable(size, align, flags, name, ctor);
3727 if (s) {
3728 s->refcount++;
3729
3730 /*
3731 * Adjust the object sizes so that we clear
3732 * the complete object on kzalloc.
3733 */
3734 s->object_size = max(s->object_size, (int)size);
3735 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3736
3737 for_each_memcg_cache(c, s) {
3738 c->object_size = s->object_size;
3739 c->inuse = max_t(int, c->inuse,
3740 ALIGN(size, sizeof(void *)));
3741 }
3742
3743 if (sysfs_slab_alias(s, name)) {
3744 s->refcount--;
3745 s = NULL;
3746 }
3747 }
3748
3749 return s;
3750 }
3751
3752 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3753 {
3754 int err;
3755
3756 err = kmem_cache_open(s, flags);
3757 if (err)
3758 return err;
3759
3760 /* Mutex is not taken during early boot */
3761 if (slab_state <= UP)
3762 return 0;
3763
3764 memcg_propagate_slab_attrs(s);
3765 err = sysfs_slab_add(s);
3766 if (err)
3767 kmem_cache_close(s);
3768
3769 return err;
3770 }
3771
3772 #ifdef CONFIG_SMP
3773 /*
3774 * Use the cpu notifier to insure that the cpu slabs are flushed when
3775 * necessary.
3776 */
3777 static int slab_cpuup_callback(struct notifier_block *nfb,
3778 unsigned long action, void *hcpu)
3779 {
3780 long cpu = (long)hcpu;
3781 struct kmem_cache *s;
3782 unsigned long flags;
3783
3784 switch (action) {
3785 case CPU_UP_CANCELED:
3786 case CPU_UP_CANCELED_FROZEN:
3787 case CPU_DEAD:
3788 case CPU_DEAD_FROZEN:
3789 mutex_lock(&slab_mutex);
3790 list_for_each_entry(s, &slab_caches, list) {
3791 local_irq_save(flags);
3792 __flush_cpu_slab(s, cpu);
3793 local_irq_restore(flags);
3794 }
3795 mutex_unlock(&slab_mutex);
3796 break;
3797 default:
3798 break;
3799 }
3800 return NOTIFY_OK;
3801 }
3802
3803 static struct notifier_block slab_notifier = {
3804 .notifier_call = slab_cpuup_callback
3805 };
3806
3807 #endif
3808
3809 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3810 {
3811 struct kmem_cache *s;
3812 void *ret;
3813
3814 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3815 return kmalloc_large(size, gfpflags);
3816
3817 s = kmalloc_slab(size, gfpflags);
3818
3819 if (unlikely(ZERO_OR_NULL_PTR(s)))
3820 return s;
3821
3822 ret = slab_alloc(s, gfpflags, caller);
3823
3824 /* Honor the call site pointer we received. */
3825 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3826
3827 return ret;
3828 }
3829
3830 #ifdef CONFIG_NUMA
3831 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3832 int node, unsigned long caller)
3833 {
3834 struct kmem_cache *s;
3835 void *ret;
3836
3837 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3838 ret = kmalloc_large_node(size, gfpflags, node);
3839
3840 trace_kmalloc_node(caller, ret,
3841 size, PAGE_SIZE << get_order(size),
3842 gfpflags, node);
3843
3844 return ret;
3845 }
3846
3847 s = kmalloc_slab(size, gfpflags);
3848
3849 if (unlikely(ZERO_OR_NULL_PTR(s)))
3850 return s;
3851
3852 ret = slab_alloc_node(s, gfpflags, node, caller);
3853
3854 /* Honor the call site pointer we received. */
3855 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3856
3857 return ret;
3858 }
3859 #endif
3860
3861 #ifdef CONFIG_SYSFS
3862 static int count_inuse(struct page *page)
3863 {
3864 return page->inuse;
3865 }
3866
3867 static int count_total(struct page *page)
3868 {
3869 return page->objects;
3870 }
3871 #endif
3872
3873 #ifdef CONFIG_SLUB_DEBUG
3874 static int validate_slab(struct kmem_cache *s, struct page *page,
3875 unsigned long *map)
3876 {
3877 void *p;
3878 void *addr = page_address(page);
3879
3880 if (!check_slab(s, page) ||
3881 !on_freelist(s, page, NULL))
3882 return 0;
3883
3884 /* Now we know that a valid freelist exists */
3885 bitmap_zero(map, page->objects);
3886
3887 get_map(s, page, map);
3888 for_each_object(p, s, addr, page->objects) {
3889 if (test_bit(slab_index(p, s, addr), map))
3890 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3891 return 0;
3892 }
3893
3894 for_each_object(p, s, addr, page->objects)
3895 if (!test_bit(slab_index(p, s, addr), map))
3896 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3897 return 0;
3898 return 1;
3899 }
3900
3901 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3902 unsigned long *map)
3903 {
3904 slab_lock(page);
3905 validate_slab(s, page, map);
3906 slab_unlock(page);
3907 }
3908
3909 static int validate_slab_node(struct kmem_cache *s,
3910 struct kmem_cache_node *n, unsigned long *map)
3911 {
3912 unsigned long count = 0;
3913 struct page *page;
3914 unsigned long flags;
3915
3916 spin_lock_irqsave(&n->list_lock, flags);
3917
3918 list_for_each_entry(page, &n->partial, lru) {
3919 validate_slab_slab(s, page, map);
3920 count++;
3921 }
3922 if (count != n->nr_partial)
3923 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
3924 s->name, count, n->nr_partial);
3925
3926 if (!(s->flags & SLAB_STORE_USER))
3927 goto out;
3928
3929 list_for_each_entry(page, &n->full, lru) {
3930 validate_slab_slab(s, page, map);
3931 count++;
3932 }
3933 if (count != atomic_long_read(&n->nr_slabs))
3934 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
3935 s->name, count, atomic_long_read(&n->nr_slabs));
3936
3937 out:
3938 spin_unlock_irqrestore(&n->list_lock, flags);
3939 return count;
3940 }
3941
3942 static long validate_slab_cache(struct kmem_cache *s)
3943 {
3944 int node;
3945 unsigned long count = 0;
3946 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3947 sizeof(unsigned long), GFP_KERNEL);
3948 struct kmem_cache_node *n;
3949
3950 if (!map)
3951 return -ENOMEM;
3952
3953 flush_all(s);
3954 for_each_kmem_cache_node(s, node, n)
3955 count += validate_slab_node(s, n, map);
3956 kfree(map);
3957 return count;
3958 }
3959 /*
3960 * Generate lists of code addresses where slabcache objects are allocated
3961 * and freed.
3962 */
3963
3964 struct location {
3965 unsigned long count;
3966 unsigned long addr;
3967 long long sum_time;
3968 long min_time;
3969 long max_time;
3970 long min_pid;
3971 long max_pid;
3972 DECLARE_BITMAP(cpus, NR_CPUS);
3973 nodemask_t nodes;
3974 };
3975
3976 struct loc_track {
3977 unsigned long max;
3978 unsigned long count;
3979 struct location *loc;
3980 };
3981
3982 static void free_loc_track(struct loc_track *t)
3983 {
3984 if (t->max)
3985 free_pages((unsigned long)t->loc,
3986 get_order(sizeof(struct location) * t->max));
3987 }
3988
3989 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3990 {
3991 struct location *l;
3992 int order;
3993
3994 order = get_order(sizeof(struct location) * max);
3995
3996 l = (void *)__get_free_pages(flags, order);
3997 if (!l)
3998 return 0;
3999
4000 if (t->count) {
4001 memcpy(l, t->loc, sizeof(struct location) * t->count);
4002 free_loc_track(t);
4003 }
4004 t->max = max;
4005 t->loc = l;
4006 return 1;
4007 }
4008
4009 static int add_location(struct loc_track *t, struct kmem_cache *s,
4010 const struct track *track)
4011 {
4012 long start, end, pos;
4013 struct location *l;
4014 unsigned long caddr;
4015 unsigned long age = jiffies - track->when;
4016
4017 start = -1;
4018 end = t->count;
4019
4020 for ( ; ; ) {
4021 pos = start + (end - start + 1) / 2;
4022
4023 /*
4024 * There is nothing at "end". If we end up there
4025 * we need to add something to before end.
4026 */
4027 if (pos == end)
4028 break;
4029
4030 caddr = t->loc[pos].addr;
4031 if (track->addr == caddr) {
4032
4033 l = &t->loc[pos];
4034 l->count++;
4035 if (track->when) {
4036 l->sum_time += age;
4037 if (age < l->min_time)
4038 l->min_time = age;
4039 if (age > l->max_time)
4040 l->max_time = age;
4041
4042 if (track->pid < l->min_pid)
4043 l->min_pid = track->pid;
4044 if (track->pid > l->max_pid)
4045 l->max_pid = track->pid;
4046
4047 cpumask_set_cpu(track->cpu,
4048 to_cpumask(l->cpus));
4049 }
4050 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4051 return 1;
4052 }
4053
4054 if (track->addr < caddr)
4055 end = pos;
4056 else
4057 start = pos;
4058 }
4059
4060 /*
4061 * Not found. Insert new tracking element.
4062 */
4063 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4064 return 0;
4065
4066 l = t->loc + pos;
4067 if (pos < t->count)
4068 memmove(l + 1, l,
4069 (t->count - pos) * sizeof(struct location));
4070 t->count++;
4071 l->count = 1;
4072 l->addr = track->addr;
4073 l->sum_time = age;
4074 l->min_time = age;
4075 l->max_time = age;
4076 l->min_pid = track->pid;
4077 l->max_pid = track->pid;
4078 cpumask_clear(to_cpumask(l->cpus));
4079 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4080 nodes_clear(l->nodes);
4081 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4082 return 1;
4083 }
4084
4085 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4086 struct page *page, enum track_item alloc,
4087 unsigned long *map)
4088 {
4089 void *addr = page_address(page);
4090 void *p;
4091
4092 bitmap_zero(map, page->objects);
4093 get_map(s, page, map);
4094
4095 for_each_object(p, s, addr, page->objects)
4096 if (!test_bit(slab_index(p, s, addr), map))
4097 add_location(t, s, get_track(s, p, alloc));
4098 }
4099
4100 static int list_locations(struct kmem_cache *s, char *buf,
4101 enum track_item alloc)
4102 {
4103 int len = 0;
4104 unsigned long i;
4105 struct loc_track t = { 0, 0, NULL };
4106 int node;
4107 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4108 sizeof(unsigned long), GFP_KERNEL);
4109 struct kmem_cache_node *n;
4110
4111 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4112 GFP_TEMPORARY)) {
4113 kfree(map);
4114 return sprintf(buf, "Out of memory\n");
4115 }
4116 /* Push back cpu slabs */
4117 flush_all(s);
4118
4119 for_each_kmem_cache_node(s, node, n) {
4120 unsigned long flags;
4121 struct page *page;
4122
4123 if (!atomic_long_read(&n->nr_slabs))
4124 continue;
4125
4126 spin_lock_irqsave(&n->list_lock, flags);
4127 list_for_each_entry(page, &n->partial, lru)
4128 process_slab(&t, s, page, alloc, map);
4129 list_for_each_entry(page, &n->full, lru)
4130 process_slab(&t, s, page, alloc, map);
4131 spin_unlock_irqrestore(&n->list_lock, flags);
4132 }
4133
4134 for (i = 0; i < t.count; i++) {
4135 struct location *l = &t.loc[i];
4136
4137 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4138 break;
4139 len += sprintf(buf + len, "%7ld ", l->count);
4140
4141 if (l->addr)
4142 len += sprintf(buf + len, "%pS", (void *)l->addr);
4143 else
4144 len += sprintf(buf + len, "<not-available>");
4145
4146 if (l->sum_time != l->min_time) {
4147 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4148 l->min_time,
4149 (long)div_u64(l->sum_time, l->count),
4150 l->max_time);
4151 } else
4152 len += sprintf(buf + len, " age=%ld",
4153 l->min_time);
4154
4155 if (l->min_pid != l->max_pid)
4156 len += sprintf(buf + len, " pid=%ld-%ld",
4157 l->min_pid, l->max_pid);
4158 else
4159 len += sprintf(buf + len, " pid=%ld",
4160 l->min_pid);
4161
4162 if (num_online_cpus() > 1 &&
4163 !cpumask_empty(to_cpumask(l->cpus)) &&
4164 len < PAGE_SIZE - 60)
4165 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4166 " cpus=%*pbl",
4167 cpumask_pr_args(to_cpumask(l->cpus)));
4168
4169 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4170 len < PAGE_SIZE - 60)
4171 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4172 " nodes=%*pbl",
4173 nodemask_pr_args(&l->nodes));
4174
4175 len += sprintf(buf + len, "\n");
4176 }
4177
4178 free_loc_track(&t);
4179 kfree(map);
4180 if (!t.count)
4181 len += sprintf(buf, "No data\n");
4182 return len;
4183 }
4184 #endif
4185
4186 #ifdef SLUB_RESILIENCY_TEST
4187 static void __init resiliency_test(void)
4188 {
4189 u8 *p;
4190
4191 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4192
4193 pr_err("SLUB resiliency testing\n");
4194 pr_err("-----------------------\n");
4195 pr_err("A. Corruption after allocation\n");
4196
4197 p = kzalloc(16, GFP_KERNEL);
4198 p[16] = 0x12;
4199 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4200 p + 16);
4201
4202 validate_slab_cache(kmalloc_caches[4]);
4203
4204 /* Hmmm... The next two are dangerous */
4205 p = kzalloc(32, GFP_KERNEL);
4206 p[32 + sizeof(void *)] = 0x34;
4207 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4208 p);
4209 pr_err("If allocated object is overwritten then not detectable\n\n");
4210
4211 validate_slab_cache(kmalloc_caches[5]);
4212 p = kzalloc(64, GFP_KERNEL);
4213 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4214 *p = 0x56;
4215 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4216 p);
4217 pr_err("If allocated object is overwritten then not detectable\n\n");
4218 validate_slab_cache(kmalloc_caches[6]);
4219
4220 pr_err("\nB. Corruption after free\n");
4221 p = kzalloc(128, GFP_KERNEL);
4222 kfree(p);
4223 *p = 0x78;
4224 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4225 validate_slab_cache(kmalloc_caches[7]);
4226
4227 p = kzalloc(256, GFP_KERNEL);
4228 kfree(p);
4229 p[50] = 0x9a;
4230 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4231 validate_slab_cache(kmalloc_caches[8]);
4232
4233 p = kzalloc(512, GFP_KERNEL);
4234 kfree(p);
4235 p[512] = 0xab;
4236 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4237 validate_slab_cache(kmalloc_caches[9]);
4238 }
4239 #else
4240 #ifdef CONFIG_SYSFS
4241 static void resiliency_test(void) {};
4242 #endif
4243 #endif
4244
4245 #ifdef CONFIG_SYSFS
4246 enum slab_stat_type {
4247 SL_ALL, /* All slabs */
4248 SL_PARTIAL, /* Only partially allocated slabs */
4249 SL_CPU, /* Only slabs used for cpu caches */
4250 SL_OBJECTS, /* Determine allocated objects not slabs */
4251 SL_TOTAL /* Determine object capacity not slabs */
4252 };
4253
4254 #define SO_ALL (1 << SL_ALL)
4255 #define SO_PARTIAL (1 << SL_PARTIAL)
4256 #define SO_CPU (1 << SL_CPU)
4257 #define SO_OBJECTS (1 << SL_OBJECTS)
4258 #define SO_TOTAL (1 << SL_TOTAL)
4259
4260 static ssize_t show_slab_objects(struct kmem_cache *s,
4261 char *buf, unsigned long flags)
4262 {
4263 unsigned long total = 0;
4264 int node;
4265 int x;
4266 unsigned long *nodes;
4267
4268 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4269 if (!nodes)
4270 return -ENOMEM;
4271
4272 if (flags & SO_CPU) {
4273 int cpu;
4274
4275 for_each_possible_cpu(cpu) {
4276 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4277 cpu);
4278 int node;
4279 struct page *page;
4280
4281 page = READ_ONCE(c->page);
4282 if (!page)
4283 continue;
4284
4285 node = page_to_nid(page);
4286 if (flags & SO_TOTAL)
4287 x = page->objects;
4288 else if (flags & SO_OBJECTS)
4289 x = page->inuse;
4290 else
4291 x = 1;
4292
4293 total += x;
4294 nodes[node] += x;
4295
4296 page = READ_ONCE(c->partial);
4297 if (page) {
4298 node = page_to_nid(page);
4299 if (flags & SO_TOTAL)
4300 WARN_ON_ONCE(1);
4301 else if (flags & SO_OBJECTS)
4302 WARN_ON_ONCE(1);
4303 else
4304 x = page->pages;
4305 total += x;
4306 nodes[node] += x;
4307 }
4308 }
4309 }
4310
4311 get_online_mems();
4312 #ifdef CONFIG_SLUB_DEBUG
4313 if (flags & SO_ALL) {
4314 struct kmem_cache_node *n;
4315
4316 for_each_kmem_cache_node(s, node, n) {
4317
4318 if (flags & SO_TOTAL)
4319 x = atomic_long_read(&n->total_objects);
4320 else if (flags & SO_OBJECTS)
4321 x = atomic_long_read(&n->total_objects) -
4322 count_partial(n, count_free);
4323 else
4324 x = atomic_long_read(&n->nr_slabs);
4325 total += x;
4326 nodes[node] += x;
4327 }
4328
4329 } else
4330 #endif
4331 if (flags & SO_PARTIAL) {
4332 struct kmem_cache_node *n;
4333
4334 for_each_kmem_cache_node(s, node, n) {
4335 if (flags & SO_TOTAL)
4336 x = count_partial(n, count_total);
4337 else if (flags & SO_OBJECTS)
4338 x = count_partial(n, count_inuse);
4339 else
4340 x = n->nr_partial;
4341 total += x;
4342 nodes[node] += x;
4343 }
4344 }
4345 x = sprintf(buf, "%lu", total);
4346 #ifdef CONFIG_NUMA
4347 for (node = 0; node < nr_node_ids; node++)
4348 if (nodes[node])
4349 x += sprintf(buf + x, " N%d=%lu",
4350 node, nodes[node]);
4351 #endif
4352 put_online_mems();
4353 kfree(nodes);
4354 return x + sprintf(buf + x, "\n");
4355 }
4356
4357 #ifdef CONFIG_SLUB_DEBUG
4358 static int any_slab_objects(struct kmem_cache *s)
4359 {
4360 int node;
4361 struct kmem_cache_node *n;
4362
4363 for_each_kmem_cache_node(s, node, n)
4364 if (atomic_long_read(&n->total_objects))
4365 return 1;
4366
4367 return 0;
4368 }
4369 #endif
4370
4371 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4372 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4373
4374 struct slab_attribute {
4375 struct attribute attr;
4376 ssize_t (*show)(struct kmem_cache *s, char *buf);
4377 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4378 };
4379
4380 #define SLAB_ATTR_RO(_name) \
4381 static struct slab_attribute _name##_attr = \
4382 __ATTR(_name, 0400, _name##_show, NULL)
4383
4384 #define SLAB_ATTR(_name) \
4385 static struct slab_attribute _name##_attr = \
4386 __ATTR(_name, 0600, _name##_show, _name##_store)
4387
4388 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4389 {
4390 return sprintf(buf, "%d\n", s->size);
4391 }
4392 SLAB_ATTR_RO(slab_size);
4393
4394 static ssize_t align_show(struct kmem_cache *s, char *buf)
4395 {
4396 return sprintf(buf, "%d\n", s->align);
4397 }
4398 SLAB_ATTR_RO(align);
4399
4400 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4401 {
4402 return sprintf(buf, "%d\n", s->object_size);
4403 }
4404 SLAB_ATTR_RO(object_size);
4405
4406 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4407 {
4408 return sprintf(buf, "%d\n", oo_objects(s->oo));
4409 }
4410 SLAB_ATTR_RO(objs_per_slab);
4411
4412 static ssize_t order_store(struct kmem_cache *s,
4413 const char *buf, size_t length)
4414 {
4415 unsigned long order;
4416 int err;
4417
4418 err = kstrtoul(buf, 10, &order);
4419 if (err)
4420 return err;
4421
4422 if (order > slub_max_order || order < slub_min_order)
4423 return -EINVAL;
4424
4425 calculate_sizes(s, order);
4426 return length;
4427 }
4428
4429 static ssize_t order_show(struct kmem_cache *s, char *buf)
4430 {
4431 return sprintf(buf, "%d\n", oo_order(s->oo));
4432 }
4433 SLAB_ATTR(order);
4434
4435 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4436 {
4437 return sprintf(buf, "%lu\n", s->min_partial);
4438 }
4439
4440 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4441 size_t length)
4442 {
4443 unsigned long min;
4444 int err;
4445
4446 err = kstrtoul(buf, 10, &min);
4447 if (err)
4448 return err;
4449
4450 set_min_partial(s, min);
4451 return length;
4452 }
4453 SLAB_ATTR(min_partial);
4454
4455 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4456 {
4457 return sprintf(buf, "%u\n", s->cpu_partial);
4458 }
4459
4460 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4461 size_t length)
4462 {
4463 unsigned long objects;
4464 int err;
4465
4466 err = kstrtoul(buf, 10, &objects);
4467 if (err)
4468 return err;
4469 if (objects && !kmem_cache_has_cpu_partial(s))
4470 return -EINVAL;
4471
4472 s->cpu_partial = objects;
4473 flush_all(s);
4474 return length;
4475 }
4476 SLAB_ATTR(cpu_partial);
4477
4478 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4479 {
4480 if (!s->ctor)
4481 return 0;
4482 return sprintf(buf, "%pS\n", s->ctor);
4483 }
4484 SLAB_ATTR_RO(ctor);
4485
4486 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4487 {
4488 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
4489 }
4490 SLAB_ATTR_RO(aliases);
4491
4492 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4493 {
4494 return show_slab_objects(s, buf, SO_PARTIAL);
4495 }
4496 SLAB_ATTR_RO(partial);
4497
4498 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4499 {
4500 return show_slab_objects(s, buf, SO_CPU);
4501 }
4502 SLAB_ATTR_RO(cpu_slabs);
4503
4504 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4505 {
4506 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4507 }
4508 SLAB_ATTR_RO(objects);
4509
4510 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4511 {
4512 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4513 }
4514 SLAB_ATTR_RO(objects_partial);
4515
4516 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4517 {
4518 int objects = 0;
4519 int pages = 0;
4520 int cpu;
4521 int len;
4522
4523 for_each_online_cpu(cpu) {
4524 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4525
4526 if (page) {
4527 pages += page->pages;
4528 objects += page->pobjects;
4529 }
4530 }
4531
4532 len = sprintf(buf, "%d(%d)", objects, pages);
4533
4534 #ifdef CONFIG_SMP
4535 for_each_online_cpu(cpu) {
4536 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4537
4538 if (page && len < PAGE_SIZE - 20)
4539 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4540 page->pobjects, page->pages);
4541 }
4542 #endif
4543 return len + sprintf(buf + len, "\n");
4544 }
4545 SLAB_ATTR_RO(slabs_cpu_partial);
4546
4547 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4548 {
4549 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4550 }
4551
4552 static ssize_t reclaim_account_store(struct kmem_cache *s,
4553 const char *buf, size_t length)
4554 {
4555 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4556 if (buf[0] == '1')
4557 s->flags |= SLAB_RECLAIM_ACCOUNT;
4558 return length;
4559 }
4560 SLAB_ATTR(reclaim_account);
4561
4562 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4563 {
4564 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4565 }
4566 SLAB_ATTR_RO(hwcache_align);
4567
4568 #ifdef CONFIG_ZONE_DMA
4569 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4570 {
4571 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4572 }
4573 SLAB_ATTR_RO(cache_dma);
4574 #endif
4575
4576 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4577 {
4578 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4579 }
4580 SLAB_ATTR_RO(destroy_by_rcu);
4581
4582 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4583 {
4584 return sprintf(buf, "%d\n", s->reserved);
4585 }
4586 SLAB_ATTR_RO(reserved);
4587
4588 #ifdef CONFIG_SLUB_DEBUG
4589 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4590 {
4591 return show_slab_objects(s, buf, SO_ALL);
4592 }
4593 SLAB_ATTR_RO(slabs);
4594
4595 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4596 {
4597 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4598 }
4599 SLAB_ATTR_RO(total_objects);
4600
4601 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4602 {
4603 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4604 }
4605
4606 static ssize_t sanity_checks_store(struct kmem_cache *s,
4607 const char *buf, size_t length)
4608 {
4609 s->flags &= ~SLAB_DEBUG_FREE;
4610 if (buf[0] == '1') {
4611 s->flags &= ~__CMPXCHG_DOUBLE;
4612 s->flags |= SLAB_DEBUG_FREE;
4613 }
4614 return length;
4615 }
4616 SLAB_ATTR(sanity_checks);
4617
4618 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4619 {
4620 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4621 }
4622
4623 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4624 size_t length)
4625 {
4626 /*
4627 * Tracing a merged cache is going to give confusing results
4628 * as well as cause other issues like converting a mergeable
4629 * cache into an umergeable one.
4630 */
4631 if (s->refcount > 1)
4632 return -EINVAL;
4633
4634 s->flags &= ~SLAB_TRACE;
4635 if (buf[0] == '1') {
4636 s->flags &= ~__CMPXCHG_DOUBLE;
4637 s->flags |= SLAB_TRACE;
4638 }
4639 return length;
4640 }
4641 SLAB_ATTR(trace);
4642
4643 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4644 {
4645 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4646 }
4647
4648 static ssize_t red_zone_store(struct kmem_cache *s,
4649 const char *buf, size_t length)
4650 {
4651 if (any_slab_objects(s))
4652 return -EBUSY;
4653
4654 s->flags &= ~SLAB_RED_ZONE;
4655 if (buf[0] == '1') {
4656 s->flags &= ~__CMPXCHG_DOUBLE;
4657 s->flags |= SLAB_RED_ZONE;
4658 }
4659 calculate_sizes(s, -1);
4660 return length;
4661 }
4662 SLAB_ATTR(red_zone);
4663
4664 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4665 {
4666 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4667 }
4668
4669 static ssize_t poison_store(struct kmem_cache *s,
4670 const char *buf, size_t length)
4671 {
4672 if (any_slab_objects(s))
4673 return -EBUSY;
4674
4675 s->flags &= ~SLAB_POISON;
4676 if (buf[0] == '1') {
4677 s->flags &= ~__CMPXCHG_DOUBLE;
4678 s->flags |= SLAB_POISON;
4679 }
4680 calculate_sizes(s, -1);
4681 return length;
4682 }
4683 SLAB_ATTR(poison);
4684
4685 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4686 {
4687 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4688 }
4689
4690 static ssize_t store_user_store(struct kmem_cache *s,
4691 const char *buf, size_t length)
4692 {
4693 if (any_slab_objects(s))
4694 return -EBUSY;
4695
4696 s->flags &= ~SLAB_STORE_USER;
4697 if (buf[0] == '1') {
4698 s->flags &= ~__CMPXCHG_DOUBLE;
4699 s->flags |= SLAB_STORE_USER;
4700 }
4701 calculate_sizes(s, -1);
4702 return length;
4703 }
4704 SLAB_ATTR(store_user);
4705
4706 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4707 {
4708 return 0;
4709 }
4710
4711 static ssize_t validate_store(struct kmem_cache *s,
4712 const char *buf, size_t length)
4713 {
4714 int ret = -EINVAL;
4715
4716 if (buf[0] == '1') {
4717 ret = validate_slab_cache(s);
4718 if (ret >= 0)
4719 ret = length;
4720 }
4721 return ret;
4722 }
4723 SLAB_ATTR(validate);
4724
4725 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4726 {
4727 if (!(s->flags & SLAB_STORE_USER))
4728 return -ENOSYS;
4729 return list_locations(s, buf, TRACK_ALLOC);
4730 }
4731 SLAB_ATTR_RO(alloc_calls);
4732
4733 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4734 {
4735 if (!(s->flags & SLAB_STORE_USER))
4736 return -ENOSYS;
4737 return list_locations(s, buf, TRACK_FREE);
4738 }
4739 SLAB_ATTR_RO(free_calls);
4740 #endif /* CONFIG_SLUB_DEBUG */
4741
4742 #ifdef CONFIG_FAILSLAB
4743 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4744 {
4745 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4746 }
4747
4748 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4749 size_t length)
4750 {
4751 if (s->refcount > 1)
4752 return -EINVAL;
4753
4754 s->flags &= ~SLAB_FAILSLAB;
4755 if (buf[0] == '1')
4756 s->flags |= SLAB_FAILSLAB;
4757 return length;
4758 }
4759 SLAB_ATTR(failslab);
4760 #endif
4761
4762 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4763 {
4764 return 0;
4765 }
4766
4767 static ssize_t shrink_store(struct kmem_cache *s,
4768 const char *buf, size_t length)
4769 {
4770 if (buf[0] == '1')
4771 kmem_cache_shrink(s);
4772 else
4773 return -EINVAL;
4774 return length;
4775 }
4776 SLAB_ATTR(shrink);
4777
4778 #ifdef CONFIG_NUMA
4779 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4780 {
4781 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4782 }
4783
4784 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4785 const char *buf, size_t length)
4786 {
4787 unsigned long ratio;
4788 int err;
4789
4790 err = kstrtoul(buf, 10, &ratio);
4791 if (err)
4792 return err;
4793
4794 if (ratio <= 100)
4795 s->remote_node_defrag_ratio = ratio * 10;
4796
4797 return length;
4798 }
4799 SLAB_ATTR(remote_node_defrag_ratio);
4800 #endif
4801
4802 #ifdef CONFIG_SLUB_STATS
4803 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4804 {
4805 unsigned long sum = 0;
4806 int cpu;
4807 int len;
4808 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4809
4810 if (!data)
4811 return -ENOMEM;
4812
4813 for_each_online_cpu(cpu) {
4814 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4815
4816 data[cpu] = x;
4817 sum += x;
4818 }
4819
4820 len = sprintf(buf, "%lu", sum);
4821
4822 #ifdef CONFIG_SMP
4823 for_each_online_cpu(cpu) {
4824 if (data[cpu] && len < PAGE_SIZE - 20)
4825 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4826 }
4827 #endif
4828 kfree(data);
4829 return len + sprintf(buf + len, "\n");
4830 }
4831
4832 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4833 {
4834 int cpu;
4835
4836 for_each_online_cpu(cpu)
4837 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4838 }
4839
4840 #define STAT_ATTR(si, text) \
4841 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4842 { \
4843 return show_stat(s, buf, si); \
4844 } \
4845 static ssize_t text##_store(struct kmem_cache *s, \
4846 const char *buf, size_t length) \
4847 { \
4848 if (buf[0] != '0') \
4849 return -EINVAL; \
4850 clear_stat(s, si); \
4851 return length; \
4852 } \
4853 SLAB_ATTR(text); \
4854
4855 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4856 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4857 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4858 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4859 STAT_ATTR(FREE_FROZEN, free_frozen);
4860 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4861 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4862 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4863 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4864 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4865 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
4866 STAT_ATTR(FREE_SLAB, free_slab);
4867 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4868 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4869 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4870 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4871 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4872 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4873 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
4874 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4875 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4876 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4877 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
4878 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
4879 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
4880 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
4881 #endif
4882
4883 static struct attribute *slab_attrs[] = {
4884 &slab_size_attr.attr,
4885 &object_size_attr.attr,
4886 &objs_per_slab_attr.attr,
4887 &order_attr.attr,
4888 &min_partial_attr.attr,
4889 &cpu_partial_attr.attr,
4890 &objects_attr.attr,
4891 &objects_partial_attr.attr,
4892 &partial_attr.attr,
4893 &cpu_slabs_attr.attr,
4894 &ctor_attr.attr,
4895 &aliases_attr.attr,
4896 &align_attr.attr,
4897 &hwcache_align_attr.attr,
4898 &reclaim_account_attr.attr,
4899 &destroy_by_rcu_attr.attr,
4900 &shrink_attr.attr,
4901 &reserved_attr.attr,
4902 &slabs_cpu_partial_attr.attr,
4903 #ifdef CONFIG_SLUB_DEBUG
4904 &total_objects_attr.attr,
4905 &slabs_attr.attr,
4906 &sanity_checks_attr.attr,
4907 &trace_attr.attr,
4908 &red_zone_attr.attr,
4909 &poison_attr.attr,
4910 &store_user_attr.attr,
4911 &validate_attr.attr,
4912 &alloc_calls_attr.attr,
4913 &free_calls_attr.attr,
4914 #endif
4915 #ifdef CONFIG_ZONE_DMA
4916 &cache_dma_attr.attr,
4917 #endif
4918 #ifdef CONFIG_NUMA
4919 &remote_node_defrag_ratio_attr.attr,
4920 #endif
4921 #ifdef CONFIG_SLUB_STATS
4922 &alloc_fastpath_attr.attr,
4923 &alloc_slowpath_attr.attr,
4924 &free_fastpath_attr.attr,
4925 &free_slowpath_attr.attr,
4926 &free_frozen_attr.attr,
4927 &free_add_partial_attr.attr,
4928 &free_remove_partial_attr.attr,
4929 &alloc_from_partial_attr.attr,
4930 &alloc_slab_attr.attr,
4931 &alloc_refill_attr.attr,
4932 &alloc_node_mismatch_attr.attr,
4933 &free_slab_attr.attr,
4934 &cpuslab_flush_attr.attr,
4935 &deactivate_full_attr.attr,
4936 &deactivate_empty_attr.attr,
4937 &deactivate_to_head_attr.attr,
4938 &deactivate_to_tail_attr.attr,
4939 &deactivate_remote_frees_attr.attr,
4940 &deactivate_bypass_attr.attr,
4941 &order_fallback_attr.attr,
4942 &cmpxchg_double_fail_attr.attr,
4943 &cmpxchg_double_cpu_fail_attr.attr,
4944 &cpu_partial_alloc_attr.attr,
4945 &cpu_partial_free_attr.attr,
4946 &cpu_partial_node_attr.attr,
4947 &cpu_partial_drain_attr.attr,
4948 #endif
4949 #ifdef CONFIG_FAILSLAB
4950 &failslab_attr.attr,
4951 #endif
4952
4953 NULL
4954 };
4955
4956 static struct attribute_group slab_attr_group = {
4957 .attrs = slab_attrs,
4958 };
4959
4960 static ssize_t slab_attr_show(struct kobject *kobj,
4961 struct attribute *attr,
4962 char *buf)
4963 {
4964 struct slab_attribute *attribute;
4965 struct kmem_cache *s;
4966 int err;
4967
4968 attribute = to_slab_attr(attr);
4969 s = to_slab(kobj);
4970
4971 if (!attribute->show)
4972 return -EIO;
4973
4974 err = attribute->show(s, buf);
4975
4976 return err;
4977 }
4978
4979 static ssize_t slab_attr_store(struct kobject *kobj,
4980 struct attribute *attr,
4981 const char *buf, size_t len)
4982 {
4983 struct slab_attribute *attribute;
4984 struct kmem_cache *s;
4985 int err;
4986
4987 attribute = to_slab_attr(attr);
4988 s = to_slab(kobj);
4989
4990 if (!attribute->store)
4991 return -EIO;
4992
4993 err = attribute->store(s, buf, len);
4994 #ifdef CONFIG_MEMCG_KMEM
4995 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
4996 struct kmem_cache *c;
4997
4998 mutex_lock(&slab_mutex);
4999 if (s->max_attr_size < len)
5000 s->max_attr_size = len;
5001
5002 /*
5003 * This is a best effort propagation, so this function's return
5004 * value will be determined by the parent cache only. This is
5005 * basically because not all attributes will have a well
5006 * defined semantics for rollbacks - most of the actions will
5007 * have permanent effects.
5008 *
5009 * Returning the error value of any of the children that fail
5010 * is not 100 % defined, in the sense that users seeing the
5011 * error code won't be able to know anything about the state of
5012 * the cache.
5013 *
5014 * Only returning the error code for the parent cache at least
5015 * has well defined semantics. The cache being written to
5016 * directly either failed or succeeded, in which case we loop
5017 * through the descendants with best-effort propagation.
5018 */
5019 for_each_memcg_cache(c, s)
5020 attribute->store(c, buf, len);
5021 mutex_unlock(&slab_mutex);
5022 }
5023 #endif
5024 return err;
5025 }
5026
5027 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5028 {
5029 #ifdef CONFIG_MEMCG_KMEM
5030 int i;
5031 char *buffer = NULL;
5032 struct kmem_cache *root_cache;
5033
5034 if (is_root_cache(s))
5035 return;
5036
5037 root_cache = s->memcg_params.root_cache;
5038
5039 /*
5040 * This mean this cache had no attribute written. Therefore, no point
5041 * in copying default values around
5042 */
5043 if (!root_cache->max_attr_size)
5044 return;
5045
5046 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5047 char mbuf[64];
5048 char *buf;
5049 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5050
5051 if (!attr || !attr->store || !attr->show)
5052 continue;
5053
5054 /*
5055 * It is really bad that we have to allocate here, so we will
5056 * do it only as a fallback. If we actually allocate, though,
5057 * we can just use the allocated buffer until the end.
5058 *
5059 * Most of the slub attributes will tend to be very small in
5060 * size, but sysfs allows buffers up to a page, so they can
5061 * theoretically happen.
5062 */
5063 if (buffer)
5064 buf = buffer;
5065 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5066 buf = mbuf;
5067 else {
5068 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5069 if (WARN_ON(!buffer))
5070 continue;
5071 buf = buffer;
5072 }
5073
5074 attr->show(root_cache, buf);
5075 attr->store(s, buf, strlen(buf));
5076 }
5077
5078 if (buffer)
5079 free_page((unsigned long)buffer);
5080 #endif
5081 }
5082
5083 static void kmem_cache_release(struct kobject *k)
5084 {
5085 slab_kmem_cache_release(to_slab(k));
5086 }
5087
5088 static const struct sysfs_ops slab_sysfs_ops = {
5089 .show = slab_attr_show,
5090 .store = slab_attr_store,
5091 };
5092
5093 static struct kobj_type slab_ktype = {
5094 .sysfs_ops = &slab_sysfs_ops,
5095 .release = kmem_cache_release,
5096 };
5097
5098 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5099 {
5100 struct kobj_type *ktype = get_ktype(kobj);
5101
5102 if (ktype == &slab_ktype)
5103 return 1;
5104 return 0;
5105 }
5106
5107 static const struct kset_uevent_ops slab_uevent_ops = {
5108 .filter = uevent_filter,
5109 };
5110
5111 static struct kset *slab_kset;
5112
5113 static inline struct kset *cache_kset(struct kmem_cache *s)
5114 {
5115 #ifdef CONFIG_MEMCG_KMEM
5116 if (!is_root_cache(s))
5117 return s->memcg_params.root_cache->memcg_kset;
5118 #endif
5119 return slab_kset;
5120 }
5121
5122 #define ID_STR_LENGTH 64
5123
5124 /* Create a unique string id for a slab cache:
5125 *
5126 * Format :[flags-]size
5127 */
5128 static char *create_unique_id(struct kmem_cache *s)
5129 {
5130 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5131 char *p = name;
5132
5133 BUG_ON(!name);
5134
5135 *p++ = ':';
5136 /*
5137 * First flags affecting slabcache operations. We will only
5138 * get here for aliasable slabs so we do not need to support
5139 * too many flags. The flags here must cover all flags that
5140 * are matched during merging to guarantee that the id is
5141 * unique.
5142 */
5143 if (s->flags & SLAB_CACHE_DMA)
5144 *p++ = 'd';
5145 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5146 *p++ = 'a';
5147 if (s->flags & SLAB_DEBUG_FREE)
5148 *p++ = 'F';
5149 if (!(s->flags & SLAB_NOTRACK))
5150 *p++ = 't';
5151 if (p != name + 1)
5152 *p++ = '-';
5153 p += sprintf(p, "%07d", s->size);
5154
5155 BUG_ON(p > name + ID_STR_LENGTH - 1);
5156 return name;
5157 }
5158
5159 static int sysfs_slab_add(struct kmem_cache *s)
5160 {
5161 int err;
5162 const char *name;
5163 int unmergeable = slab_unmergeable(s);
5164
5165 if (unmergeable) {
5166 /*
5167 * Slabcache can never be merged so we can use the name proper.
5168 * This is typically the case for debug situations. In that
5169 * case we can catch duplicate names easily.
5170 */
5171 sysfs_remove_link(&slab_kset->kobj, s->name);
5172 name = s->name;
5173 } else {
5174 /*
5175 * Create a unique name for the slab as a target
5176 * for the symlinks.
5177 */
5178 name = create_unique_id(s);
5179 }
5180
5181 s->kobj.kset = cache_kset(s);
5182 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5183 if (err)
5184 goto out_put_kobj;
5185
5186 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5187 if (err)
5188 goto out_del_kobj;
5189
5190 #ifdef CONFIG_MEMCG_KMEM
5191 if (is_root_cache(s)) {
5192 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5193 if (!s->memcg_kset) {
5194 err = -ENOMEM;
5195 goto out_del_kobj;
5196 }
5197 }
5198 #endif
5199
5200 kobject_uevent(&s->kobj, KOBJ_ADD);
5201 if (!unmergeable) {
5202 /* Setup first alias */
5203 sysfs_slab_alias(s, s->name);
5204 }
5205 out:
5206 if (!unmergeable)
5207 kfree(name);
5208 return err;
5209 out_del_kobj:
5210 kobject_del(&s->kobj);
5211 out_put_kobj:
5212 kobject_put(&s->kobj);
5213 goto out;
5214 }
5215
5216 void sysfs_slab_remove(struct kmem_cache *s)
5217 {
5218 if (slab_state < FULL)
5219 /*
5220 * Sysfs has not been setup yet so no need to remove the
5221 * cache from sysfs.
5222 */
5223 return;
5224
5225 #ifdef CONFIG_MEMCG_KMEM
5226 kset_unregister(s->memcg_kset);
5227 #endif
5228 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5229 kobject_del(&s->kobj);
5230 kobject_put(&s->kobj);
5231 }
5232
5233 /*
5234 * Need to buffer aliases during bootup until sysfs becomes
5235 * available lest we lose that information.
5236 */
5237 struct saved_alias {
5238 struct kmem_cache *s;
5239 const char *name;
5240 struct saved_alias *next;
5241 };
5242
5243 static struct saved_alias *alias_list;
5244
5245 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5246 {
5247 struct saved_alias *al;
5248
5249 if (slab_state == FULL) {
5250 /*
5251 * If we have a leftover link then remove it.
5252 */
5253 sysfs_remove_link(&slab_kset->kobj, name);
5254 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5255 }
5256
5257 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5258 if (!al)
5259 return -ENOMEM;
5260
5261 al->s = s;
5262 al->name = name;
5263 al->next = alias_list;
5264 alias_list = al;
5265 return 0;
5266 }
5267
5268 static int __init slab_sysfs_init(void)
5269 {
5270 struct kmem_cache *s;
5271 int err;
5272
5273 mutex_lock(&slab_mutex);
5274
5275 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5276 if (!slab_kset) {
5277 mutex_unlock(&slab_mutex);
5278 pr_err("Cannot register slab subsystem.\n");
5279 return -ENOSYS;
5280 }
5281
5282 slab_state = FULL;
5283
5284 list_for_each_entry(s, &slab_caches, list) {
5285 err = sysfs_slab_add(s);
5286 if (err)
5287 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5288 s->name);
5289 }
5290
5291 while (alias_list) {
5292 struct saved_alias *al = alias_list;
5293
5294 alias_list = alias_list->next;
5295 err = sysfs_slab_alias(al->s, al->name);
5296 if (err)
5297 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5298 al->name);
5299 kfree(al);
5300 }
5301
5302 mutex_unlock(&slab_mutex);
5303 resiliency_test();
5304 return 0;
5305 }
5306
5307 __initcall(slab_sysfs_init);
5308 #endif /* CONFIG_SYSFS */
5309
5310 /*
5311 * The /proc/slabinfo ABI
5312 */
5313 #ifdef CONFIG_SLABINFO
5314 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5315 {
5316 unsigned long nr_slabs = 0;
5317 unsigned long nr_objs = 0;
5318 unsigned long nr_free = 0;
5319 int node;
5320 struct kmem_cache_node *n;
5321
5322 for_each_kmem_cache_node(s, node, n) {
5323 nr_slabs += node_nr_slabs(n);
5324 nr_objs += node_nr_objs(n);
5325 nr_free += count_partial(n, count_free);
5326 }
5327
5328 sinfo->active_objs = nr_objs - nr_free;
5329 sinfo->num_objs = nr_objs;
5330 sinfo->active_slabs = nr_slabs;
5331 sinfo->num_slabs = nr_slabs;
5332 sinfo->objects_per_slab = oo_objects(s->oo);
5333 sinfo->cache_order = oo_order(s->oo);
5334 }
5335
5336 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5337 {
5338 }
5339
5340 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5341 size_t count, loff_t *ppos)
5342 {
5343 return -EIO;
5344 }
5345 #endif /* CONFIG_SLABINFO */