xref: /openbmc/linux/mm/slub.c (revision 6774def6)
1 /*
2  * SLUB: A slab allocator that limits cache line use instead of queuing
3  * objects in per cpu and per node lists.
4  *
5  * The allocator synchronizes using per slab locks or atomic operatios
6  * and only uses a centralized lock to manage a pool of partial slabs.
7  *
8  * (C) 2007 SGI, Christoph Lameter
9  * (C) 2011 Linux Foundation, Christoph Lameter
10  */
11 
12 #include <linux/mm.h>
13 #include <linux/swap.h> /* struct reclaim_state */
14 #include <linux/module.h>
15 #include <linux/bit_spinlock.h>
16 #include <linux/interrupt.h>
17 #include <linux/bitops.h>
18 #include <linux/slab.h>
19 #include "slab.h"
20 #include <linux/proc_fs.h>
21 #include <linux/notifier.h>
22 #include <linux/seq_file.h>
23 #include <linux/kmemcheck.h>
24 #include <linux/cpu.h>
25 #include <linux/cpuset.h>
26 #include <linux/mempolicy.h>
27 #include <linux/ctype.h>
28 #include <linux/debugobjects.h>
29 #include <linux/kallsyms.h>
30 #include <linux/memory.h>
31 #include <linux/math64.h>
32 #include <linux/fault-inject.h>
33 #include <linux/stacktrace.h>
34 #include <linux/prefetch.h>
35 #include <linux/memcontrol.h>
36 
37 #include <trace/events/kmem.h>
38 
39 #include "internal.h"
40 
41 /*
42  * Lock order:
43  *   1. slab_mutex (Global Mutex)
44  *   2. node->list_lock
45  *   3. slab_lock(page) (Only on some arches and for debugging)
46  *
47  *   slab_mutex
48  *
49  *   The role of the slab_mutex is to protect the list of all the slabs
50  *   and to synchronize major metadata changes to slab cache structures.
51  *
52  *   The slab_lock is only used for debugging and on arches that do not
53  *   have the ability to do a cmpxchg_double. It only protects the second
54  *   double word in the page struct. Meaning
55  *	A. page->freelist	-> List of object free in a page
56  *	B. page->counters	-> Counters of objects
57  *	C. page->frozen		-> frozen state
58  *
59  *   If a slab is frozen then it is exempt from list management. It is not
60  *   on any list. The processor that froze the slab is the one who can
61  *   perform list operations on the page. Other processors may put objects
62  *   onto the freelist but the processor that froze the slab is the only
63  *   one that can retrieve the objects from the page's freelist.
64  *
65  *   The list_lock protects the partial and full list on each node and
66  *   the partial slab counter. If taken then no new slabs may be added or
67  *   removed from the lists nor make the number of partial slabs be modified.
68  *   (Note that the total number of slabs is an atomic value that may be
69  *   modified without taking the list lock).
70  *
71  *   The list_lock is a centralized lock and thus we avoid taking it as
72  *   much as possible. As long as SLUB does not have to handle partial
73  *   slabs, operations can continue without any centralized lock. F.e.
74  *   allocating a long series of objects that fill up slabs does not require
75  *   the list lock.
76  *   Interrupts are disabled during allocation and deallocation in order to
77  *   make the slab allocator safe to use in the context of an irq. In addition
78  *   interrupts are disabled to ensure that the processor does not change
79  *   while handling per_cpu slabs, due to kernel preemption.
80  *
81  * SLUB assigns one slab for allocation to each processor.
82  * Allocations only occur from these slabs called cpu slabs.
83  *
84  * Slabs with free elements are kept on a partial list and during regular
85  * operations no list for full slabs is used. If an object in a full slab is
86  * freed then the slab will show up again on the partial lists.
87  * We track full slabs for debugging purposes though because otherwise we
88  * cannot scan all objects.
89  *
90  * Slabs are freed when they become empty. Teardown and setup is
91  * minimal so we rely on the page allocators per cpu caches for
92  * fast frees and allocs.
93  *
94  * Overloading of page flags that are otherwise used for LRU management.
95  *
96  * PageActive 		The slab is frozen and exempt from list processing.
97  * 			This means that the slab is dedicated to a purpose
98  * 			such as satisfying allocations for a specific
99  * 			processor. Objects may be freed in the slab while
100  * 			it is frozen but slab_free will then skip the usual
101  * 			list operations. It is up to the processor holding
102  * 			the slab to integrate the slab into the slab lists
103  * 			when the slab is no longer needed.
104  *
105  * 			One use of this flag is to mark slabs that are
106  * 			used for allocations. Then such a slab becomes a cpu
107  * 			slab. The cpu slab may be equipped with an additional
108  * 			freelist that allows lockless access to
109  * 			free objects in addition to the regular freelist
110  * 			that requires the slab lock.
111  *
112  * PageError		Slab requires special handling due to debug
113  * 			options set. This moves	slab handling out of
114  * 			the fast path and disables lockless freelists.
115  */
116 
117 static inline int kmem_cache_debug(struct kmem_cache *s)
118 {
119 #ifdef CONFIG_SLUB_DEBUG
120 	return unlikely(s->flags & SLAB_DEBUG_FLAGS);
121 #else
122 	return 0;
123 #endif
124 }
125 
126 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
127 {
128 #ifdef CONFIG_SLUB_CPU_PARTIAL
129 	return !kmem_cache_debug(s);
130 #else
131 	return false;
132 #endif
133 }
134 
135 /*
136  * Issues still to be resolved:
137  *
138  * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
139  *
140  * - Variable sizing of the per node arrays
141  */
142 
143 /* Enable to test recovery from slab corruption on boot */
144 #undef SLUB_RESILIENCY_TEST
145 
146 /* Enable to log cmpxchg failures */
147 #undef SLUB_DEBUG_CMPXCHG
148 
149 /*
150  * Mininum number of partial slabs. These will be left on the partial
151  * lists even if they are empty. kmem_cache_shrink may reclaim them.
152  */
153 #define MIN_PARTIAL 5
154 
155 /*
156  * Maximum number of desirable partial slabs.
157  * The existence of more partial slabs makes kmem_cache_shrink
158  * sort the partial list by the number of objects in use.
159  */
160 #define MAX_PARTIAL 10
161 
162 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
163 				SLAB_POISON | SLAB_STORE_USER)
164 
165 /*
166  * Debugging flags that require metadata to be stored in the slab.  These get
167  * disabled when slub_debug=O is used and a cache's min order increases with
168  * metadata.
169  */
170 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
171 
172 #define OO_SHIFT	16
173 #define OO_MASK		((1 << OO_SHIFT) - 1)
174 #define MAX_OBJS_PER_PAGE	32767 /* since page.objects is u15 */
175 
176 /* Internal SLUB flags */
177 #define __OBJECT_POISON		0x80000000UL /* Poison object */
178 #define __CMPXCHG_DOUBLE	0x40000000UL /* Use cmpxchg_double */
179 
180 #ifdef CONFIG_SMP
181 static struct notifier_block slab_notifier;
182 #endif
183 
184 /*
185  * Tracking user of a slab.
186  */
187 #define TRACK_ADDRS_COUNT 16
188 struct track {
189 	unsigned long addr;	/* Called from address */
190 #ifdef CONFIG_STACKTRACE
191 	unsigned long addrs[TRACK_ADDRS_COUNT];	/* Called from address */
192 #endif
193 	int cpu;		/* Was running on cpu */
194 	int pid;		/* Pid context */
195 	unsigned long when;	/* When did the operation occur */
196 };
197 
198 enum track_item { TRACK_ALLOC, TRACK_FREE };
199 
200 #ifdef CONFIG_SYSFS
201 static int sysfs_slab_add(struct kmem_cache *);
202 static int sysfs_slab_alias(struct kmem_cache *, const char *);
203 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
204 #else
205 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
206 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
207 							{ return 0; }
208 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
209 #endif
210 
211 static inline void stat(const struct kmem_cache *s, enum stat_item si)
212 {
213 #ifdef CONFIG_SLUB_STATS
214 	/*
215 	 * The rmw is racy on a preemptible kernel but this is acceptable, so
216 	 * avoid this_cpu_add()'s irq-disable overhead.
217 	 */
218 	raw_cpu_inc(s->cpu_slab->stat[si]);
219 #endif
220 }
221 
222 /********************************************************************
223  * 			Core slab cache functions
224  *******************************************************************/
225 
226 /* Verify that a pointer has an address that is valid within a slab page */
227 static inline int check_valid_pointer(struct kmem_cache *s,
228 				struct page *page, const void *object)
229 {
230 	void *base;
231 
232 	if (!object)
233 		return 1;
234 
235 	base = page_address(page);
236 	if (object < base || object >= base + page->objects * s->size ||
237 		(object - base) % s->size) {
238 		return 0;
239 	}
240 
241 	return 1;
242 }
243 
244 static inline void *get_freepointer(struct kmem_cache *s, void *object)
245 {
246 	return *(void **)(object + s->offset);
247 }
248 
249 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
250 {
251 	prefetch(object + s->offset);
252 }
253 
254 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
255 {
256 	void *p;
257 
258 #ifdef CONFIG_DEBUG_PAGEALLOC
259 	probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
260 #else
261 	p = get_freepointer(s, object);
262 #endif
263 	return p;
264 }
265 
266 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
267 {
268 	*(void **)(object + s->offset) = fp;
269 }
270 
271 /* Loop over all objects in a slab */
272 #define for_each_object(__p, __s, __addr, __objects) \
273 	for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
274 			__p += (__s)->size)
275 
276 #define for_each_object_idx(__p, __idx, __s, __addr, __objects) \
277 	for (__p = (__addr), __idx = 1; __idx <= __objects;\
278 			__p += (__s)->size, __idx++)
279 
280 /* Determine object index from a given position */
281 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
282 {
283 	return (p - addr) / s->size;
284 }
285 
286 static inline size_t slab_ksize(const struct kmem_cache *s)
287 {
288 #ifdef CONFIG_SLUB_DEBUG
289 	/*
290 	 * Debugging requires use of the padding between object
291 	 * and whatever may come after it.
292 	 */
293 	if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
294 		return s->object_size;
295 
296 #endif
297 	/*
298 	 * If we have the need to store the freelist pointer
299 	 * back there or track user information then we can
300 	 * only use the space before that information.
301 	 */
302 	if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
303 		return s->inuse;
304 	/*
305 	 * Else we can use all the padding etc for the allocation
306 	 */
307 	return s->size;
308 }
309 
310 static inline int order_objects(int order, unsigned long size, int reserved)
311 {
312 	return ((PAGE_SIZE << order) - reserved) / size;
313 }
314 
315 static inline struct kmem_cache_order_objects oo_make(int order,
316 		unsigned long size, int reserved)
317 {
318 	struct kmem_cache_order_objects x = {
319 		(order << OO_SHIFT) + order_objects(order, size, reserved)
320 	};
321 
322 	return x;
323 }
324 
325 static inline int oo_order(struct kmem_cache_order_objects x)
326 {
327 	return x.x >> OO_SHIFT;
328 }
329 
330 static inline int oo_objects(struct kmem_cache_order_objects x)
331 {
332 	return x.x & OO_MASK;
333 }
334 
335 /*
336  * Per slab locking using the pagelock
337  */
338 static __always_inline void slab_lock(struct page *page)
339 {
340 	bit_spin_lock(PG_locked, &page->flags);
341 }
342 
343 static __always_inline void slab_unlock(struct page *page)
344 {
345 	__bit_spin_unlock(PG_locked, &page->flags);
346 }
347 
348 static inline void set_page_slub_counters(struct page *page, unsigned long counters_new)
349 {
350 	struct page tmp;
351 	tmp.counters = counters_new;
352 	/*
353 	 * page->counters can cover frozen/inuse/objects as well
354 	 * as page->_count.  If we assign to ->counters directly
355 	 * we run the risk of losing updates to page->_count, so
356 	 * be careful and only assign to the fields we need.
357 	 */
358 	page->frozen  = tmp.frozen;
359 	page->inuse   = tmp.inuse;
360 	page->objects = tmp.objects;
361 }
362 
363 /* Interrupts must be disabled (for the fallback code to work right) */
364 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
365 		void *freelist_old, unsigned long counters_old,
366 		void *freelist_new, unsigned long counters_new,
367 		const char *n)
368 {
369 	VM_BUG_ON(!irqs_disabled());
370 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
371     defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
372 	if (s->flags & __CMPXCHG_DOUBLE) {
373 		if (cmpxchg_double(&page->freelist, &page->counters,
374 				   freelist_old, counters_old,
375 				   freelist_new, counters_new))
376 			return 1;
377 	} else
378 #endif
379 	{
380 		slab_lock(page);
381 		if (page->freelist == freelist_old &&
382 					page->counters == counters_old) {
383 			page->freelist = freelist_new;
384 			set_page_slub_counters(page, counters_new);
385 			slab_unlock(page);
386 			return 1;
387 		}
388 		slab_unlock(page);
389 	}
390 
391 	cpu_relax();
392 	stat(s, CMPXCHG_DOUBLE_FAIL);
393 
394 #ifdef SLUB_DEBUG_CMPXCHG
395 	pr_info("%s %s: cmpxchg double redo ", n, s->name);
396 #endif
397 
398 	return 0;
399 }
400 
401 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
402 		void *freelist_old, unsigned long counters_old,
403 		void *freelist_new, unsigned long counters_new,
404 		const char *n)
405 {
406 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
407     defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
408 	if (s->flags & __CMPXCHG_DOUBLE) {
409 		if (cmpxchg_double(&page->freelist, &page->counters,
410 				   freelist_old, counters_old,
411 				   freelist_new, counters_new))
412 			return 1;
413 	} else
414 #endif
415 	{
416 		unsigned long flags;
417 
418 		local_irq_save(flags);
419 		slab_lock(page);
420 		if (page->freelist == freelist_old &&
421 					page->counters == counters_old) {
422 			page->freelist = freelist_new;
423 			set_page_slub_counters(page, counters_new);
424 			slab_unlock(page);
425 			local_irq_restore(flags);
426 			return 1;
427 		}
428 		slab_unlock(page);
429 		local_irq_restore(flags);
430 	}
431 
432 	cpu_relax();
433 	stat(s, CMPXCHG_DOUBLE_FAIL);
434 
435 #ifdef SLUB_DEBUG_CMPXCHG
436 	pr_info("%s %s: cmpxchg double redo ", n, s->name);
437 #endif
438 
439 	return 0;
440 }
441 
442 #ifdef CONFIG_SLUB_DEBUG
443 /*
444  * Determine a map of object in use on a page.
445  *
446  * Node listlock must be held to guarantee that the page does
447  * not vanish from under us.
448  */
449 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
450 {
451 	void *p;
452 	void *addr = page_address(page);
453 
454 	for (p = page->freelist; p; p = get_freepointer(s, p))
455 		set_bit(slab_index(p, s, addr), map);
456 }
457 
458 /*
459  * Debug settings:
460  */
461 #ifdef CONFIG_SLUB_DEBUG_ON
462 static int slub_debug = DEBUG_DEFAULT_FLAGS;
463 #else
464 static int slub_debug;
465 #endif
466 
467 static char *slub_debug_slabs;
468 static int disable_higher_order_debug;
469 
470 /*
471  * Object debugging
472  */
473 static void print_section(char *text, u8 *addr, unsigned int length)
474 {
475 	print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
476 			length, 1);
477 }
478 
479 static struct track *get_track(struct kmem_cache *s, void *object,
480 	enum track_item alloc)
481 {
482 	struct track *p;
483 
484 	if (s->offset)
485 		p = object + s->offset + sizeof(void *);
486 	else
487 		p = object + s->inuse;
488 
489 	return p + alloc;
490 }
491 
492 static void set_track(struct kmem_cache *s, void *object,
493 			enum track_item alloc, unsigned long addr)
494 {
495 	struct track *p = get_track(s, object, alloc);
496 
497 	if (addr) {
498 #ifdef CONFIG_STACKTRACE
499 		struct stack_trace trace;
500 		int i;
501 
502 		trace.nr_entries = 0;
503 		trace.max_entries = TRACK_ADDRS_COUNT;
504 		trace.entries = p->addrs;
505 		trace.skip = 3;
506 		save_stack_trace(&trace);
507 
508 		/* See rant in lockdep.c */
509 		if (trace.nr_entries != 0 &&
510 		    trace.entries[trace.nr_entries - 1] == ULONG_MAX)
511 			trace.nr_entries--;
512 
513 		for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
514 			p->addrs[i] = 0;
515 #endif
516 		p->addr = addr;
517 		p->cpu = smp_processor_id();
518 		p->pid = current->pid;
519 		p->when = jiffies;
520 	} else
521 		memset(p, 0, sizeof(struct track));
522 }
523 
524 static void init_tracking(struct kmem_cache *s, void *object)
525 {
526 	if (!(s->flags & SLAB_STORE_USER))
527 		return;
528 
529 	set_track(s, object, TRACK_FREE, 0UL);
530 	set_track(s, object, TRACK_ALLOC, 0UL);
531 }
532 
533 static void print_track(const char *s, struct track *t)
534 {
535 	if (!t->addr)
536 		return;
537 
538 	pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
539 	       s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
540 #ifdef CONFIG_STACKTRACE
541 	{
542 		int i;
543 		for (i = 0; i < TRACK_ADDRS_COUNT; i++)
544 			if (t->addrs[i])
545 				pr_err("\t%pS\n", (void *)t->addrs[i]);
546 			else
547 				break;
548 	}
549 #endif
550 }
551 
552 static void print_tracking(struct kmem_cache *s, void *object)
553 {
554 	if (!(s->flags & SLAB_STORE_USER))
555 		return;
556 
557 	print_track("Allocated", get_track(s, object, TRACK_ALLOC));
558 	print_track("Freed", get_track(s, object, TRACK_FREE));
559 }
560 
561 static void print_page_info(struct page *page)
562 {
563 	pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
564 	       page, page->objects, page->inuse, page->freelist, page->flags);
565 
566 }
567 
568 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
569 {
570 	struct va_format vaf;
571 	va_list args;
572 
573 	va_start(args, fmt);
574 	vaf.fmt = fmt;
575 	vaf.va = &args;
576 	pr_err("=============================================================================\n");
577 	pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
578 	pr_err("-----------------------------------------------------------------------------\n\n");
579 
580 	add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
581 	va_end(args);
582 }
583 
584 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
585 {
586 	struct va_format vaf;
587 	va_list args;
588 
589 	va_start(args, fmt);
590 	vaf.fmt = fmt;
591 	vaf.va = &args;
592 	pr_err("FIX %s: %pV\n", s->name, &vaf);
593 	va_end(args);
594 }
595 
596 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
597 {
598 	unsigned int off;	/* Offset of last byte */
599 	u8 *addr = page_address(page);
600 
601 	print_tracking(s, p);
602 
603 	print_page_info(page);
604 
605 	pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
606 	       p, p - addr, get_freepointer(s, p));
607 
608 	if (p > addr + 16)
609 		print_section("Bytes b4 ", p - 16, 16);
610 
611 	print_section("Object ", p, min_t(unsigned long, s->object_size,
612 				PAGE_SIZE));
613 	if (s->flags & SLAB_RED_ZONE)
614 		print_section("Redzone ", p + s->object_size,
615 			s->inuse - s->object_size);
616 
617 	if (s->offset)
618 		off = s->offset + sizeof(void *);
619 	else
620 		off = s->inuse;
621 
622 	if (s->flags & SLAB_STORE_USER)
623 		off += 2 * sizeof(struct track);
624 
625 	if (off != s->size)
626 		/* Beginning of the filler is the free pointer */
627 		print_section("Padding ", p + off, s->size - off);
628 
629 	dump_stack();
630 }
631 
632 static void object_err(struct kmem_cache *s, struct page *page,
633 			u8 *object, char *reason)
634 {
635 	slab_bug(s, "%s", reason);
636 	print_trailer(s, page, object);
637 }
638 
639 static void slab_err(struct kmem_cache *s, struct page *page,
640 			const char *fmt, ...)
641 {
642 	va_list args;
643 	char buf[100];
644 
645 	va_start(args, fmt);
646 	vsnprintf(buf, sizeof(buf), fmt, args);
647 	va_end(args);
648 	slab_bug(s, "%s", buf);
649 	print_page_info(page);
650 	dump_stack();
651 }
652 
653 static void init_object(struct kmem_cache *s, void *object, u8 val)
654 {
655 	u8 *p = object;
656 
657 	if (s->flags & __OBJECT_POISON) {
658 		memset(p, POISON_FREE, s->object_size - 1);
659 		p[s->object_size - 1] = POISON_END;
660 	}
661 
662 	if (s->flags & SLAB_RED_ZONE)
663 		memset(p + s->object_size, val, s->inuse - s->object_size);
664 }
665 
666 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
667 						void *from, void *to)
668 {
669 	slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
670 	memset(from, data, to - from);
671 }
672 
673 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
674 			u8 *object, char *what,
675 			u8 *start, unsigned int value, unsigned int bytes)
676 {
677 	u8 *fault;
678 	u8 *end;
679 
680 	fault = memchr_inv(start, value, bytes);
681 	if (!fault)
682 		return 1;
683 
684 	end = start + bytes;
685 	while (end > fault && end[-1] == value)
686 		end--;
687 
688 	slab_bug(s, "%s overwritten", what);
689 	pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
690 					fault, end - 1, fault[0], value);
691 	print_trailer(s, page, object);
692 
693 	restore_bytes(s, what, value, fault, end);
694 	return 0;
695 }
696 
697 /*
698  * Object layout:
699  *
700  * object address
701  * 	Bytes of the object to be managed.
702  * 	If the freepointer may overlay the object then the free
703  * 	pointer is the first word of the object.
704  *
705  * 	Poisoning uses 0x6b (POISON_FREE) and the last byte is
706  * 	0xa5 (POISON_END)
707  *
708  * object + s->object_size
709  * 	Padding to reach word boundary. This is also used for Redzoning.
710  * 	Padding is extended by another word if Redzoning is enabled and
711  * 	object_size == inuse.
712  *
713  * 	We fill with 0xbb (RED_INACTIVE) for inactive objects and with
714  * 	0xcc (RED_ACTIVE) for objects in use.
715  *
716  * object + s->inuse
717  * 	Meta data starts here.
718  *
719  * 	A. Free pointer (if we cannot overwrite object on free)
720  * 	B. Tracking data for SLAB_STORE_USER
721  * 	C. Padding to reach required alignment boundary or at mininum
722  * 		one word if debugging is on to be able to detect writes
723  * 		before the word boundary.
724  *
725  *	Padding is done using 0x5a (POISON_INUSE)
726  *
727  * object + s->size
728  * 	Nothing is used beyond s->size.
729  *
730  * If slabcaches are merged then the object_size and inuse boundaries are mostly
731  * ignored. And therefore no slab options that rely on these boundaries
732  * may be used with merged slabcaches.
733  */
734 
735 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
736 {
737 	unsigned long off = s->inuse;	/* The end of info */
738 
739 	if (s->offset)
740 		/* Freepointer is placed after the object. */
741 		off += sizeof(void *);
742 
743 	if (s->flags & SLAB_STORE_USER)
744 		/* We also have user information there */
745 		off += 2 * sizeof(struct track);
746 
747 	if (s->size == off)
748 		return 1;
749 
750 	return check_bytes_and_report(s, page, p, "Object padding",
751 				p + off, POISON_INUSE, s->size - off);
752 }
753 
754 /* Check the pad bytes at the end of a slab page */
755 static int slab_pad_check(struct kmem_cache *s, struct page *page)
756 {
757 	u8 *start;
758 	u8 *fault;
759 	u8 *end;
760 	int length;
761 	int remainder;
762 
763 	if (!(s->flags & SLAB_POISON))
764 		return 1;
765 
766 	start = page_address(page);
767 	length = (PAGE_SIZE << compound_order(page)) - s->reserved;
768 	end = start + length;
769 	remainder = length % s->size;
770 	if (!remainder)
771 		return 1;
772 
773 	fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
774 	if (!fault)
775 		return 1;
776 	while (end > fault && end[-1] == POISON_INUSE)
777 		end--;
778 
779 	slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
780 	print_section("Padding ", end - remainder, remainder);
781 
782 	restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
783 	return 0;
784 }
785 
786 static int check_object(struct kmem_cache *s, struct page *page,
787 					void *object, u8 val)
788 {
789 	u8 *p = object;
790 	u8 *endobject = object + s->object_size;
791 
792 	if (s->flags & SLAB_RED_ZONE) {
793 		if (!check_bytes_and_report(s, page, object, "Redzone",
794 			endobject, val, s->inuse - s->object_size))
795 			return 0;
796 	} else {
797 		if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
798 			check_bytes_and_report(s, page, p, "Alignment padding",
799 				endobject, POISON_INUSE,
800 				s->inuse - s->object_size);
801 		}
802 	}
803 
804 	if (s->flags & SLAB_POISON) {
805 		if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
806 			(!check_bytes_and_report(s, page, p, "Poison", p,
807 					POISON_FREE, s->object_size - 1) ||
808 			 !check_bytes_and_report(s, page, p, "Poison",
809 				p + s->object_size - 1, POISON_END, 1)))
810 			return 0;
811 		/*
812 		 * check_pad_bytes cleans up on its own.
813 		 */
814 		check_pad_bytes(s, page, p);
815 	}
816 
817 	if (!s->offset && val == SLUB_RED_ACTIVE)
818 		/*
819 		 * Object and freepointer overlap. Cannot check
820 		 * freepointer while object is allocated.
821 		 */
822 		return 1;
823 
824 	/* Check free pointer validity */
825 	if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
826 		object_err(s, page, p, "Freepointer corrupt");
827 		/*
828 		 * No choice but to zap it and thus lose the remainder
829 		 * of the free objects in this slab. May cause
830 		 * another error because the object count is now wrong.
831 		 */
832 		set_freepointer(s, p, NULL);
833 		return 0;
834 	}
835 	return 1;
836 }
837 
838 static int check_slab(struct kmem_cache *s, struct page *page)
839 {
840 	int maxobj;
841 
842 	VM_BUG_ON(!irqs_disabled());
843 
844 	if (!PageSlab(page)) {
845 		slab_err(s, page, "Not a valid slab page");
846 		return 0;
847 	}
848 
849 	maxobj = order_objects(compound_order(page), s->size, s->reserved);
850 	if (page->objects > maxobj) {
851 		slab_err(s, page, "objects %u > max %u",
852 			s->name, page->objects, maxobj);
853 		return 0;
854 	}
855 	if (page->inuse > page->objects) {
856 		slab_err(s, page, "inuse %u > max %u",
857 			s->name, page->inuse, page->objects);
858 		return 0;
859 	}
860 	/* Slab_pad_check fixes things up after itself */
861 	slab_pad_check(s, page);
862 	return 1;
863 }
864 
865 /*
866  * Determine if a certain object on a page is on the freelist. Must hold the
867  * slab lock to guarantee that the chains are in a consistent state.
868  */
869 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
870 {
871 	int nr = 0;
872 	void *fp;
873 	void *object = NULL;
874 	unsigned long max_objects;
875 
876 	fp = page->freelist;
877 	while (fp && nr <= page->objects) {
878 		if (fp == search)
879 			return 1;
880 		if (!check_valid_pointer(s, page, fp)) {
881 			if (object) {
882 				object_err(s, page, object,
883 					"Freechain corrupt");
884 				set_freepointer(s, object, NULL);
885 			} else {
886 				slab_err(s, page, "Freepointer corrupt");
887 				page->freelist = NULL;
888 				page->inuse = page->objects;
889 				slab_fix(s, "Freelist cleared");
890 				return 0;
891 			}
892 			break;
893 		}
894 		object = fp;
895 		fp = get_freepointer(s, object);
896 		nr++;
897 	}
898 
899 	max_objects = order_objects(compound_order(page), s->size, s->reserved);
900 	if (max_objects > MAX_OBJS_PER_PAGE)
901 		max_objects = MAX_OBJS_PER_PAGE;
902 
903 	if (page->objects != max_objects) {
904 		slab_err(s, page, "Wrong number of objects. Found %d but "
905 			"should be %d", page->objects, max_objects);
906 		page->objects = max_objects;
907 		slab_fix(s, "Number of objects adjusted.");
908 	}
909 	if (page->inuse != page->objects - nr) {
910 		slab_err(s, page, "Wrong object count. Counter is %d but "
911 			"counted were %d", page->inuse, page->objects - nr);
912 		page->inuse = page->objects - nr;
913 		slab_fix(s, "Object count adjusted.");
914 	}
915 	return search == NULL;
916 }
917 
918 static void trace(struct kmem_cache *s, struct page *page, void *object,
919 								int alloc)
920 {
921 	if (s->flags & SLAB_TRACE) {
922 		pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
923 			s->name,
924 			alloc ? "alloc" : "free",
925 			object, page->inuse,
926 			page->freelist);
927 
928 		if (!alloc)
929 			print_section("Object ", (void *)object,
930 					s->object_size);
931 
932 		dump_stack();
933 	}
934 }
935 
936 /*
937  * Tracking of fully allocated slabs for debugging purposes.
938  */
939 static void add_full(struct kmem_cache *s,
940 	struct kmem_cache_node *n, struct page *page)
941 {
942 	if (!(s->flags & SLAB_STORE_USER))
943 		return;
944 
945 	lockdep_assert_held(&n->list_lock);
946 	list_add(&page->lru, &n->full);
947 }
948 
949 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
950 {
951 	if (!(s->flags & SLAB_STORE_USER))
952 		return;
953 
954 	lockdep_assert_held(&n->list_lock);
955 	list_del(&page->lru);
956 }
957 
958 /* Tracking of the number of slabs for debugging purposes */
959 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
960 {
961 	struct kmem_cache_node *n = get_node(s, node);
962 
963 	return atomic_long_read(&n->nr_slabs);
964 }
965 
966 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
967 {
968 	return atomic_long_read(&n->nr_slabs);
969 }
970 
971 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
972 {
973 	struct kmem_cache_node *n = get_node(s, node);
974 
975 	/*
976 	 * May be called early in order to allocate a slab for the
977 	 * kmem_cache_node structure. Solve the chicken-egg
978 	 * dilemma by deferring the increment of the count during
979 	 * bootstrap (see early_kmem_cache_node_alloc).
980 	 */
981 	if (likely(n)) {
982 		atomic_long_inc(&n->nr_slabs);
983 		atomic_long_add(objects, &n->total_objects);
984 	}
985 }
986 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
987 {
988 	struct kmem_cache_node *n = get_node(s, node);
989 
990 	atomic_long_dec(&n->nr_slabs);
991 	atomic_long_sub(objects, &n->total_objects);
992 }
993 
994 /* Object debug checks for alloc/free paths */
995 static void setup_object_debug(struct kmem_cache *s, struct page *page,
996 								void *object)
997 {
998 	if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
999 		return;
1000 
1001 	init_object(s, object, SLUB_RED_INACTIVE);
1002 	init_tracking(s, object);
1003 }
1004 
1005 static noinline int alloc_debug_processing(struct kmem_cache *s,
1006 					struct page *page,
1007 					void *object, unsigned long addr)
1008 {
1009 	if (!check_slab(s, page))
1010 		goto bad;
1011 
1012 	if (!check_valid_pointer(s, page, object)) {
1013 		object_err(s, page, object, "Freelist Pointer check fails");
1014 		goto bad;
1015 	}
1016 
1017 	if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1018 		goto bad;
1019 
1020 	/* Success perform special debug activities for allocs */
1021 	if (s->flags & SLAB_STORE_USER)
1022 		set_track(s, object, TRACK_ALLOC, addr);
1023 	trace(s, page, object, 1);
1024 	init_object(s, object, SLUB_RED_ACTIVE);
1025 	return 1;
1026 
1027 bad:
1028 	if (PageSlab(page)) {
1029 		/*
1030 		 * If this is a slab page then lets do the best we can
1031 		 * to avoid issues in the future. Marking all objects
1032 		 * as used avoids touching the remaining objects.
1033 		 */
1034 		slab_fix(s, "Marking all objects used");
1035 		page->inuse = page->objects;
1036 		page->freelist = NULL;
1037 	}
1038 	return 0;
1039 }
1040 
1041 static noinline struct kmem_cache_node *free_debug_processing(
1042 	struct kmem_cache *s, struct page *page, void *object,
1043 	unsigned long addr, unsigned long *flags)
1044 {
1045 	struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1046 
1047 	spin_lock_irqsave(&n->list_lock, *flags);
1048 	slab_lock(page);
1049 
1050 	if (!check_slab(s, page))
1051 		goto fail;
1052 
1053 	if (!check_valid_pointer(s, page, object)) {
1054 		slab_err(s, page, "Invalid object pointer 0x%p", object);
1055 		goto fail;
1056 	}
1057 
1058 	if (on_freelist(s, page, object)) {
1059 		object_err(s, page, object, "Object already free");
1060 		goto fail;
1061 	}
1062 
1063 	if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1064 		goto out;
1065 
1066 	if (unlikely(s != page->slab_cache)) {
1067 		if (!PageSlab(page)) {
1068 			slab_err(s, page, "Attempt to free object(0x%p) "
1069 				"outside of slab", object);
1070 		} else if (!page->slab_cache) {
1071 			pr_err("SLUB <none>: no slab for object 0x%p.\n",
1072 			       object);
1073 			dump_stack();
1074 		} else
1075 			object_err(s, page, object,
1076 					"page slab pointer corrupt.");
1077 		goto fail;
1078 	}
1079 
1080 	if (s->flags & SLAB_STORE_USER)
1081 		set_track(s, object, TRACK_FREE, addr);
1082 	trace(s, page, object, 0);
1083 	init_object(s, object, SLUB_RED_INACTIVE);
1084 out:
1085 	slab_unlock(page);
1086 	/*
1087 	 * Keep node_lock to preserve integrity
1088 	 * until the object is actually freed
1089 	 */
1090 	return n;
1091 
1092 fail:
1093 	slab_unlock(page);
1094 	spin_unlock_irqrestore(&n->list_lock, *flags);
1095 	slab_fix(s, "Object at 0x%p not freed", object);
1096 	return NULL;
1097 }
1098 
1099 static int __init setup_slub_debug(char *str)
1100 {
1101 	slub_debug = DEBUG_DEFAULT_FLAGS;
1102 	if (*str++ != '=' || !*str)
1103 		/*
1104 		 * No options specified. Switch on full debugging.
1105 		 */
1106 		goto out;
1107 
1108 	if (*str == ',')
1109 		/*
1110 		 * No options but restriction on slabs. This means full
1111 		 * debugging for slabs matching a pattern.
1112 		 */
1113 		goto check_slabs;
1114 
1115 	if (tolower(*str) == 'o') {
1116 		/*
1117 		 * Avoid enabling debugging on caches if its minimum order
1118 		 * would increase as a result.
1119 		 */
1120 		disable_higher_order_debug = 1;
1121 		goto out;
1122 	}
1123 
1124 	slub_debug = 0;
1125 	if (*str == '-')
1126 		/*
1127 		 * Switch off all debugging measures.
1128 		 */
1129 		goto out;
1130 
1131 	/*
1132 	 * Determine which debug features should be switched on
1133 	 */
1134 	for (; *str && *str != ','; str++) {
1135 		switch (tolower(*str)) {
1136 		case 'f':
1137 			slub_debug |= SLAB_DEBUG_FREE;
1138 			break;
1139 		case 'z':
1140 			slub_debug |= SLAB_RED_ZONE;
1141 			break;
1142 		case 'p':
1143 			slub_debug |= SLAB_POISON;
1144 			break;
1145 		case 'u':
1146 			slub_debug |= SLAB_STORE_USER;
1147 			break;
1148 		case 't':
1149 			slub_debug |= SLAB_TRACE;
1150 			break;
1151 		case 'a':
1152 			slub_debug |= SLAB_FAILSLAB;
1153 			break;
1154 		default:
1155 			pr_err("slub_debug option '%c' unknown. skipped\n",
1156 			       *str);
1157 		}
1158 	}
1159 
1160 check_slabs:
1161 	if (*str == ',')
1162 		slub_debug_slabs = str + 1;
1163 out:
1164 	return 1;
1165 }
1166 
1167 __setup("slub_debug", setup_slub_debug);
1168 
1169 unsigned long kmem_cache_flags(unsigned long object_size,
1170 	unsigned long flags, const char *name,
1171 	void (*ctor)(void *))
1172 {
1173 	/*
1174 	 * Enable debugging if selected on the kernel commandline.
1175 	 */
1176 	if (slub_debug && (!slub_debug_slabs || (name &&
1177 		!strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1178 		flags |= slub_debug;
1179 
1180 	return flags;
1181 }
1182 #else
1183 static inline void setup_object_debug(struct kmem_cache *s,
1184 			struct page *page, void *object) {}
1185 
1186 static inline int alloc_debug_processing(struct kmem_cache *s,
1187 	struct page *page, void *object, unsigned long addr) { return 0; }
1188 
1189 static inline struct kmem_cache_node *free_debug_processing(
1190 	struct kmem_cache *s, struct page *page, void *object,
1191 	unsigned long addr, unsigned long *flags) { return NULL; }
1192 
1193 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1194 			{ return 1; }
1195 static inline int check_object(struct kmem_cache *s, struct page *page,
1196 			void *object, u8 val) { return 1; }
1197 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1198 					struct page *page) {}
1199 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1200 					struct page *page) {}
1201 unsigned long kmem_cache_flags(unsigned long object_size,
1202 	unsigned long flags, const char *name,
1203 	void (*ctor)(void *))
1204 {
1205 	return flags;
1206 }
1207 #define slub_debug 0
1208 
1209 #define disable_higher_order_debug 0
1210 
1211 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1212 							{ return 0; }
1213 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1214 							{ return 0; }
1215 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1216 							int objects) {}
1217 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1218 							int objects) {}
1219 
1220 #endif /* CONFIG_SLUB_DEBUG */
1221 
1222 /*
1223  * Hooks for other subsystems that check memory allocations. In a typical
1224  * production configuration these hooks all should produce no code at all.
1225  */
1226 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1227 {
1228 	kmemleak_alloc(ptr, size, 1, flags);
1229 }
1230 
1231 static inline void kfree_hook(const void *x)
1232 {
1233 	kmemleak_free(x);
1234 }
1235 
1236 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1237 {
1238 	flags &= gfp_allowed_mask;
1239 	lockdep_trace_alloc(flags);
1240 	might_sleep_if(flags & __GFP_WAIT);
1241 
1242 	return should_failslab(s->object_size, flags, s->flags);
1243 }
1244 
1245 static inline void slab_post_alloc_hook(struct kmem_cache *s,
1246 					gfp_t flags, void *object)
1247 {
1248 	flags &= gfp_allowed_mask;
1249 	kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
1250 	kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
1251 }
1252 
1253 static inline void slab_free_hook(struct kmem_cache *s, void *x)
1254 {
1255 	kmemleak_free_recursive(x, s->flags);
1256 
1257 	/*
1258 	 * Trouble is that we may no longer disable interrupts in the fast path
1259 	 * So in order to make the debug calls that expect irqs to be
1260 	 * disabled we need to disable interrupts temporarily.
1261 	 */
1262 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1263 	{
1264 		unsigned long flags;
1265 
1266 		local_irq_save(flags);
1267 		kmemcheck_slab_free(s, x, s->object_size);
1268 		debug_check_no_locks_freed(x, s->object_size);
1269 		local_irq_restore(flags);
1270 	}
1271 #endif
1272 	if (!(s->flags & SLAB_DEBUG_OBJECTS))
1273 		debug_check_no_obj_freed(x, s->object_size);
1274 }
1275 
1276 /*
1277  * Slab allocation and freeing
1278  */
1279 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1280 		gfp_t flags, int node, struct kmem_cache_order_objects oo)
1281 {
1282 	struct page *page;
1283 	int order = oo_order(oo);
1284 
1285 	flags |= __GFP_NOTRACK;
1286 
1287 	if (memcg_charge_slab(s, flags, order))
1288 		return NULL;
1289 
1290 	if (node == NUMA_NO_NODE)
1291 		page = alloc_pages(flags, order);
1292 	else
1293 		page = alloc_pages_exact_node(node, flags, order);
1294 
1295 	if (!page)
1296 		memcg_uncharge_slab(s, order);
1297 
1298 	return page;
1299 }
1300 
1301 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1302 {
1303 	struct page *page;
1304 	struct kmem_cache_order_objects oo = s->oo;
1305 	gfp_t alloc_gfp;
1306 
1307 	flags &= gfp_allowed_mask;
1308 
1309 	if (flags & __GFP_WAIT)
1310 		local_irq_enable();
1311 
1312 	flags |= s->allocflags;
1313 
1314 	/*
1315 	 * Let the initial higher-order allocation fail under memory pressure
1316 	 * so we fall-back to the minimum order allocation.
1317 	 */
1318 	alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1319 
1320 	page = alloc_slab_page(s, alloc_gfp, node, oo);
1321 	if (unlikely(!page)) {
1322 		oo = s->min;
1323 		alloc_gfp = flags;
1324 		/*
1325 		 * Allocation may have failed due to fragmentation.
1326 		 * Try a lower order alloc if possible
1327 		 */
1328 		page = alloc_slab_page(s, alloc_gfp, node, oo);
1329 
1330 		if (page)
1331 			stat(s, ORDER_FALLBACK);
1332 	}
1333 
1334 	if (kmemcheck_enabled && page
1335 		&& !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1336 		int pages = 1 << oo_order(oo);
1337 
1338 		kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node);
1339 
1340 		/*
1341 		 * Objects from caches that have a constructor don't get
1342 		 * cleared when they're allocated, so we need to do it here.
1343 		 */
1344 		if (s->ctor)
1345 			kmemcheck_mark_uninitialized_pages(page, pages);
1346 		else
1347 			kmemcheck_mark_unallocated_pages(page, pages);
1348 	}
1349 
1350 	if (flags & __GFP_WAIT)
1351 		local_irq_disable();
1352 	if (!page)
1353 		return NULL;
1354 
1355 	page->objects = oo_objects(oo);
1356 	mod_zone_page_state(page_zone(page),
1357 		(s->flags & SLAB_RECLAIM_ACCOUNT) ?
1358 		NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1359 		1 << oo_order(oo));
1360 
1361 	return page;
1362 }
1363 
1364 static void setup_object(struct kmem_cache *s, struct page *page,
1365 				void *object)
1366 {
1367 	setup_object_debug(s, page, object);
1368 	if (unlikely(s->ctor))
1369 		s->ctor(object);
1370 }
1371 
1372 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1373 {
1374 	struct page *page;
1375 	void *start;
1376 	void *p;
1377 	int order;
1378 	int idx;
1379 
1380 	BUG_ON(flags & GFP_SLAB_BUG_MASK);
1381 
1382 	page = allocate_slab(s,
1383 		flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1384 	if (!page)
1385 		goto out;
1386 
1387 	order = compound_order(page);
1388 	inc_slabs_node(s, page_to_nid(page), page->objects);
1389 	page->slab_cache = s;
1390 	__SetPageSlab(page);
1391 	if (page->pfmemalloc)
1392 		SetPageSlabPfmemalloc(page);
1393 
1394 	start = page_address(page);
1395 
1396 	if (unlikely(s->flags & SLAB_POISON))
1397 		memset(start, POISON_INUSE, PAGE_SIZE << order);
1398 
1399 	for_each_object_idx(p, idx, s, start, page->objects) {
1400 		setup_object(s, page, p);
1401 		if (likely(idx < page->objects))
1402 			set_freepointer(s, p, p + s->size);
1403 		else
1404 			set_freepointer(s, p, NULL);
1405 	}
1406 
1407 	page->freelist = start;
1408 	page->inuse = page->objects;
1409 	page->frozen = 1;
1410 out:
1411 	return page;
1412 }
1413 
1414 static void __free_slab(struct kmem_cache *s, struct page *page)
1415 {
1416 	int order = compound_order(page);
1417 	int pages = 1 << order;
1418 
1419 	if (kmem_cache_debug(s)) {
1420 		void *p;
1421 
1422 		slab_pad_check(s, page);
1423 		for_each_object(p, s, page_address(page),
1424 						page->objects)
1425 			check_object(s, page, p, SLUB_RED_INACTIVE);
1426 	}
1427 
1428 	kmemcheck_free_shadow(page, compound_order(page));
1429 
1430 	mod_zone_page_state(page_zone(page),
1431 		(s->flags & SLAB_RECLAIM_ACCOUNT) ?
1432 		NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1433 		-pages);
1434 
1435 	__ClearPageSlabPfmemalloc(page);
1436 	__ClearPageSlab(page);
1437 
1438 	page_mapcount_reset(page);
1439 	if (current->reclaim_state)
1440 		current->reclaim_state->reclaimed_slab += pages;
1441 	__free_pages(page, order);
1442 	memcg_uncharge_slab(s, order);
1443 }
1444 
1445 #define need_reserve_slab_rcu						\
1446 	(sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1447 
1448 static void rcu_free_slab(struct rcu_head *h)
1449 {
1450 	struct page *page;
1451 
1452 	if (need_reserve_slab_rcu)
1453 		page = virt_to_head_page(h);
1454 	else
1455 		page = container_of((struct list_head *)h, struct page, lru);
1456 
1457 	__free_slab(page->slab_cache, page);
1458 }
1459 
1460 static void free_slab(struct kmem_cache *s, struct page *page)
1461 {
1462 	if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1463 		struct rcu_head *head;
1464 
1465 		if (need_reserve_slab_rcu) {
1466 			int order = compound_order(page);
1467 			int offset = (PAGE_SIZE << order) - s->reserved;
1468 
1469 			VM_BUG_ON(s->reserved != sizeof(*head));
1470 			head = page_address(page) + offset;
1471 		} else {
1472 			/*
1473 			 * RCU free overloads the RCU head over the LRU
1474 			 */
1475 			head = (void *)&page->lru;
1476 		}
1477 
1478 		call_rcu(head, rcu_free_slab);
1479 	} else
1480 		__free_slab(s, page);
1481 }
1482 
1483 static void discard_slab(struct kmem_cache *s, struct page *page)
1484 {
1485 	dec_slabs_node(s, page_to_nid(page), page->objects);
1486 	free_slab(s, page);
1487 }
1488 
1489 /*
1490  * Management of partially allocated slabs.
1491  */
1492 static inline void
1493 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1494 {
1495 	n->nr_partial++;
1496 	if (tail == DEACTIVATE_TO_TAIL)
1497 		list_add_tail(&page->lru, &n->partial);
1498 	else
1499 		list_add(&page->lru, &n->partial);
1500 }
1501 
1502 static inline void add_partial(struct kmem_cache_node *n,
1503 				struct page *page, int tail)
1504 {
1505 	lockdep_assert_held(&n->list_lock);
1506 	__add_partial(n, page, tail);
1507 }
1508 
1509 static inline void
1510 __remove_partial(struct kmem_cache_node *n, struct page *page)
1511 {
1512 	list_del(&page->lru);
1513 	n->nr_partial--;
1514 }
1515 
1516 static inline void remove_partial(struct kmem_cache_node *n,
1517 					struct page *page)
1518 {
1519 	lockdep_assert_held(&n->list_lock);
1520 	__remove_partial(n, page);
1521 }
1522 
1523 /*
1524  * Remove slab from the partial list, freeze it and
1525  * return the pointer to the freelist.
1526  *
1527  * Returns a list of objects or NULL if it fails.
1528  */
1529 static inline void *acquire_slab(struct kmem_cache *s,
1530 		struct kmem_cache_node *n, struct page *page,
1531 		int mode, int *objects)
1532 {
1533 	void *freelist;
1534 	unsigned long counters;
1535 	struct page new;
1536 
1537 	lockdep_assert_held(&n->list_lock);
1538 
1539 	/*
1540 	 * Zap the freelist and set the frozen bit.
1541 	 * The old freelist is the list of objects for the
1542 	 * per cpu allocation list.
1543 	 */
1544 	freelist = page->freelist;
1545 	counters = page->counters;
1546 	new.counters = counters;
1547 	*objects = new.objects - new.inuse;
1548 	if (mode) {
1549 		new.inuse = page->objects;
1550 		new.freelist = NULL;
1551 	} else {
1552 		new.freelist = freelist;
1553 	}
1554 
1555 	VM_BUG_ON(new.frozen);
1556 	new.frozen = 1;
1557 
1558 	if (!__cmpxchg_double_slab(s, page,
1559 			freelist, counters,
1560 			new.freelist, new.counters,
1561 			"acquire_slab"))
1562 		return NULL;
1563 
1564 	remove_partial(n, page);
1565 	WARN_ON(!freelist);
1566 	return freelist;
1567 }
1568 
1569 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1570 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1571 
1572 /*
1573  * Try to allocate a partial slab from a specific node.
1574  */
1575 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1576 				struct kmem_cache_cpu *c, gfp_t flags)
1577 {
1578 	struct page *page, *page2;
1579 	void *object = NULL;
1580 	int available = 0;
1581 	int objects;
1582 
1583 	/*
1584 	 * Racy check. If we mistakenly see no partial slabs then we
1585 	 * just allocate an empty slab. If we mistakenly try to get a
1586 	 * partial slab and there is none available then get_partials()
1587 	 * will return NULL.
1588 	 */
1589 	if (!n || !n->nr_partial)
1590 		return NULL;
1591 
1592 	spin_lock(&n->list_lock);
1593 	list_for_each_entry_safe(page, page2, &n->partial, lru) {
1594 		void *t;
1595 
1596 		if (!pfmemalloc_match(page, flags))
1597 			continue;
1598 
1599 		t = acquire_slab(s, n, page, object == NULL, &objects);
1600 		if (!t)
1601 			break;
1602 
1603 		available += objects;
1604 		if (!object) {
1605 			c->page = page;
1606 			stat(s, ALLOC_FROM_PARTIAL);
1607 			object = t;
1608 		} else {
1609 			put_cpu_partial(s, page, 0);
1610 			stat(s, CPU_PARTIAL_NODE);
1611 		}
1612 		if (!kmem_cache_has_cpu_partial(s)
1613 			|| available > s->cpu_partial / 2)
1614 			break;
1615 
1616 	}
1617 	spin_unlock(&n->list_lock);
1618 	return object;
1619 }
1620 
1621 /*
1622  * Get a page from somewhere. Search in increasing NUMA distances.
1623  */
1624 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1625 		struct kmem_cache_cpu *c)
1626 {
1627 #ifdef CONFIG_NUMA
1628 	struct zonelist *zonelist;
1629 	struct zoneref *z;
1630 	struct zone *zone;
1631 	enum zone_type high_zoneidx = gfp_zone(flags);
1632 	void *object;
1633 	unsigned int cpuset_mems_cookie;
1634 
1635 	/*
1636 	 * The defrag ratio allows a configuration of the tradeoffs between
1637 	 * inter node defragmentation and node local allocations. A lower
1638 	 * defrag_ratio increases the tendency to do local allocations
1639 	 * instead of attempting to obtain partial slabs from other nodes.
1640 	 *
1641 	 * If the defrag_ratio is set to 0 then kmalloc() always
1642 	 * returns node local objects. If the ratio is higher then kmalloc()
1643 	 * may return off node objects because partial slabs are obtained
1644 	 * from other nodes and filled up.
1645 	 *
1646 	 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1647 	 * defrag_ratio = 1000) then every (well almost) allocation will
1648 	 * first attempt to defrag slab caches on other nodes. This means
1649 	 * scanning over all nodes to look for partial slabs which may be
1650 	 * expensive if we do it every time we are trying to find a slab
1651 	 * with available objects.
1652 	 */
1653 	if (!s->remote_node_defrag_ratio ||
1654 			get_cycles() % 1024 > s->remote_node_defrag_ratio)
1655 		return NULL;
1656 
1657 	do {
1658 		cpuset_mems_cookie = read_mems_allowed_begin();
1659 		zonelist = node_zonelist(mempolicy_slab_node(), flags);
1660 		for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1661 			struct kmem_cache_node *n;
1662 
1663 			n = get_node(s, zone_to_nid(zone));
1664 
1665 			if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1666 					n->nr_partial > s->min_partial) {
1667 				object = get_partial_node(s, n, c, flags);
1668 				if (object) {
1669 					/*
1670 					 * Don't check read_mems_allowed_retry()
1671 					 * here - if mems_allowed was updated in
1672 					 * parallel, that was a harmless race
1673 					 * between allocation and the cpuset
1674 					 * update
1675 					 */
1676 					return object;
1677 				}
1678 			}
1679 		}
1680 	} while (read_mems_allowed_retry(cpuset_mems_cookie));
1681 #endif
1682 	return NULL;
1683 }
1684 
1685 /*
1686  * Get a partial page, lock it and return it.
1687  */
1688 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1689 		struct kmem_cache_cpu *c)
1690 {
1691 	void *object;
1692 	int searchnode = node;
1693 
1694 	if (node == NUMA_NO_NODE)
1695 		searchnode = numa_mem_id();
1696 	else if (!node_present_pages(node))
1697 		searchnode = node_to_mem_node(node);
1698 
1699 	object = get_partial_node(s, get_node(s, searchnode), c, flags);
1700 	if (object || node != NUMA_NO_NODE)
1701 		return object;
1702 
1703 	return get_any_partial(s, flags, c);
1704 }
1705 
1706 #ifdef CONFIG_PREEMPT
1707 /*
1708  * Calculate the next globally unique transaction for disambiguiation
1709  * during cmpxchg. The transactions start with the cpu number and are then
1710  * incremented by CONFIG_NR_CPUS.
1711  */
1712 #define TID_STEP  roundup_pow_of_two(CONFIG_NR_CPUS)
1713 #else
1714 /*
1715  * No preemption supported therefore also no need to check for
1716  * different cpus.
1717  */
1718 #define TID_STEP 1
1719 #endif
1720 
1721 static inline unsigned long next_tid(unsigned long tid)
1722 {
1723 	return tid + TID_STEP;
1724 }
1725 
1726 static inline unsigned int tid_to_cpu(unsigned long tid)
1727 {
1728 	return tid % TID_STEP;
1729 }
1730 
1731 static inline unsigned long tid_to_event(unsigned long tid)
1732 {
1733 	return tid / TID_STEP;
1734 }
1735 
1736 static inline unsigned int init_tid(int cpu)
1737 {
1738 	return cpu;
1739 }
1740 
1741 static inline void note_cmpxchg_failure(const char *n,
1742 		const struct kmem_cache *s, unsigned long tid)
1743 {
1744 #ifdef SLUB_DEBUG_CMPXCHG
1745 	unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1746 
1747 	pr_info("%s %s: cmpxchg redo ", n, s->name);
1748 
1749 #ifdef CONFIG_PREEMPT
1750 	if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1751 		pr_warn("due to cpu change %d -> %d\n",
1752 			tid_to_cpu(tid), tid_to_cpu(actual_tid));
1753 	else
1754 #endif
1755 	if (tid_to_event(tid) != tid_to_event(actual_tid))
1756 		pr_warn("due to cpu running other code. Event %ld->%ld\n",
1757 			tid_to_event(tid), tid_to_event(actual_tid));
1758 	else
1759 		pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1760 			actual_tid, tid, next_tid(tid));
1761 #endif
1762 	stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1763 }
1764 
1765 static void init_kmem_cache_cpus(struct kmem_cache *s)
1766 {
1767 	int cpu;
1768 
1769 	for_each_possible_cpu(cpu)
1770 		per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1771 }
1772 
1773 /*
1774  * Remove the cpu slab
1775  */
1776 static void deactivate_slab(struct kmem_cache *s, struct page *page,
1777 				void *freelist)
1778 {
1779 	enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1780 	struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1781 	int lock = 0;
1782 	enum slab_modes l = M_NONE, m = M_NONE;
1783 	void *nextfree;
1784 	int tail = DEACTIVATE_TO_HEAD;
1785 	struct page new;
1786 	struct page old;
1787 
1788 	if (page->freelist) {
1789 		stat(s, DEACTIVATE_REMOTE_FREES);
1790 		tail = DEACTIVATE_TO_TAIL;
1791 	}
1792 
1793 	/*
1794 	 * Stage one: Free all available per cpu objects back
1795 	 * to the page freelist while it is still frozen. Leave the
1796 	 * last one.
1797 	 *
1798 	 * There is no need to take the list->lock because the page
1799 	 * is still frozen.
1800 	 */
1801 	while (freelist && (nextfree = get_freepointer(s, freelist))) {
1802 		void *prior;
1803 		unsigned long counters;
1804 
1805 		do {
1806 			prior = page->freelist;
1807 			counters = page->counters;
1808 			set_freepointer(s, freelist, prior);
1809 			new.counters = counters;
1810 			new.inuse--;
1811 			VM_BUG_ON(!new.frozen);
1812 
1813 		} while (!__cmpxchg_double_slab(s, page,
1814 			prior, counters,
1815 			freelist, new.counters,
1816 			"drain percpu freelist"));
1817 
1818 		freelist = nextfree;
1819 	}
1820 
1821 	/*
1822 	 * Stage two: Ensure that the page is unfrozen while the
1823 	 * list presence reflects the actual number of objects
1824 	 * during unfreeze.
1825 	 *
1826 	 * We setup the list membership and then perform a cmpxchg
1827 	 * with the count. If there is a mismatch then the page
1828 	 * is not unfrozen but the page is on the wrong list.
1829 	 *
1830 	 * Then we restart the process which may have to remove
1831 	 * the page from the list that we just put it on again
1832 	 * because the number of objects in the slab may have
1833 	 * changed.
1834 	 */
1835 redo:
1836 
1837 	old.freelist = page->freelist;
1838 	old.counters = page->counters;
1839 	VM_BUG_ON(!old.frozen);
1840 
1841 	/* Determine target state of the slab */
1842 	new.counters = old.counters;
1843 	if (freelist) {
1844 		new.inuse--;
1845 		set_freepointer(s, freelist, old.freelist);
1846 		new.freelist = freelist;
1847 	} else
1848 		new.freelist = old.freelist;
1849 
1850 	new.frozen = 0;
1851 
1852 	if (!new.inuse && n->nr_partial >= s->min_partial)
1853 		m = M_FREE;
1854 	else if (new.freelist) {
1855 		m = M_PARTIAL;
1856 		if (!lock) {
1857 			lock = 1;
1858 			/*
1859 			 * Taking the spinlock removes the possiblity
1860 			 * that acquire_slab() will see a slab page that
1861 			 * is frozen
1862 			 */
1863 			spin_lock(&n->list_lock);
1864 		}
1865 	} else {
1866 		m = M_FULL;
1867 		if (kmem_cache_debug(s) && !lock) {
1868 			lock = 1;
1869 			/*
1870 			 * This also ensures that the scanning of full
1871 			 * slabs from diagnostic functions will not see
1872 			 * any frozen slabs.
1873 			 */
1874 			spin_lock(&n->list_lock);
1875 		}
1876 	}
1877 
1878 	if (l != m) {
1879 
1880 		if (l == M_PARTIAL)
1881 
1882 			remove_partial(n, page);
1883 
1884 		else if (l == M_FULL)
1885 
1886 			remove_full(s, n, page);
1887 
1888 		if (m == M_PARTIAL) {
1889 
1890 			add_partial(n, page, tail);
1891 			stat(s, tail);
1892 
1893 		} else if (m == M_FULL) {
1894 
1895 			stat(s, DEACTIVATE_FULL);
1896 			add_full(s, n, page);
1897 
1898 		}
1899 	}
1900 
1901 	l = m;
1902 	if (!__cmpxchg_double_slab(s, page,
1903 				old.freelist, old.counters,
1904 				new.freelist, new.counters,
1905 				"unfreezing slab"))
1906 		goto redo;
1907 
1908 	if (lock)
1909 		spin_unlock(&n->list_lock);
1910 
1911 	if (m == M_FREE) {
1912 		stat(s, DEACTIVATE_EMPTY);
1913 		discard_slab(s, page);
1914 		stat(s, FREE_SLAB);
1915 	}
1916 }
1917 
1918 /*
1919  * Unfreeze all the cpu partial slabs.
1920  *
1921  * This function must be called with interrupts disabled
1922  * for the cpu using c (or some other guarantee must be there
1923  * to guarantee no concurrent accesses).
1924  */
1925 static void unfreeze_partials(struct kmem_cache *s,
1926 		struct kmem_cache_cpu *c)
1927 {
1928 #ifdef CONFIG_SLUB_CPU_PARTIAL
1929 	struct kmem_cache_node *n = NULL, *n2 = NULL;
1930 	struct page *page, *discard_page = NULL;
1931 
1932 	while ((page = c->partial)) {
1933 		struct page new;
1934 		struct page old;
1935 
1936 		c->partial = page->next;
1937 
1938 		n2 = get_node(s, page_to_nid(page));
1939 		if (n != n2) {
1940 			if (n)
1941 				spin_unlock(&n->list_lock);
1942 
1943 			n = n2;
1944 			spin_lock(&n->list_lock);
1945 		}
1946 
1947 		do {
1948 
1949 			old.freelist = page->freelist;
1950 			old.counters = page->counters;
1951 			VM_BUG_ON(!old.frozen);
1952 
1953 			new.counters = old.counters;
1954 			new.freelist = old.freelist;
1955 
1956 			new.frozen = 0;
1957 
1958 		} while (!__cmpxchg_double_slab(s, page,
1959 				old.freelist, old.counters,
1960 				new.freelist, new.counters,
1961 				"unfreezing slab"));
1962 
1963 		if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
1964 			page->next = discard_page;
1965 			discard_page = page;
1966 		} else {
1967 			add_partial(n, page, DEACTIVATE_TO_TAIL);
1968 			stat(s, FREE_ADD_PARTIAL);
1969 		}
1970 	}
1971 
1972 	if (n)
1973 		spin_unlock(&n->list_lock);
1974 
1975 	while (discard_page) {
1976 		page = discard_page;
1977 		discard_page = discard_page->next;
1978 
1979 		stat(s, DEACTIVATE_EMPTY);
1980 		discard_slab(s, page);
1981 		stat(s, FREE_SLAB);
1982 	}
1983 #endif
1984 }
1985 
1986 /*
1987  * Put a page that was just frozen (in __slab_free) into a partial page
1988  * slot if available. This is done without interrupts disabled and without
1989  * preemption disabled. The cmpxchg is racy and may put the partial page
1990  * onto a random cpus partial slot.
1991  *
1992  * If we did not find a slot then simply move all the partials to the
1993  * per node partial list.
1994  */
1995 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
1996 {
1997 #ifdef CONFIG_SLUB_CPU_PARTIAL
1998 	struct page *oldpage;
1999 	int pages;
2000 	int pobjects;
2001 
2002 	do {
2003 		pages = 0;
2004 		pobjects = 0;
2005 		oldpage = this_cpu_read(s->cpu_slab->partial);
2006 
2007 		if (oldpage) {
2008 			pobjects = oldpage->pobjects;
2009 			pages = oldpage->pages;
2010 			if (drain && pobjects > s->cpu_partial) {
2011 				unsigned long flags;
2012 				/*
2013 				 * partial array is full. Move the existing
2014 				 * set to the per node partial list.
2015 				 */
2016 				local_irq_save(flags);
2017 				unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2018 				local_irq_restore(flags);
2019 				oldpage = NULL;
2020 				pobjects = 0;
2021 				pages = 0;
2022 				stat(s, CPU_PARTIAL_DRAIN);
2023 			}
2024 		}
2025 
2026 		pages++;
2027 		pobjects += page->objects - page->inuse;
2028 
2029 		page->pages = pages;
2030 		page->pobjects = pobjects;
2031 		page->next = oldpage;
2032 
2033 	} while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2034 								!= oldpage);
2035 #endif
2036 }
2037 
2038 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2039 {
2040 	stat(s, CPUSLAB_FLUSH);
2041 	deactivate_slab(s, c->page, c->freelist);
2042 
2043 	c->tid = next_tid(c->tid);
2044 	c->page = NULL;
2045 	c->freelist = NULL;
2046 }
2047 
2048 /*
2049  * Flush cpu slab.
2050  *
2051  * Called from IPI handler with interrupts disabled.
2052  */
2053 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2054 {
2055 	struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2056 
2057 	if (likely(c)) {
2058 		if (c->page)
2059 			flush_slab(s, c);
2060 
2061 		unfreeze_partials(s, c);
2062 	}
2063 }
2064 
2065 static void flush_cpu_slab(void *d)
2066 {
2067 	struct kmem_cache *s = d;
2068 
2069 	__flush_cpu_slab(s, smp_processor_id());
2070 }
2071 
2072 static bool has_cpu_slab(int cpu, void *info)
2073 {
2074 	struct kmem_cache *s = info;
2075 	struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2076 
2077 	return c->page || c->partial;
2078 }
2079 
2080 static void flush_all(struct kmem_cache *s)
2081 {
2082 	on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2083 }
2084 
2085 /*
2086  * Check if the objects in a per cpu structure fit numa
2087  * locality expectations.
2088  */
2089 static inline int node_match(struct page *page, int node)
2090 {
2091 #ifdef CONFIG_NUMA
2092 	if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2093 		return 0;
2094 #endif
2095 	return 1;
2096 }
2097 
2098 #ifdef CONFIG_SLUB_DEBUG
2099 static int count_free(struct page *page)
2100 {
2101 	return page->objects - page->inuse;
2102 }
2103 
2104 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2105 {
2106 	return atomic_long_read(&n->total_objects);
2107 }
2108 #endif /* CONFIG_SLUB_DEBUG */
2109 
2110 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2111 static unsigned long count_partial(struct kmem_cache_node *n,
2112 					int (*get_count)(struct page *))
2113 {
2114 	unsigned long flags;
2115 	unsigned long x = 0;
2116 	struct page *page;
2117 
2118 	spin_lock_irqsave(&n->list_lock, flags);
2119 	list_for_each_entry(page, &n->partial, lru)
2120 		x += get_count(page);
2121 	spin_unlock_irqrestore(&n->list_lock, flags);
2122 	return x;
2123 }
2124 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2125 
2126 static noinline void
2127 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2128 {
2129 #ifdef CONFIG_SLUB_DEBUG
2130 	static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2131 				      DEFAULT_RATELIMIT_BURST);
2132 	int node;
2133 	struct kmem_cache_node *n;
2134 
2135 	if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2136 		return;
2137 
2138 	pr_warn("SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2139 		nid, gfpflags);
2140 	pr_warn("  cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2141 		s->name, s->object_size, s->size, oo_order(s->oo),
2142 		oo_order(s->min));
2143 
2144 	if (oo_order(s->min) > get_order(s->object_size))
2145 		pr_warn("  %s debugging increased min order, use slub_debug=O to disable.\n",
2146 			s->name);
2147 
2148 	for_each_kmem_cache_node(s, node, n) {
2149 		unsigned long nr_slabs;
2150 		unsigned long nr_objs;
2151 		unsigned long nr_free;
2152 
2153 		nr_free  = count_partial(n, count_free);
2154 		nr_slabs = node_nr_slabs(n);
2155 		nr_objs  = node_nr_objs(n);
2156 
2157 		pr_warn("  node %d: slabs: %ld, objs: %ld, free: %ld\n",
2158 			node, nr_slabs, nr_objs, nr_free);
2159 	}
2160 #endif
2161 }
2162 
2163 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2164 			int node, struct kmem_cache_cpu **pc)
2165 {
2166 	void *freelist;
2167 	struct kmem_cache_cpu *c = *pc;
2168 	struct page *page;
2169 
2170 	freelist = get_partial(s, flags, node, c);
2171 
2172 	if (freelist)
2173 		return freelist;
2174 
2175 	page = new_slab(s, flags, node);
2176 	if (page) {
2177 		c = raw_cpu_ptr(s->cpu_slab);
2178 		if (c->page)
2179 			flush_slab(s, c);
2180 
2181 		/*
2182 		 * No other reference to the page yet so we can
2183 		 * muck around with it freely without cmpxchg
2184 		 */
2185 		freelist = page->freelist;
2186 		page->freelist = NULL;
2187 
2188 		stat(s, ALLOC_SLAB);
2189 		c->page = page;
2190 		*pc = c;
2191 	} else
2192 		freelist = NULL;
2193 
2194 	return freelist;
2195 }
2196 
2197 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2198 {
2199 	if (unlikely(PageSlabPfmemalloc(page)))
2200 		return gfp_pfmemalloc_allowed(gfpflags);
2201 
2202 	return true;
2203 }
2204 
2205 /*
2206  * Check the page->freelist of a page and either transfer the freelist to the
2207  * per cpu freelist or deactivate the page.
2208  *
2209  * The page is still frozen if the return value is not NULL.
2210  *
2211  * If this function returns NULL then the page has been unfrozen.
2212  *
2213  * This function must be called with interrupt disabled.
2214  */
2215 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2216 {
2217 	struct page new;
2218 	unsigned long counters;
2219 	void *freelist;
2220 
2221 	do {
2222 		freelist = page->freelist;
2223 		counters = page->counters;
2224 
2225 		new.counters = counters;
2226 		VM_BUG_ON(!new.frozen);
2227 
2228 		new.inuse = page->objects;
2229 		new.frozen = freelist != NULL;
2230 
2231 	} while (!__cmpxchg_double_slab(s, page,
2232 		freelist, counters,
2233 		NULL, new.counters,
2234 		"get_freelist"));
2235 
2236 	return freelist;
2237 }
2238 
2239 /*
2240  * Slow path. The lockless freelist is empty or we need to perform
2241  * debugging duties.
2242  *
2243  * Processing is still very fast if new objects have been freed to the
2244  * regular freelist. In that case we simply take over the regular freelist
2245  * as the lockless freelist and zap the regular freelist.
2246  *
2247  * If that is not working then we fall back to the partial lists. We take the
2248  * first element of the freelist as the object to allocate now and move the
2249  * rest of the freelist to the lockless freelist.
2250  *
2251  * And if we were unable to get a new slab from the partial slab lists then
2252  * we need to allocate a new slab. This is the slowest path since it involves
2253  * a call to the page allocator and the setup of a new slab.
2254  */
2255 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2256 			  unsigned long addr, struct kmem_cache_cpu *c)
2257 {
2258 	void *freelist;
2259 	struct page *page;
2260 	unsigned long flags;
2261 
2262 	local_irq_save(flags);
2263 #ifdef CONFIG_PREEMPT
2264 	/*
2265 	 * We may have been preempted and rescheduled on a different
2266 	 * cpu before disabling interrupts. Need to reload cpu area
2267 	 * pointer.
2268 	 */
2269 	c = this_cpu_ptr(s->cpu_slab);
2270 #endif
2271 
2272 	page = c->page;
2273 	if (!page)
2274 		goto new_slab;
2275 redo:
2276 
2277 	if (unlikely(!node_match(page, node))) {
2278 		int searchnode = node;
2279 
2280 		if (node != NUMA_NO_NODE && !node_present_pages(node))
2281 			searchnode = node_to_mem_node(node);
2282 
2283 		if (unlikely(!node_match(page, searchnode))) {
2284 			stat(s, ALLOC_NODE_MISMATCH);
2285 			deactivate_slab(s, page, c->freelist);
2286 			c->page = NULL;
2287 			c->freelist = NULL;
2288 			goto new_slab;
2289 		}
2290 	}
2291 
2292 	/*
2293 	 * By rights, we should be searching for a slab page that was
2294 	 * PFMEMALLOC but right now, we are losing the pfmemalloc
2295 	 * information when the page leaves the per-cpu allocator
2296 	 */
2297 	if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2298 		deactivate_slab(s, page, c->freelist);
2299 		c->page = NULL;
2300 		c->freelist = NULL;
2301 		goto new_slab;
2302 	}
2303 
2304 	/* must check again c->freelist in case of cpu migration or IRQ */
2305 	freelist = c->freelist;
2306 	if (freelist)
2307 		goto load_freelist;
2308 
2309 	freelist = get_freelist(s, page);
2310 
2311 	if (!freelist) {
2312 		c->page = NULL;
2313 		stat(s, DEACTIVATE_BYPASS);
2314 		goto new_slab;
2315 	}
2316 
2317 	stat(s, ALLOC_REFILL);
2318 
2319 load_freelist:
2320 	/*
2321 	 * freelist is pointing to the list of objects to be used.
2322 	 * page is pointing to the page from which the objects are obtained.
2323 	 * That page must be frozen for per cpu allocations to work.
2324 	 */
2325 	VM_BUG_ON(!c->page->frozen);
2326 	c->freelist = get_freepointer(s, freelist);
2327 	c->tid = next_tid(c->tid);
2328 	local_irq_restore(flags);
2329 	return freelist;
2330 
2331 new_slab:
2332 
2333 	if (c->partial) {
2334 		page = c->page = c->partial;
2335 		c->partial = page->next;
2336 		stat(s, CPU_PARTIAL_ALLOC);
2337 		c->freelist = NULL;
2338 		goto redo;
2339 	}
2340 
2341 	freelist = new_slab_objects(s, gfpflags, node, &c);
2342 
2343 	if (unlikely(!freelist)) {
2344 		slab_out_of_memory(s, gfpflags, node);
2345 		local_irq_restore(flags);
2346 		return NULL;
2347 	}
2348 
2349 	page = c->page;
2350 	if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2351 		goto load_freelist;
2352 
2353 	/* Only entered in the debug case */
2354 	if (kmem_cache_debug(s) &&
2355 			!alloc_debug_processing(s, page, freelist, addr))
2356 		goto new_slab;	/* Slab failed checks. Next slab needed */
2357 
2358 	deactivate_slab(s, page, get_freepointer(s, freelist));
2359 	c->page = NULL;
2360 	c->freelist = NULL;
2361 	local_irq_restore(flags);
2362 	return freelist;
2363 }
2364 
2365 /*
2366  * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2367  * have the fastpath folded into their functions. So no function call
2368  * overhead for requests that can be satisfied on the fastpath.
2369  *
2370  * The fastpath works by first checking if the lockless freelist can be used.
2371  * If not then __slab_alloc is called for slow processing.
2372  *
2373  * Otherwise we can simply pick the next object from the lockless free list.
2374  */
2375 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2376 		gfp_t gfpflags, int node, unsigned long addr)
2377 {
2378 	void **object;
2379 	struct kmem_cache_cpu *c;
2380 	struct page *page;
2381 	unsigned long tid;
2382 
2383 	if (slab_pre_alloc_hook(s, gfpflags))
2384 		return NULL;
2385 
2386 	s = memcg_kmem_get_cache(s, gfpflags);
2387 redo:
2388 	/*
2389 	 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2390 	 * enabled. We may switch back and forth between cpus while
2391 	 * reading from one cpu area. That does not matter as long
2392 	 * as we end up on the original cpu again when doing the cmpxchg.
2393 	 *
2394 	 * Preemption is disabled for the retrieval of the tid because that
2395 	 * must occur from the current processor. We cannot allow rescheduling
2396 	 * on a different processor between the determination of the pointer
2397 	 * and the retrieval of the tid.
2398 	 */
2399 	preempt_disable();
2400 	c = this_cpu_ptr(s->cpu_slab);
2401 
2402 	/*
2403 	 * The transaction ids are globally unique per cpu and per operation on
2404 	 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2405 	 * occurs on the right processor and that there was no operation on the
2406 	 * linked list in between.
2407 	 */
2408 	tid = c->tid;
2409 	preempt_enable();
2410 
2411 	object = c->freelist;
2412 	page = c->page;
2413 	if (unlikely(!object || !node_match(page, node))) {
2414 		object = __slab_alloc(s, gfpflags, node, addr, c);
2415 		stat(s, ALLOC_SLOWPATH);
2416 	} else {
2417 		void *next_object = get_freepointer_safe(s, object);
2418 
2419 		/*
2420 		 * The cmpxchg will only match if there was no additional
2421 		 * operation and if we are on the right processor.
2422 		 *
2423 		 * The cmpxchg does the following atomically (without lock
2424 		 * semantics!)
2425 		 * 1. Relocate first pointer to the current per cpu area.
2426 		 * 2. Verify that tid and freelist have not been changed
2427 		 * 3. If they were not changed replace tid and freelist
2428 		 *
2429 		 * Since this is without lock semantics the protection is only
2430 		 * against code executing on this cpu *not* from access by
2431 		 * other cpus.
2432 		 */
2433 		if (unlikely(!this_cpu_cmpxchg_double(
2434 				s->cpu_slab->freelist, s->cpu_slab->tid,
2435 				object, tid,
2436 				next_object, next_tid(tid)))) {
2437 
2438 			note_cmpxchg_failure("slab_alloc", s, tid);
2439 			goto redo;
2440 		}
2441 		prefetch_freepointer(s, next_object);
2442 		stat(s, ALLOC_FASTPATH);
2443 	}
2444 
2445 	if (unlikely(gfpflags & __GFP_ZERO) && object)
2446 		memset(object, 0, s->object_size);
2447 
2448 	slab_post_alloc_hook(s, gfpflags, object);
2449 
2450 	return object;
2451 }
2452 
2453 static __always_inline void *slab_alloc(struct kmem_cache *s,
2454 		gfp_t gfpflags, unsigned long addr)
2455 {
2456 	return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2457 }
2458 
2459 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2460 {
2461 	void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2462 
2463 	trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2464 				s->size, gfpflags);
2465 
2466 	return ret;
2467 }
2468 EXPORT_SYMBOL(kmem_cache_alloc);
2469 
2470 #ifdef CONFIG_TRACING
2471 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2472 {
2473 	void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2474 	trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2475 	return ret;
2476 }
2477 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2478 #endif
2479 
2480 #ifdef CONFIG_NUMA
2481 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2482 {
2483 	void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2484 
2485 	trace_kmem_cache_alloc_node(_RET_IP_, ret,
2486 				    s->object_size, s->size, gfpflags, node);
2487 
2488 	return ret;
2489 }
2490 EXPORT_SYMBOL(kmem_cache_alloc_node);
2491 
2492 #ifdef CONFIG_TRACING
2493 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2494 				    gfp_t gfpflags,
2495 				    int node, size_t size)
2496 {
2497 	void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2498 
2499 	trace_kmalloc_node(_RET_IP_, ret,
2500 			   size, s->size, gfpflags, node);
2501 	return ret;
2502 }
2503 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2504 #endif
2505 #endif
2506 
2507 /*
2508  * Slow patch handling. This may still be called frequently since objects
2509  * have a longer lifetime than the cpu slabs in most processing loads.
2510  *
2511  * So we still attempt to reduce cache line usage. Just take the slab
2512  * lock and free the item. If there is no additional partial page
2513  * handling required then we can return immediately.
2514  */
2515 static void __slab_free(struct kmem_cache *s, struct page *page,
2516 			void *x, unsigned long addr)
2517 {
2518 	void *prior;
2519 	void **object = (void *)x;
2520 	int was_frozen;
2521 	struct page new;
2522 	unsigned long counters;
2523 	struct kmem_cache_node *n = NULL;
2524 	unsigned long uninitialized_var(flags);
2525 
2526 	stat(s, FREE_SLOWPATH);
2527 
2528 	if (kmem_cache_debug(s) &&
2529 		!(n = free_debug_processing(s, page, x, addr, &flags)))
2530 		return;
2531 
2532 	do {
2533 		if (unlikely(n)) {
2534 			spin_unlock_irqrestore(&n->list_lock, flags);
2535 			n = NULL;
2536 		}
2537 		prior = page->freelist;
2538 		counters = page->counters;
2539 		set_freepointer(s, object, prior);
2540 		new.counters = counters;
2541 		was_frozen = new.frozen;
2542 		new.inuse--;
2543 		if ((!new.inuse || !prior) && !was_frozen) {
2544 
2545 			if (kmem_cache_has_cpu_partial(s) && !prior) {
2546 
2547 				/*
2548 				 * Slab was on no list before and will be
2549 				 * partially empty
2550 				 * We can defer the list move and instead
2551 				 * freeze it.
2552 				 */
2553 				new.frozen = 1;
2554 
2555 			} else { /* Needs to be taken off a list */
2556 
2557 	                        n = get_node(s, page_to_nid(page));
2558 				/*
2559 				 * Speculatively acquire the list_lock.
2560 				 * If the cmpxchg does not succeed then we may
2561 				 * drop the list_lock without any processing.
2562 				 *
2563 				 * Otherwise the list_lock will synchronize with
2564 				 * other processors updating the list of slabs.
2565 				 */
2566 				spin_lock_irqsave(&n->list_lock, flags);
2567 
2568 			}
2569 		}
2570 
2571 	} while (!cmpxchg_double_slab(s, page,
2572 		prior, counters,
2573 		object, new.counters,
2574 		"__slab_free"));
2575 
2576 	if (likely(!n)) {
2577 
2578 		/*
2579 		 * If we just froze the page then put it onto the
2580 		 * per cpu partial list.
2581 		 */
2582 		if (new.frozen && !was_frozen) {
2583 			put_cpu_partial(s, page, 1);
2584 			stat(s, CPU_PARTIAL_FREE);
2585 		}
2586 		/*
2587 		 * The list lock was not taken therefore no list
2588 		 * activity can be necessary.
2589 		 */
2590                 if (was_frozen)
2591                         stat(s, FREE_FROZEN);
2592                 return;
2593         }
2594 
2595 	if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2596 		goto slab_empty;
2597 
2598 	/*
2599 	 * Objects left in the slab. If it was not on the partial list before
2600 	 * then add it.
2601 	 */
2602 	if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2603 		if (kmem_cache_debug(s))
2604 			remove_full(s, n, page);
2605 		add_partial(n, page, DEACTIVATE_TO_TAIL);
2606 		stat(s, FREE_ADD_PARTIAL);
2607 	}
2608 	spin_unlock_irqrestore(&n->list_lock, flags);
2609 	return;
2610 
2611 slab_empty:
2612 	if (prior) {
2613 		/*
2614 		 * Slab on the partial list.
2615 		 */
2616 		remove_partial(n, page);
2617 		stat(s, FREE_REMOVE_PARTIAL);
2618 	} else {
2619 		/* Slab must be on the full list */
2620 		remove_full(s, n, page);
2621 	}
2622 
2623 	spin_unlock_irqrestore(&n->list_lock, flags);
2624 	stat(s, FREE_SLAB);
2625 	discard_slab(s, page);
2626 }
2627 
2628 /*
2629  * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2630  * can perform fastpath freeing without additional function calls.
2631  *
2632  * The fastpath is only possible if we are freeing to the current cpu slab
2633  * of this processor. This typically the case if we have just allocated
2634  * the item before.
2635  *
2636  * If fastpath is not possible then fall back to __slab_free where we deal
2637  * with all sorts of special processing.
2638  */
2639 static __always_inline void slab_free(struct kmem_cache *s,
2640 			struct page *page, void *x, unsigned long addr)
2641 {
2642 	void **object = (void *)x;
2643 	struct kmem_cache_cpu *c;
2644 	unsigned long tid;
2645 
2646 	slab_free_hook(s, x);
2647 
2648 redo:
2649 	/*
2650 	 * Determine the currently cpus per cpu slab.
2651 	 * The cpu may change afterward. However that does not matter since
2652 	 * data is retrieved via this pointer. If we are on the same cpu
2653 	 * during the cmpxchg then the free will succedd.
2654 	 */
2655 	preempt_disable();
2656 	c = this_cpu_ptr(s->cpu_slab);
2657 
2658 	tid = c->tid;
2659 	preempt_enable();
2660 
2661 	if (likely(page == c->page)) {
2662 		set_freepointer(s, object, c->freelist);
2663 
2664 		if (unlikely(!this_cpu_cmpxchg_double(
2665 				s->cpu_slab->freelist, s->cpu_slab->tid,
2666 				c->freelist, tid,
2667 				object, next_tid(tid)))) {
2668 
2669 			note_cmpxchg_failure("slab_free", s, tid);
2670 			goto redo;
2671 		}
2672 		stat(s, FREE_FASTPATH);
2673 	} else
2674 		__slab_free(s, page, x, addr);
2675 
2676 }
2677 
2678 void kmem_cache_free(struct kmem_cache *s, void *x)
2679 {
2680 	s = cache_from_obj(s, x);
2681 	if (!s)
2682 		return;
2683 	slab_free(s, virt_to_head_page(x), x, _RET_IP_);
2684 	trace_kmem_cache_free(_RET_IP_, x);
2685 }
2686 EXPORT_SYMBOL(kmem_cache_free);
2687 
2688 /*
2689  * Object placement in a slab is made very easy because we always start at
2690  * offset 0. If we tune the size of the object to the alignment then we can
2691  * get the required alignment by putting one properly sized object after
2692  * another.
2693  *
2694  * Notice that the allocation order determines the sizes of the per cpu
2695  * caches. Each processor has always one slab available for allocations.
2696  * Increasing the allocation order reduces the number of times that slabs
2697  * must be moved on and off the partial lists and is therefore a factor in
2698  * locking overhead.
2699  */
2700 
2701 /*
2702  * Mininum / Maximum order of slab pages. This influences locking overhead
2703  * and slab fragmentation. A higher order reduces the number of partial slabs
2704  * and increases the number of allocations possible without having to
2705  * take the list_lock.
2706  */
2707 static int slub_min_order;
2708 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2709 static int slub_min_objects;
2710 
2711 /*
2712  * Calculate the order of allocation given an slab object size.
2713  *
2714  * The order of allocation has significant impact on performance and other
2715  * system components. Generally order 0 allocations should be preferred since
2716  * order 0 does not cause fragmentation in the page allocator. Larger objects
2717  * be problematic to put into order 0 slabs because there may be too much
2718  * unused space left. We go to a higher order if more than 1/16th of the slab
2719  * would be wasted.
2720  *
2721  * In order to reach satisfactory performance we must ensure that a minimum
2722  * number of objects is in one slab. Otherwise we may generate too much
2723  * activity on the partial lists which requires taking the list_lock. This is
2724  * less a concern for large slabs though which are rarely used.
2725  *
2726  * slub_max_order specifies the order where we begin to stop considering the
2727  * number of objects in a slab as critical. If we reach slub_max_order then
2728  * we try to keep the page order as low as possible. So we accept more waste
2729  * of space in favor of a small page order.
2730  *
2731  * Higher order allocations also allow the placement of more objects in a
2732  * slab and thereby reduce object handling overhead. If the user has
2733  * requested a higher mininum order then we start with that one instead of
2734  * the smallest order which will fit the object.
2735  */
2736 static inline int slab_order(int size, int min_objects,
2737 				int max_order, int fract_leftover, int reserved)
2738 {
2739 	int order;
2740 	int rem;
2741 	int min_order = slub_min_order;
2742 
2743 	if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2744 		return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2745 
2746 	for (order = max(min_order,
2747 				fls(min_objects * size - 1) - PAGE_SHIFT);
2748 			order <= max_order; order++) {
2749 
2750 		unsigned long slab_size = PAGE_SIZE << order;
2751 
2752 		if (slab_size < min_objects * size + reserved)
2753 			continue;
2754 
2755 		rem = (slab_size - reserved) % size;
2756 
2757 		if (rem <= slab_size / fract_leftover)
2758 			break;
2759 
2760 	}
2761 
2762 	return order;
2763 }
2764 
2765 static inline int calculate_order(int size, int reserved)
2766 {
2767 	int order;
2768 	int min_objects;
2769 	int fraction;
2770 	int max_objects;
2771 
2772 	/*
2773 	 * Attempt to find best configuration for a slab. This
2774 	 * works by first attempting to generate a layout with
2775 	 * the best configuration and backing off gradually.
2776 	 *
2777 	 * First we reduce the acceptable waste in a slab. Then
2778 	 * we reduce the minimum objects required in a slab.
2779 	 */
2780 	min_objects = slub_min_objects;
2781 	if (!min_objects)
2782 		min_objects = 4 * (fls(nr_cpu_ids) + 1);
2783 	max_objects = order_objects(slub_max_order, size, reserved);
2784 	min_objects = min(min_objects, max_objects);
2785 
2786 	while (min_objects > 1) {
2787 		fraction = 16;
2788 		while (fraction >= 4) {
2789 			order = slab_order(size, min_objects,
2790 					slub_max_order, fraction, reserved);
2791 			if (order <= slub_max_order)
2792 				return order;
2793 			fraction /= 2;
2794 		}
2795 		min_objects--;
2796 	}
2797 
2798 	/*
2799 	 * We were unable to place multiple objects in a slab. Now
2800 	 * lets see if we can place a single object there.
2801 	 */
2802 	order = slab_order(size, 1, slub_max_order, 1, reserved);
2803 	if (order <= slub_max_order)
2804 		return order;
2805 
2806 	/*
2807 	 * Doh this slab cannot be placed using slub_max_order.
2808 	 */
2809 	order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2810 	if (order < MAX_ORDER)
2811 		return order;
2812 	return -ENOSYS;
2813 }
2814 
2815 static void
2816 init_kmem_cache_node(struct kmem_cache_node *n)
2817 {
2818 	n->nr_partial = 0;
2819 	spin_lock_init(&n->list_lock);
2820 	INIT_LIST_HEAD(&n->partial);
2821 #ifdef CONFIG_SLUB_DEBUG
2822 	atomic_long_set(&n->nr_slabs, 0);
2823 	atomic_long_set(&n->total_objects, 0);
2824 	INIT_LIST_HEAD(&n->full);
2825 #endif
2826 }
2827 
2828 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2829 {
2830 	BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2831 			KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
2832 
2833 	/*
2834 	 * Must align to double word boundary for the double cmpxchg
2835 	 * instructions to work; see __pcpu_double_call_return_bool().
2836 	 */
2837 	s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2838 				     2 * sizeof(void *));
2839 
2840 	if (!s->cpu_slab)
2841 		return 0;
2842 
2843 	init_kmem_cache_cpus(s);
2844 
2845 	return 1;
2846 }
2847 
2848 static struct kmem_cache *kmem_cache_node;
2849 
2850 /*
2851  * No kmalloc_node yet so do it by hand. We know that this is the first
2852  * slab on the node for this slabcache. There are no concurrent accesses
2853  * possible.
2854  *
2855  * Note that this function only works on the kmem_cache_node
2856  * when allocating for the kmem_cache_node. This is used for bootstrapping
2857  * memory on a fresh node that has no slab structures yet.
2858  */
2859 static void early_kmem_cache_node_alloc(int node)
2860 {
2861 	struct page *page;
2862 	struct kmem_cache_node *n;
2863 
2864 	BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2865 
2866 	page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2867 
2868 	BUG_ON(!page);
2869 	if (page_to_nid(page) != node) {
2870 		pr_err("SLUB: Unable to allocate memory from node %d\n", node);
2871 		pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
2872 	}
2873 
2874 	n = page->freelist;
2875 	BUG_ON(!n);
2876 	page->freelist = get_freepointer(kmem_cache_node, n);
2877 	page->inuse = 1;
2878 	page->frozen = 0;
2879 	kmem_cache_node->node[node] = n;
2880 #ifdef CONFIG_SLUB_DEBUG
2881 	init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2882 	init_tracking(kmem_cache_node, n);
2883 #endif
2884 	init_kmem_cache_node(n);
2885 	inc_slabs_node(kmem_cache_node, node, page->objects);
2886 
2887 	/*
2888 	 * No locks need to be taken here as it has just been
2889 	 * initialized and there is no concurrent access.
2890 	 */
2891 	__add_partial(n, page, DEACTIVATE_TO_HEAD);
2892 }
2893 
2894 static void free_kmem_cache_nodes(struct kmem_cache *s)
2895 {
2896 	int node;
2897 	struct kmem_cache_node *n;
2898 
2899 	for_each_kmem_cache_node(s, node, n) {
2900 		kmem_cache_free(kmem_cache_node, n);
2901 		s->node[node] = NULL;
2902 	}
2903 }
2904 
2905 static int init_kmem_cache_nodes(struct kmem_cache *s)
2906 {
2907 	int node;
2908 
2909 	for_each_node_state(node, N_NORMAL_MEMORY) {
2910 		struct kmem_cache_node *n;
2911 
2912 		if (slab_state == DOWN) {
2913 			early_kmem_cache_node_alloc(node);
2914 			continue;
2915 		}
2916 		n = kmem_cache_alloc_node(kmem_cache_node,
2917 						GFP_KERNEL, node);
2918 
2919 		if (!n) {
2920 			free_kmem_cache_nodes(s);
2921 			return 0;
2922 		}
2923 
2924 		s->node[node] = n;
2925 		init_kmem_cache_node(n);
2926 	}
2927 	return 1;
2928 }
2929 
2930 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2931 {
2932 	if (min < MIN_PARTIAL)
2933 		min = MIN_PARTIAL;
2934 	else if (min > MAX_PARTIAL)
2935 		min = MAX_PARTIAL;
2936 	s->min_partial = min;
2937 }
2938 
2939 /*
2940  * calculate_sizes() determines the order and the distribution of data within
2941  * a slab object.
2942  */
2943 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2944 {
2945 	unsigned long flags = s->flags;
2946 	unsigned long size = s->object_size;
2947 	int order;
2948 
2949 	/*
2950 	 * Round up object size to the next word boundary. We can only
2951 	 * place the free pointer at word boundaries and this determines
2952 	 * the possible location of the free pointer.
2953 	 */
2954 	size = ALIGN(size, sizeof(void *));
2955 
2956 #ifdef CONFIG_SLUB_DEBUG
2957 	/*
2958 	 * Determine if we can poison the object itself. If the user of
2959 	 * the slab may touch the object after free or before allocation
2960 	 * then we should never poison the object itself.
2961 	 */
2962 	if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2963 			!s->ctor)
2964 		s->flags |= __OBJECT_POISON;
2965 	else
2966 		s->flags &= ~__OBJECT_POISON;
2967 
2968 
2969 	/*
2970 	 * If we are Redzoning then check if there is some space between the
2971 	 * end of the object and the free pointer. If not then add an
2972 	 * additional word to have some bytes to store Redzone information.
2973 	 */
2974 	if ((flags & SLAB_RED_ZONE) && size == s->object_size)
2975 		size += sizeof(void *);
2976 #endif
2977 
2978 	/*
2979 	 * With that we have determined the number of bytes in actual use
2980 	 * by the object. This is the potential offset to the free pointer.
2981 	 */
2982 	s->inuse = size;
2983 
2984 	if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2985 		s->ctor)) {
2986 		/*
2987 		 * Relocate free pointer after the object if it is not
2988 		 * permitted to overwrite the first word of the object on
2989 		 * kmem_cache_free.
2990 		 *
2991 		 * This is the case if we do RCU, have a constructor or
2992 		 * destructor or are poisoning the objects.
2993 		 */
2994 		s->offset = size;
2995 		size += sizeof(void *);
2996 	}
2997 
2998 #ifdef CONFIG_SLUB_DEBUG
2999 	if (flags & SLAB_STORE_USER)
3000 		/*
3001 		 * Need to store information about allocs and frees after
3002 		 * the object.
3003 		 */
3004 		size += 2 * sizeof(struct track);
3005 
3006 	if (flags & SLAB_RED_ZONE)
3007 		/*
3008 		 * Add some empty padding so that we can catch
3009 		 * overwrites from earlier objects rather than let
3010 		 * tracking information or the free pointer be
3011 		 * corrupted if a user writes before the start
3012 		 * of the object.
3013 		 */
3014 		size += sizeof(void *);
3015 #endif
3016 
3017 	/*
3018 	 * SLUB stores one object immediately after another beginning from
3019 	 * offset 0. In order to align the objects we have to simply size
3020 	 * each object to conform to the alignment.
3021 	 */
3022 	size = ALIGN(size, s->align);
3023 	s->size = size;
3024 	if (forced_order >= 0)
3025 		order = forced_order;
3026 	else
3027 		order = calculate_order(size, s->reserved);
3028 
3029 	if (order < 0)
3030 		return 0;
3031 
3032 	s->allocflags = 0;
3033 	if (order)
3034 		s->allocflags |= __GFP_COMP;
3035 
3036 	if (s->flags & SLAB_CACHE_DMA)
3037 		s->allocflags |= GFP_DMA;
3038 
3039 	if (s->flags & SLAB_RECLAIM_ACCOUNT)
3040 		s->allocflags |= __GFP_RECLAIMABLE;
3041 
3042 	/*
3043 	 * Determine the number of objects per slab
3044 	 */
3045 	s->oo = oo_make(order, size, s->reserved);
3046 	s->min = oo_make(get_order(size), size, s->reserved);
3047 	if (oo_objects(s->oo) > oo_objects(s->max))
3048 		s->max = s->oo;
3049 
3050 	return !!oo_objects(s->oo);
3051 }
3052 
3053 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3054 {
3055 	s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3056 	s->reserved = 0;
3057 
3058 	if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3059 		s->reserved = sizeof(struct rcu_head);
3060 
3061 	if (!calculate_sizes(s, -1))
3062 		goto error;
3063 	if (disable_higher_order_debug) {
3064 		/*
3065 		 * Disable debugging flags that store metadata if the min slab
3066 		 * order increased.
3067 		 */
3068 		if (get_order(s->size) > get_order(s->object_size)) {
3069 			s->flags &= ~DEBUG_METADATA_FLAGS;
3070 			s->offset = 0;
3071 			if (!calculate_sizes(s, -1))
3072 				goto error;
3073 		}
3074 	}
3075 
3076 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3077     defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3078 	if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3079 		/* Enable fast mode */
3080 		s->flags |= __CMPXCHG_DOUBLE;
3081 #endif
3082 
3083 	/*
3084 	 * The larger the object size is, the more pages we want on the partial
3085 	 * list to avoid pounding the page allocator excessively.
3086 	 */
3087 	set_min_partial(s, ilog2(s->size) / 2);
3088 
3089 	/*
3090 	 * cpu_partial determined the maximum number of objects kept in the
3091 	 * per cpu partial lists of a processor.
3092 	 *
3093 	 * Per cpu partial lists mainly contain slabs that just have one
3094 	 * object freed. If they are used for allocation then they can be
3095 	 * filled up again with minimal effort. The slab will never hit the
3096 	 * per node partial lists and therefore no locking will be required.
3097 	 *
3098 	 * This setting also determines
3099 	 *
3100 	 * A) The number of objects from per cpu partial slabs dumped to the
3101 	 *    per node list when we reach the limit.
3102 	 * B) The number of objects in cpu partial slabs to extract from the
3103 	 *    per node list when we run out of per cpu objects. We only fetch
3104 	 *    50% to keep some capacity around for frees.
3105 	 */
3106 	if (!kmem_cache_has_cpu_partial(s))
3107 		s->cpu_partial = 0;
3108 	else if (s->size >= PAGE_SIZE)
3109 		s->cpu_partial = 2;
3110 	else if (s->size >= 1024)
3111 		s->cpu_partial = 6;
3112 	else if (s->size >= 256)
3113 		s->cpu_partial = 13;
3114 	else
3115 		s->cpu_partial = 30;
3116 
3117 #ifdef CONFIG_NUMA
3118 	s->remote_node_defrag_ratio = 1000;
3119 #endif
3120 	if (!init_kmem_cache_nodes(s))
3121 		goto error;
3122 
3123 	if (alloc_kmem_cache_cpus(s))
3124 		return 0;
3125 
3126 	free_kmem_cache_nodes(s);
3127 error:
3128 	if (flags & SLAB_PANIC)
3129 		panic("Cannot create slab %s size=%lu realsize=%u "
3130 			"order=%u offset=%u flags=%lx\n",
3131 			s->name, (unsigned long)s->size, s->size,
3132 			oo_order(s->oo), s->offset, flags);
3133 	return -EINVAL;
3134 }
3135 
3136 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3137 							const char *text)
3138 {
3139 #ifdef CONFIG_SLUB_DEBUG
3140 	void *addr = page_address(page);
3141 	void *p;
3142 	unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3143 				     sizeof(long), GFP_ATOMIC);
3144 	if (!map)
3145 		return;
3146 	slab_err(s, page, text, s->name);
3147 	slab_lock(page);
3148 
3149 	get_map(s, page, map);
3150 	for_each_object(p, s, addr, page->objects) {
3151 
3152 		if (!test_bit(slab_index(p, s, addr), map)) {
3153 			pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3154 			print_tracking(s, p);
3155 		}
3156 	}
3157 	slab_unlock(page);
3158 	kfree(map);
3159 #endif
3160 }
3161 
3162 /*
3163  * Attempt to free all partial slabs on a node.
3164  * This is called from kmem_cache_close(). We must be the last thread
3165  * using the cache and therefore we do not need to lock anymore.
3166  */
3167 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3168 {
3169 	struct page *page, *h;
3170 
3171 	list_for_each_entry_safe(page, h, &n->partial, lru) {
3172 		if (!page->inuse) {
3173 			__remove_partial(n, page);
3174 			discard_slab(s, page);
3175 		} else {
3176 			list_slab_objects(s, page,
3177 			"Objects remaining in %s on kmem_cache_close()");
3178 		}
3179 	}
3180 }
3181 
3182 /*
3183  * Release all resources used by a slab cache.
3184  */
3185 static inline int kmem_cache_close(struct kmem_cache *s)
3186 {
3187 	int node;
3188 	struct kmem_cache_node *n;
3189 
3190 	flush_all(s);
3191 	/* Attempt to free all objects */
3192 	for_each_kmem_cache_node(s, node, n) {
3193 		free_partial(s, n);
3194 		if (n->nr_partial || slabs_node(s, node))
3195 			return 1;
3196 	}
3197 	free_percpu(s->cpu_slab);
3198 	free_kmem_cache_nodes(s);
3199 	return 0;
3200 }
3201 
3202 int __kmem_cache_shutdown(struct kmem_cache *s)
3203 {
3204 	return kmem_cache_close(s);
3205 }
3206 
3207 /********************************************************************
3208  *		Kmalloc subsystem
3209  *******************************************************************/
3210 
3211 static int __init setup_slub_min_order(char *str)
3212 {
3213 	get_option(&str, &slub_min_order);
3214 
3215 	return 1;
3216 }
3217 
3218 __setup("slub_min_order=", setup_slub_min_order);
3219 
3220 static int __init setup_slub_max_order(char *str)
3221 {
3222 	get_option(&str, &slub_max_order);
3223 	slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3224 
3225 	return 1;
3226 }
3227 
3228 __setup("slub_max_order=", setup_slub_max_order);
3229 
3230 static int __init setup_slub_min_objects(char *str)
3231 {
3232 	get_option(&str, &slub_min_objects);
3233 
3234 	return 1;
3235 }
3236 
3237 __setup("slub_min_objects=", setup_slub_min_objects);
3238 
3239 void *__kmalloc(size_t size, gfp_t flags)
3240 {
3241 	struct kmem_cache *s;
3242 	void *ret;
3243 
3244 	if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3245 		return kmalloc_large(size, flags);
3246 
3247 	s = kmalloc_slab(size, flags);
3248 
3249 	if (unlikely(ZERO_OR_NULL_PTR(s)))
3250 		return s;
3251 
3252 	ret = slab_alloc(s, flags, _RET_IP_);
3253 
3254 	trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3255 
3256 	return ret;
3257 }
3258 EXPORT_SYMBOL(__kmalloc);
3259 
3260 #ifdef CONFIG_NUMA
3261 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3262 {
3263 	struct page *page;
3264 	void *ptr = NULL;
3265 
3266 	flags |= __GFP_COMP | __GFP_NOTRACK;
3267 	page = alloc_kmem_pages_node(node, flags, get_order(size));
3268 	if (page)
3269 		ptr = page_address(page);
3270 
3271 	kmalloc_large_node_hook(ptr, size, flags);
3272 	return ptr;
3273 }
3274 
3275 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3276 {
3277 	struct kmem_cache *s;
3278 	void *ret;
3279 
3280 	if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3281 		ret = kmalloc_large_node(size, flags, node);
3282 
3283 		trace_kmalloc_node(_RET_IP_, ret,
3284 				   size, PAGE_SIZE << get_order(size),
3285 				   flags, node);
3286 
3287 		return ret;
3288 	}
3289 
3290 	s = kmalloc_slab(size, flags);
3291 
3292 	if (unlikely(ZERO_OR_NULL_PTR(s)))
3293 		return s;
3294 
3295 	ret = slab_alloc_node(s, flags, node, _RET_IP_);
3296 
3297 	trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3298 
3299 	return ret;
3300 }
3301 EXPORT_SYMBOL(__kmalloc_node);
3302 #endif
3303 
3304 size_t ksize(const void *object)
3305 {
3306 	struct page *page;
3307 
3308 	if (unlikely(object == ZERO_SIZE_PTR))
3309 		return 0;
3310 
3311 	page = virt_to_head_page(object);
3312 
3313 	if (unlikely(!PageSlab(page))) {
3314 		WARN_ON(!PageCompound(page));
3315 		return PAGE_SIZE << compound_order(page);
3316 	}
3317 
3318 	return slab_ksize(page->slab_cache);
3319 }
3320 EXPORT_SYMBOL(ksize);
3321 
3322 void kfree(const void *x)
3323 {
3324 	struct page *page;
3325 	void *object = (void *)x;
3326 
3327 	trace_kfree(_RET_IP_, x);
3328 
3329 	if (unlikely(ZERO_OR_NULL_PTR(x)))
3330 		return;
3331 
3332 	page = virt_to_head_page(x);
3333 	if (unlikely(!PageSlab(page))) {
3334 		BUG_ON(!PageCompound(page));
3335 		kfree_hook(x);
3336 		__free_kmem_pages(page, compound_order(page));
3337 		return;
3338 	}
3339 	slab_free(page->slab_cache, page, object, _RET_IP_);
3340 }
3341 EXPORT_SYMBOL(kfree);
3342 
3343 /*
3344  * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3345  * the remaining slabs by the number of items in use. The slabs with the
3346  * most items in use come first. New allocations will then fill those up
3347  * and thus they can be removed from the partial lists.
3348  *
3349  * The slabs with the least items are placed last. This results in them
3350  * being allocated from last increasing the chance that the last objects
3351  * are freed in them.
3352  */
3353 int __kmem_cache_shrink(struct kmem_cache *s)
3354 {
3355 	int node;
3356 	int i;
3357 	struct kmem_cache_node *n;
3358 	struct page *page;
3359 	struct page *t;
3360 	int objects = oo_objects(s->max);
3361 	struct list_head *slabs_by_inuse =
3362 		kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3363 	unsigned long flags;
3364 
3365 	if (!slabs_by_inuse)
3366 		return -ENOMEM;
3367 
3368 	flush_all(s);
3369 	for_each_kmem_cache_node(s, node, n) {
3370 		if (!n->nr_partial)
3371 			continue;
3372 
3373 		for (i = 0; i < objects; i++)
3374 			INIT_LIST_HEAD(slabs_by_inuse + i);
3375 
3376 		spin_lock_irqsave(&n->list_lock, flags);
3377 
3378 		/*
3379 		 * Build lists indexed by the items in use in each slab.
3380 		 *
3381 		 * Note that concurrent frees may occur while we hold the
3382 		 * list_lock. page->inuse here is the upper limit.
3383 		 */
3384 		list_for_each_entry_safe(page, t, &n->partial, lru) {
3385 			list_move(&page->lru, slabs_by_inuse + page->inuse);
3386 			if (!page->inuse)
3387 				n->nr_partial--;
3388 		}
3389 
3390 		/*
3391 		 * Rebuild the partial list with the slabs filled up most
3392 		 * first and the least used slabs at the end.
3393 		 */
3394 		for (i = objects - 1; i > 0; i--)
3395 			list_splice(slabs_by_inuse + i, n->partial.prev);
3396 
3397 		spin_unlock_irqrestore(&n->list_lock, flags);
3398 
3399 		/* Release empty slabs */
3400 		list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3401 			discard_slab(s, page);
3402 	}
3403 
3404 	kfree(slabs_by_inuse);
3405 	return 0;
3406 }
3407 
3408 static int slab_mem_going_offline_callback(void *arg)
3409 {
3410 	struct kmem_cache *s;
3411 
3412 	mutex_lock(&slab_mutex);
3413 	list_for_each_entry(s, &slab_caches, list)
3414 		__kmem_cache_shrink(s);
3415 	mutex_unlock(&slab_mutex);
3416 
3417 	return 0;
3418 }
3419 
3420 static void slab_mem_offline_callback(void *arg)
3421 {
3422 	struct kmem_cache_node *n;
3423 	struct kmem_cache *s;
3424 	struct memory_notify *marg = arg;
3425 	int offline_node;
3426 
3427 	offline_node = marg->status_change_nid_normal;
3428 
3429 	/*
3430 	 * If the node still has available memory. we need kmem_cache_node
3431 	 * for it yet.
3432 	 */
3433 	if (offline_node < 0)
3434 		return;
3435 
3436 	mutex_lock(&slab_mutex);
3437 	list_for_each_entry(s, &slab_caches, list) {
3438 		n = get_node(s, offline_node);
3439 		if (n) {
3440 			/*
3441 			 * if n->nr_slabs > 0, slabs still exist on the node
3442 			 * that is going down. We were unable to free them,
3443 			 * and offline_pages() function shouldn't call this
3444 			 * callback. So, we must fail.
3445 			 */
3446 			BUG_ON(slabs_node(s, offline_node));
3447 
3448 			s->node[offline_node] = NULL;
3449 			kmem_cache_free(kmem_cache_node, n);
3450 		}
3451 	}
3452 	mutex_unlock(&slab_mutex);
3453 }
3454 
3455 static int slab_mem_going_online_callback(void *arg)
3456 {
3457 	struct kmem_cache_node *n;
3458 	struct kmem_cache *s;
3459 	struct memory_notify *marg = arg;
3460 	int nid = marg->status_change_nid_normal;
3461 	int ret = 0;
3462 
3463 	/*
3464 	 * If the node's memory is already available, then kmem_cache_node is
3465 	 * already created. Nothing to do.
3466 	 */
3467 	if (nid < 0)
3468 		return 0;
3469 
3470 	/*
3471 	 * We are bringing a node online. No memory is available yet. We must
3472 	 * allocate a kmem_cache_node structure in order to bring the node
3473 	 * online.
3474 	 */
3475 	mutex_lock(&slab_mutex);
3476 	list_for_each_entry(s, &slab_caches, list) {
3477 		/*
3478 		 * XXX: kmem_cache_alloc_node will fallback to other nodes
3479 		 *      since memory is not yet available from the node that
3480 		 *      is brought up.
3481 		 */
3482 		n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3483 		if (!n) {
3484 			ret = -ENOMEM;
3485 			goto out;
3486 		}
3487 		init_kmem_cache_node(n);
3488 		s->node[nid] = n;
3489 	}
3490 out:
3491 	mutex_unlock(&slab_mutex);
3492 	return ret;
3493 }
3494 
3495 static int slab_memory_callback(struct notifier_block *self,
3496 				unsigned long action, void *arg)
3497 {
3498 	int ret = 0;
3499 
3500 	switch (action) {
3501 	case MEM_GOING_ONLINE:
3502 		ret = slab_mem_going_online_callback(arg);
3503 		break;
3504 	case MEM_GOING_OFFLINE:
3505 		ret = slab_mem_going_offline_callback(arg);
3506 		break;
3507 	case MEM_OFFLINE:
3508 	case MEM_CANCEL_ONLINE:
3509 		slab_mem_offline_callback(arg);
3510 		break;
3511 	case MEM_ONLINE:
3512 	case MEM_CANCEL_OFFLINE:
3513 		break;
3514 	}
3515 	if (ret)
3516 		ret = notifier_from_errno(ret);
3517 	else
3518 		ret = NOTIFY_OK;
3519 	return ret;
3520 }
3521 
3522 static struct notifier_block slab_memory_callback_nb = {
3523 	.notifier_call = slab_memory_callback,
3524 	.priority = SLAB_CALLBACK_PRI,
3525 };
3526 
3527 /********************************************************************
3528  *			Basic setup of slabs
3529  *******************************************************************/
3530 
3531 /*
3532  * Used for early kmem_cache structures that were allocated using
3533  * the page allocator. Allocate them properly then fix up the pointers
3534  * that may be pointing to the wrong kmem_cache structure.
3535  */
3536 
3537 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3538 {
3539 	int node;
3540 	struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3541 	struct kmem_cache_node *n;
3542 
3543 	memcpy(s, static_cache, kmem_cache->object_size);
3544 
3545 	/*
3546 	 * This runs very early, and only the boot processor is supposed to be
3547 	 * up.  Even if it weren't true, IRQs are not up so we couldn't fire
3548 	 * IPIs around.
3549 	 */
3550 	__flush_cpu_slab(s, smp_processor_id());
3551 	for_each_kmem_cache_node(s, node, n) {
3552 		struct page *p;
3553 
3554 		list_for_each_entry(p, &n->partial, lru)
3555 			p->slab_cache = s;
3556 
3557 #ifdef CONFIG_SLUB_DEBUG
3558 		list_for_each_entry(p, &n->full, lru)
3559 			p->slab_cache = s;
3560 #endif
3561 	}
3562 	list_add(&s->list, &slab_caches);
3563 	return s;
3564 }
3565 
3566 void __init kmem_cache_init(void)
3567 {
3568 	static __initdata struct kmem_cache boot_kmem_cache,
3569 		boot_kmem_cache_node;
3570 
3571 	if (debug_guardpage_minorder())
3572 		slub_max_order = 0;
3573 
3574 	kmem_cache_node = &boot_kmem_cache_node;
3575 	kmem_cache = &boot_kmem_cache;
3576 
3577 	create_boot_cache(kmem_cache_node, "kmem_cache_node",
3578 		sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3579 
3580 	register_hotmemory_notifier(&slab_memory_callback_nb);
3581 
3582 	/* Able to allocate the per node structures */
3583 	slab_state = PARTIAL;
3584 
3585 	create_boot_cache(kmem_cache, "kmem_cache",
3586 			offsetof(struct kmem_cache, node) +
3587 				nr_node_ids * sizeof(struct kmem_cache_node *),
3588 		       SLAB_HWCACHE_ALIGN);
3589 
3590 	kmem_cache = bootstrap(&boot_kmem_cache);
3591 
3592 	/*
3593 	 * Allocate kmem_cache_node properly from the kmem_cache slab.
3594 	 * kmem_cache_node is separately allocated so no need to
3595 	 * update any list pointers.
3596 	 */
3597 	kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3598 
3599 	/* Now we can use the kmem_cache to allocate kmalloc slabs */
3600 	create_kmalloc_caches(0);
3601 
3602 #ifdef CONFIG_SMP
3603 	register_cpu_notifier(&slab_notifier);
3604 #endif
3605 
3606 	pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
3607 		cache_line_size(),
3608 		slub_min_order, slub_max_order, slub_min_objects,
3609 		nr_cpu_ids, nr_node_ids);
3610 }
3611 
3612 void __init kmem_cache_init_late(void)
3613 {
3614 }
3615 
3616 struct kmem_cache *
3617 __kmem_cache_alias(const char *name, size_t size, size_t align,
3618 		   unsigned long flags, void (*ctor)(void *))
3619 {
3620 	struct kmem_cache *s;
3621 
3622 	s = find_mergeable(size, align, flags, name, ctor);
3623 	if (s) {
3624 		int i;
3625 		struct kmem_cache *c;
3626 
3627 		s->refcount++;
3628 
3629 		/*
3630 		 * Adjust the object sizes so that we clear
3631 		 * the complete object on kzalloc.
3632 		 */
3633 		s->object_size = max(s->object_size, (int)size);
3634 		s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3635 
3636 		for_each_memcg_cache_index(i) {
3637 			c = cache_from_memcg_idx(s, i);
3638 			if (!c)
3639 				continue;
3640 			c->object_size = s->object_size;
3641 			c->inuse = max_t(int, c->inuse,
3642 					 ALIGN(size, sizeof(void *)));
3643 		}
3644 
3645 		if (sysfs_slab_alias(s, name)) {
3646 			s->refcount--;
3647 			s = NULL;
3648 		}
3649 	}
3650 
3651 	return s;
3652 }
3653 
3654 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3655 {
3656 	int err;
3657 
3658 	err = kmem_cache_open(s, flags);
3659 	if (err)
3660 		return err;
3661 
3662 	/* Mutex is not taken during early boot */
3663 	if (slab_state <= UP)
3664 		return 0;
3665 
3666 	memcg_propagate_slab_attrs(s);
3667 	err = sysfs_slab_add(s);
3668 	if (err)
3669 		kmem_cache_close(s);
3670 
3671 	return err;
3672 }
3673 
3674 #ifdef CONFIG_SMP
3675 /*
3676  * Use the cpu notifier to insure that the cpu slabs are flushed when
3677  * necessary.
3678  */
3679 static int slab_cpuup_callback(struct notifier_block *nfb,
3680 		unsigned long action, void *hcpu)
3681 {
3682 	long cpu = (long)hcpu;
3683 	struct kmem_cache *s;
3684 	unsigned long flags;
3685 
3686 	switch (action) {
3687 	case CPU_UP_CANCELED:
3688 	case CPU_UP_CANCELED_FROZEN:
3689 	case CPU_DEAD:
3690 	case CPU_DEAD_FROZEN:
3691 		mutex_lock(&slab_mutex);
3692 		list_for_each_entry(s, &slab_caches, list) {
3693 			local_irq_save(flags);
3694 			__flush_cpu_slab(s, cpu);
3695 			local_irq_restore(flags);
3696 		}
3697 		mutex_unlock(&slab_mutex);
3698 		break;
3699 	default:
3700 		break;
3701 	}
3702 	return NOTIFY_OK;
3703 }
3704 
3705 static struct notifier_block slab_notifier = {
3706 	.notifier_call = slab_cpuup_callback
3707 };
3708 
3709 #endif
3710 
3711 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3712 {
3713 	struct kmem_cache *s;
3714 	void *ret;
3715 
3716 	if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3717 		return kmalloc_large(size, gfpflags);
3718 
3719 	s = kmalloc_slab(size, gfpflags);
3720 
3721 	if (unlikely(ZERO_OR_NULL_PTR(s)))
3722 		return s;
3723 
3724 	ret = slab_alloc(s, gfpflags, caller);
3725 
3726 	/* Honor the call site pointer we received. */
3727 	trace_kmalloc(caller, ret, size, s->size, gfpflags);
3728 
3729 	return ret;
3730 }
3731 
3732 #ifdef CONFIG_NUMA
3733 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3734 					int node, unsigned long caller)
3735 {
3736 	struct kmem_cache *s;
3737 	void *ret;
3738 
3739 	if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3740 		ret = kmalloc_large_node(size, gfpflags, node);
3741 
3742 		trace_kmalloc_node(caller, ret,
3743 				   size, PAGE_SIZE << get_order(size),
3744 				   gfpflags, node);
3745 
3746 		return ret;
3747 	}
3748 
3749 	s = kmalloc_slab(size, gfpflags);
3750 
3751 	if (unlikely(ZERO_OR_NULL_PTR(s)))
3752 		return s;
3753 
3754 	ret = slab_alloc_node(s, gfpflags, node, caller);
3755 
3756 	/* Honor the call site pointer we received. */
3757 	trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3758 
3759 	return ret;
3760 }
3761 #endif
3762 
3763 #ifdef CONFIG_SYSFS
3764 static int count_inuse(struct page *page)
3765 {
3766 	return page->inuse;
3767 }
3768 
3769 static int count_total(struct page *page)
3770 {
3771 	return page->objects;
3772 }
3773 #endif
3774 
3775 #ifdef CONFIG_SLUB_DEBUG
3776 static int validate_slab(struct kmem_cache *s, struct page *page,
3777 						unsigned long *map)
3778 {
3779 	void *p;
3780 	void *addr = page_address(page);
3781 
3782 	if (!check_slab(s, page) ||
3783 			!on_freelist(s, page, NULL))
3784 		return 0;
3785 
3786 	/* Now we know that a valid freelist exists */
3787 	bitmap_zero(map, page->objects);
3788 
3789 	get_map(s, page, map);
3790 	for_each_object(p, s, addr, page->objects) {
3791 		if (test_bit(slab_index(p, s, addr), map))
3792 			if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3793 				return 0;
3794 	}
3795 
3796 	for_each_object(p, s, addr, page->objects)
3797 		if (!test_bit(slab_index(p, s, addr), map))
3798 			if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3799 				return 0;
3800 	return 1;
3801 }
3802 
3803 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3804 						unsigned long *map)
3805 {
3806 	slab_lock(page);
3807 	validate_slab(s, page, map);
3808 	slab_unlock(page);
3809 }
3810 
3811 static int validate_slab_node(struct kmem_cache *s,
3812 		struct kmem_cache_node *n, unsigned long *map)
3813 {
3814 	unsigned long count = 0;
3815 	struct page *page;
3816 	unsigned long flags;
3817 
3818 	spin_lock_irqsave(&n->list_lock, flags);
3819 
3820 	list_for_each_entry(page, &n->partial, lru) {
3821 		validate_slab_slab(s, page, map);
3822 		count++;
3823 	}
3824 	if (count != n->nr_partial)
3825 		pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
3826 		       s->name, count, n->nr_partial);
3827 
3828 	if (!(s->flags & SLAB_STORE_USER))
3829 		goto out;
3830 
3831 	list_for_each_entry(page, &n->full, lru) {
3832 		validate_slab_slab(s, page, map);
3833 		count++;
3834 	}
3835 	if (count != atomic_long_read(&n->nr_slabs))
3836 		pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
3837 		       s->name, count, atomic_long_read(&n->nr_slabs));
3838 
3839 out:
3840 	spin_unlock_irqrestore(&n->list_lock, flags);
3841 	return count;
3842 }
3843 
3844 static long validate_slab_cache(struct kmem_cache *s)
3845 {
3846 	int node;
3847 	unsigned long count = 0;
3848 	unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3849 				sizeof(unsigned long), GFP_KERNEL);
3850 	struct kmem_cache_node *n;
3851 
3852 	if (!map)
3853 		return -ENOMEM;
3854 
3855 	flush_all(s);
3856 	for_each_kmem_cache_node(s, node, n)
3857 		count += validate_slab_node(s, n, map);
3858 	kfree(map);
3859 	return count;
3860 }
3861 /*
3862  * Generate lists of code addresses where slabcache objects are allocated
3863  * and freed.
3864  */
3865 
3866 struct location {
3867 	unsigned long count;
3868 	unsigned long addr;
3869 	long long sum_time;
3870 	long min_time;
3871 	long max_time;
3872 	long min_pid;
3873 	long max_pid;
3874 	DECLARE_BITMAP(cpus, NR_CPUS);
3875 	nodemask_t nodes;
3876 };
3877 
3878 struct loc_track {
3879 	unsigned long max;
3880 	unsigned long count;
3881 	struct location *loc;
3882 };
3883 
3884 static void free_loc_track(struct loc_track *t)
3885 {
3886 	if (t->max)
3887 		free_pages((unsigned long)t->loc,
3888 			get_order(sizeof(struct location) * t->max));
3889 }
3890 
3891 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3892 {
3893 	struct location *l;
3894 	int order;
3895 
3896 	order = get_order(sizeof(struct location) * max);
3897 
3898 	l = (void *)__get_free_pages(flags, order);
3899 	if (!l)
3900 		return 0;
3901 
3902 	if (t->count) {
3903 		memcpy(l, t->loc, sizeof(struct location) * t->count);
3904 		free_loc_track(t);
3905 	}
3906 	t->max = max;
3907 	t->loc = l;
3908 	return 1;
3909 }
3910 
3911 static int add_location(struct loc_track *t, struct kmem_cache *s,
3912 				const struct track *track)
3913 {
3914 	long start, end, pos;
3915 	struct location *l;
3916 	unsigned long caddr;
3917 	unsigned long age = jiffies - track->when;
3918 
3919 	start = -1;
3920 	end = t->count;
3921 
3922 	for ( ; ; ) {
3923 		pos = start + (end - start + 1) / 2;
3924 
3925 		/*
3926 		 * There is nothing at "end". If we end up there
3927 		 * we need to add something to before end.
3928 		 */
3929 		if (pos == end)
3930 			break;
3931 
3932 		caddr = t->loc[pos].addr;
3933 		if (track->addr == caddr) {
3934 
3935 			l = &t->loc[pos];
3936 			l->count++;
3937 			if (track->when) {
3938 				l->sum_time += age;
3939 				if (age < l->min_time)
3940 					l->min_time = age;
3941 				if (age > l->max_time)
3942 					l->max_time = age;
3943 
3944 				if (track->pid < l->min_pid)
3945 					l->min_pid = track->pid;
3946 				if (track->pid > l->max_pid)
3947 					l->max_pid = track->pid;
3948 
3949 				cpumask_set_cpu(track->cpu,
3950 						to_cpumask(l->cpus));
3951 			}
3952 			node_set(page_to_nid(virt_to_page(track)), l->nodes);
3953 			return 1;
3954 		}
3955 
3956 		if (track->addr < caddr)
3957 			end = pos;
3958 		else
3959 			start = pos;
3960 	}
3961 
3962 	/*
3963 	 * Not found. Insert new tracking element.
3964 	 */
3965 	if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3966 		return 0;
3967 
3968 	l = t->loc + pos;
3969 	if (pos < t->count)
3970 		memmove(l + 1, l,
3971 			(t->count - pos) * sizeof(struct location));
3972 	t->count++;
3973 	l->count = 1;
3974 	l->addr = track->addr;
3975 	l->sum_time = age;
3976 	l->min_time = age;
3977 	l->max_time = age;
3978 	l->min_pid = track->pid;
3979 	l->max_pid = track->pid;
3980 	cpumask_clear(to_cpumask(l->cpus));
3981 	cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
3982 	nodes_clear(l->nodes);
3983 	node_set(page_to_nid(virt_to_page(track)), l->nodes);
3984 	return 1;
3985 }
3986 
3987 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3988 		struct page *page, enum track_item alloc,
3989 		unsigned long *map)
3990 {
3991 	void *addr = page_address(page);
3992 	void *p;
3993 
3994 	bitmap_zero(map, page->objects);
3995 	get_map(s, page, map);
3996 
3997 	for_each_object(p, s, addr, page->objects)
3998 		if (!test_bit(slab_index(p, s, addr), map))
3999 			add_location(t, s, get_track(s, p, alloc));
4000 }
4001 
4002 static int list_locations(struct kmem_cache *s, char *buf,
4003 					enum track_item alloc)
4004 {
4005 	int len = 0;
4006 	unsigned long i;
4007 	struct loc_track t = { 0, 0, NULL };
4008 	int node;
4009 	unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4010 				     sizeof(unsigned long), GFP_KERNEL);
4011 	struct kmem_cache_node *n;
4012 
4013 	if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4014 				     GFP_TEMPORARY)) {
4015 		kfree(map);
4016 		return sprintf(buf, "Out of memory\n");
4017 	}
4018 	/* Push back cpu slabs */
4019 	flush_all(s);
4020 
4021 	for_each_kmem_cache_node(s, node, n) {
4022 		unsigned long flags;
4023 		struct page *page;
4024 
4025 		if (!atomic_long_read(&n->nr_slabs))
4026 			continue;
4027 
4028 		spin_lock_irqsave(&n->list_lock, flags);
4029 		list_for_each_entry(page, &n->partial, lru)
4030 			process_slab(&t, s, page, alloc, map);
4031 		list_for_each_entry(page, &n->full, lru)
4032 			process_slab(&t, s, page, alloc, map);
4033 		spin_unlock_irqrestore(&n->list_lock, flags);
4034 	}
4035 
4036 	for (i = 0; i < t.count; i++) {
4037 		struct location *l = &t.loc[i];
4038 
4039 		if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4040 			break;
4041 		len += sprintf(buf + len, "%7ld ", l->count);
4042 
4043 		if (l->addr)
4044 			len += sprintf(buf + len, "%pS", (void *)l->addr);
4045 		else
4046 			len += sprintf(buf + len, "<not-available>");
4047 
4048 		if (l->sum_time != l->min_time) {
4049 			len += sprintf(buf + len, " age=%ld/%ld/%ld",
4050 				l->min_time,
4051 				(long)div_u64(l->sum_time, l->count),
4052 				l->max_time);
4053 		} else
4054 			len += sprintf(buf + len, " age=%ld",
4055 				l->min_time);
4056 
4057 		if (l->min_pid != l->max_pid)
4058 			len += sprintf(buf + len, " pid=%ld-%ld",
4059 				l->min_pid, l->max_pid);
4060 		else
4061 			len += sprintf(buf + len, " pid=%ld",
4062 				l->min_pid);
4063 
4064 		if (num_online_cpus() > 1 &&
4065 				!cpumask_empty(to_cpumask(l->cpus)) &&
4066 				len < PAGE_SIZE - 60) {
4067 			len += sprintf(buf + len, " cpus=");
4068 			len += cpulist_scnprintf(buf + len,
4069 						 PAGE_SIZE - len - 50,
4070 						 to_cpumask(l->cpus));
4071 		}
4072 
4073 		if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4074 				len < PAGE_SIZE - 60) {
4075 			len += sprintf(buf + len, " nodes=");
4076 			len += nodelist_scnprintf(buf + len,
4077 						  PAGE_SIZE - len - 50,
4078 						  l->nodes);
4079 		}
4080 
4081 		len += sprintf(buf + len, "\n");
4082 	}
4083 
4084 	free_loc_track(&t);
4085 	kfree(map);
4086 	if (!t.count)
4087 		len += sprintf(buf, "No data\n");
4088 	return len;
4089 }
4090 #endif
4091 
4092 #ifdef SLUB_RESILIENCY_TEST
4093 static void __init resiliency_test(void)
4094 {
4095 	u8 *p;
4096 
4097 	BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4098 
4099 	pr_err("SLUB resiliency testing\n");
4100 	pr_err("-----------------------\n");
4101 	pr_err("A. Corruption after allocation\n");
4102 
4103 	p = kzalloc(16, GFP_KERNEL);
4104 	p[16] = 0x12;
4105 	pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4106 	       p + 16);
4107 
4108 	validate_slab_cache(kmalloc_caches[4]);
4109 
4110 	/* Hmmm... The next two are dangerous */
4111 	p = kzalloc(32, GFP_KERNEL);
4112 	p[32 + sizeof(void *)] = 0x34;
4113 	pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4114 	       p);
4115 	pr_err("If allocated object is overwritten then not detectable\n\n");
4116 
4117 	validate_slab_cache(kmalloc_caches[5]);
4118 	p = kzalloc(64, GFP_KERNEL);
4119 	p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4120 	*p = 0x56;
4121 	pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4122 	       p);
4123 	pr_err("If allocated object is overwritten then not detectable\n\n");
4124 	validate_slab_cache(kmalloc_caches[6]);
4125 
4126 	pr_err("\nB. Corruption after free\n");
4127 	p = kzalloc(128, GFP_KERNEL);
4128 	kfree(p);
4129 	*p = 0x78;
4130 	pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4131 	validate_slab_cache(kmalloc_caches[7]);
4132 
4133 	p = kzalloc(256, GFP_KERNEL);
4134 	kfree(p);
4135 	p[50] = 0x9a;
4136 	pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4137 	validate_slab_cache(kmalloc_caches[8]);
4138 
4139 	p = kzalloc(512, GFP_KERNEL);
4140 	kfree(p);
4141 	p[512] = 0xab;
4142 	pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4143 	validate_slab_cache(kmalloc_caches[9]);
4144 }
4145 #else
4146 #ifdef CONFIG_SYSFS
4147 static void resiliency_test(void) {};
4148 #endif
4149 #endif
4150 
4151 #ifdef CONFIG_SYSFS
4152 enum slab_stat_type {
4153 	SL_ALL,			/* All slabs */
4154 	SL_PARTIAL,		/* Only partially allocated slabs */
4155 	SL_CPU,			/* Only slabs used for cpu caches */
4156 	SL_OBJECTS,		/* Determine allocated objects not slabs */
4157 	SL_TOTAL		/* Determine object capacity not slabs */
4158 };
4159 
4160 #define SO_ALL		(1 << SL_ALL)
4161 #define SO_PARTIAL	(1 << SL_PARTIAL)
4162 #define SO_CPU		(1 << SL_CPU)
4163 #define SO_OBJECTS	(1 << SL_OBJECTS)
4164 #define SO_TOTAL	(1 << SL_TOTAL)
4165 
4166 static ssize_t show_slab_objects(struct kmem_cache *s,
4167 			    char *buf, unsigned long flags)
4168 {
4169 	unsigned long total = 0;
4170 	int node;
4171 	int x;
4172 	unsigned long *nodes;
4173 
4174 	nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4175 	if (!nodes)
4176 		return -ENOMEM;
4177 
4178 	if (flags & SO_CPU) {
4179 		int cpu;
4180 
4181 		for_each_possible_cpu(cpu) {
4182 			struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4183 							       cpu);
4184 			int node;
4185 			struct page *page;
4186 
4187 			page = ACCESS_ONCE(c->page);
4188 			if (!page)
4189 				continue;
4190 
4191 			node = page_to_nid(page);
4192 			if (flags & SO_TOTAL)
4193 				x = page->objects;
4194 			else if (flags & SO_OBJECTS)
4195 				x = page->inuse;
4196 			else
4197 				x = 1;
4198 
4199 			total += x;
4200 			nodes[node] += x;
4201 
4202 			page = ACCESS_ONCE(c->partial);
4203 			if (page) {
4204 				node = page_to_nid(page);
4205 				if (flags & SO_TOTAL)
4206 					WARN_ON_ONCE(1);
4207 				else if (flags & SO_OBJECTS)
4208 					WARN_ON_ONCE(1);
4209 				else
4210 					x = page->pages;
4211 				total += x;
4212 				nodes[node] += x;
4213 			}
4214 		}
4215 	}
4216 
4217 	get_online_mems();
4218 #ifdef CONFIG_SLUB_DEBUG
4219 	if (flags & SO_ALL) {
4220 		struct kmem_cache_node *n;
4221 
4222 		for_each_kmem_cache_node(s, node, n) {
4223 
4224 			if (flags & SO_TOTAL)
4225 				x = atomic_long_read(&n->total_objects);
4226 			else if (flags & SO_OBJECTS)
4227 				x = atomic_long_read(&n->total_objects) -
4228 					count_partial(n, count_free);
4229 			else
4230 				x = atomic_long_read(&n->nr_slabs);
4231 			total += x;
4232 			nodes[node] += x;
4233 		}
4234 
4235 	} else
4236 #endif
4237 	if (flags & SO_PARTIAL) {
4238 		struct kmem_cache_node *n;
4239 
4240 		for_each_kmem_cache_node(s, node, n) {
4241 			if (flags & SO_TOTAL)
4242 				x = count_partial(n, count_total);
4243 			else if (flags & SO_OBJECTS)
4244 				x = count_partial(n, count_inuse);
4245 			else
4246 				x = n->nr_partial;
4247 			total += x;
4248 			nodes[node] += x;
4249 		}
4250 	}
4251 	x = sprintf(buf, "%lu", total);
4252 #ifdef CONFIG_NUMA
4253 	for (node = 0; node < nr_node_ids; node++)
4254 		if (nodes[node])
4255 			x += sprintf(buf + x, " N%d=%lu",
4256 					node, nodes[node]);
4257 #endif
4258 	put_online_mems();
4259 	kfree(nodes);
4260 	return x + sprintf(buf + x, "\n");
4261 }
4262 
4263 #ifdef CONFIG_SLUB_DEBUG
4264 static int any_slab_objects(struct kmem_cache *s)
4265 {
4266 	int node;
4267 	struct kmem_cache_node *n;
4268 
4269 	for_each_kmem_cache_node(s, node, n)
4270 		if (atomic_long_read(&n->total_objects))
4271 			return 1;
4272 
4273 	return 0;
4274 }
4275 #endif
4276 
4277 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4278 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4279 
4280 struct slab_attribute {
4281 	struct attribute attr;
4282 	ssize_t (*show)(struct kmem_cache *s, char *buf);
4283 	ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4284 };
4285 
4286 #define SLAB_ATTR_RO(_name) \
4287 	static struct slab_attribute _name##_attr = \
4288 	__ATTR(_name, 0400, _name##_show, NULL)
4289 
4290 #define SLAB_ATTR(_name) \
4291 	static struct slab_attribute _name##_attr =  \
4292 	__ATTR(_name, 0600, _name##_show, _name##_store)
4293 
4294 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4295 {
4296 	return sprintf(buf, "%d\n", s->size);
4297 }
4298 SLAB_ATTR_RO(slab_size);
4299 
4300 static ssize_t align_show(struct kmem_cache *s, char *buf)
4301 {
4302 	return sprintf(buf, "%d\n", s->align);
4303 }
4304 SLAB_ATTR_RO(align);
4305 
4306 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4307 {
4308 	return sprintf(buf, "%d\n", s->object_size);
4309 }
4310 SLAB_ATTR_RO(object_size);
4311 
4312 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4313 {
4314 	return sprintf(buf, "%d\n", oo_objects(s->oo));
4315 }
4316 SLAB_ATTR_RO(objs_per_slab);
4317 
4318 static ssize_t order_store(struct kmem_cache *s,
4319 				const char *buf, size_t length)
4320 {
4321 	unsigned long order;
4322 	int err;
4323 
4324 	err = kstrtoul(buf, 10, &order);
4325 	if (err)
4326 		return err;
4327 
4328 	if (order > slub_max_order || order < slub_min_order)
4329 		return -EINVAL;
4330 
4331 	calculate_sizes(s, order);
4332 	return length;
4333 }
4334 
4335 static ssize_t order_show(struct kmem_cache *s, char *buf)
4336 {
4337 	return sprintf(buf, "%d\n", oo_order(s->oo));
4338 }
4339 SLAB_ATTR(order);
4340 
4341 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4342 {
4343 	return sprintf(buf, "%lu\n", s->min_partial);
4344 }
4345 
4346 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4347 				 size_t length)
4348 {
4349 	unsigned long min;
4350 	int err;
4351 
4352 	err = kstrtoul(buf, 10, &min);
4353 	if (err)
4354 		return err;
4355 
4356 	set_min_partial(s, min);
4357 	return length;
4358 }
4359 SLAB_ATTR(min_partial);
4360 
4361 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4362 {
4363 	return sprintf(buf, "%u\n", s->cpu_partial);
4364 }
4365 
4366 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4367 				 size_t length)
4368 {
4369 	unsigned long objects;
4370 	int err;
4371 
4372 	err = kstrtoul(buf, 10, &objects);
4373 	if (err)
4374 		return err;
4375 	if (objects && !kmem_cache_has_cpu_partial(s))
4376 		return -EINVAL;
4377 
4378 	s->cpu_partial = objects;
4379 	flush_all(s);
4380 	return length;
4381 }
4382 SLAB_ATTR(cpu_partial);
4383 
4384 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4385 {
4386 	if (!s->ctor)
4387 		return 0;
4388 	return sprintf(buf, "%pS\n", s->ctor);
4389 }
4390 SLAB_ATTR_RO(ctor);
4391 
4392 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4393 {
4394 	return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
4395 }
4396 SLAB_ATTR_RO(aliases);
4397 
4398 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4399 {
4400 	return show_slab_objects(s, buf, SO_PARTIAL);
4401 }
4402 SLAB_ATTR_RO(partial);
4403 
4404 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4405 {
4406 	return show_slab_objects(s, buf, SO_CPU);
4407 }
4408 SLAB_ATTR_RO(cpu_slabs);
4409 
4410 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4411 {
4412 	return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4413 }
4414 SLAB_ATTR_RO(objects);
4415 
4416 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4417 {
4418 	return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4419 }
4420 SLAB_ATTR_RO(objects_partial);
4421 
4422 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4423 {
4424 	int objects = 0;
4425 	int pages = 0;
4426 	int cpu;
4427 	int len;
4428 
4429 	for_each_online_cpu(cpu) {
4430 		struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4431 
4432 		if (page) {
4433 			pages += page->pages;
4434 			objects += page->pobjects;
4435 		}
4436 	}
4437 
4438 	len = sprintf(buf, "%d(%d)", objects, pages);
4439 
4440 #ifdef CONFIG_SMP
4441 	for_each_online_cpu(cpu) {
4442 		struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4443 
4444 		if (page && len < PAGE_SIZE - 20)
4445 			len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4446 				page->pobjects, page->pages);
4447 	}
4448 #endif
4449 	return len + sprintf(buf + len, "\n");
4450 }
4451 SLAB_ATTR_RO(slabs_cpu_partial);
4452 
4453 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4454 {
4455 	return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4456 }
4457 
4458 static ssize_t reclaim_account_store(struct kmem_cache *s,
4459 				const char *buf, size_t length)
4460 {
4461 	s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4462 	if (buf[0] == '1')
4463 		s->flags |= SLAB_RECLAIM_ACCOUNT;
4464 	return length;
4465 }
4466 SLAB_ATTR(reclaim_account);
4467 
4468 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4469 {
4470 	return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4471 }
4472 SLAB_ATTR_RO(hwcache_align);
4473 
4474 #ifdef CONFIG_ZONE_DMA
4475 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4476 {
4477 	return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4478 }
4479 SLAB_ATTR_RO(cache_dma);
4480 #endif
4481 
4482 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4483 {
4484 	return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4485 }
4486 SLAB_ATTR_RO(destroy_by_rcu);
4487 
4488 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4489 {
4490 	return sprintf(buf, "%d\n", s->reserved);
4491 }
4492 SLAB_ATTR_RO(reserved);
4493 
4494 #ifdef CONFIG_SLUB_DEBUG
4495 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4496 {
4497 	return show_slab_objects(s, buf, SO_ALL);
4498 }
4499 SLAB_ATTR_RO(slabs);
4500 
4501 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4502 {
4503 	return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4504 }
4505 SLAB_ATTR_RO(total_objects);
4506 
4507 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4508 {
4509 	return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4510 }
4511 
4512 static ssize_t sanity_checks_store(struct kmem_cache *s,
4513 				const char *buf, size_t length)
4514 {
4515 	s->flags &= ~SLAB_DEBUG_FREE;
4516 	if (buf[0] == '1') {
4517 		s->flags &= ~__CMPXCHG_DOUBLE;
4518 		s->flags |= SLAB_DEBUG_FREE;
4519 	}
4520 	return length;
4521 }
4522 SLAB_ATTR(sanity_checks);
4523 
4524 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4525 {
4526 	return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4527 }
4528 
4529 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4530 							size_t length)
4531 {
4532 	/*
4533 	 * Tracing a merged cache is going to give confusing results
4534 	 * as well as cause other issues like converting a mergeable
4535 	 * cache into an umergeable one.
4536 	 */
4537 	if (s->refcount > 1)
4538 		return -EINVAL;
4539 
4540 	s->flags &= ~SLAB_TRACE;
4541 	if (buf[0] == '1') {
4542 		s->flags &= ~__CMPXCHG_DOUBLE;
4543 		s->flags |= SLAB_TRACE;
4544 	}
4545 	return length;
4546 }
4547 SLAB_ATTR(trace);
4548 
4549 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4550 {
4551 	return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4552 }
4553 
4554 static ssize_t red_zone_store(struct kmem_cache *s,
4555 				const char *buf, size_t length)
4556 {
4557 	if (any_slab_objects(s))
4558 		return -EBUSY;
4559 
4560 	s->flags &= ~SLAB_RED_ZONE;
4561 	if (buf[0] == '1') {
4562 		s->flags &= ~__CMPXCHG_DOUBLE;
4563 		s->flags |= SLAB_RED_ZONE;
4564 	}
4565 	calculate_sizes(s, -1);
4566 	return length;
4567 }
4568 SLAB_ATTR(red_zone);
4569 
4570 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4571 {
4572 	return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4573 }
4574 
4575 static ssize_t poison_store(struct kmem_cache *s,
4576 				const char *buf, size_t length)
4577 {
4578 	if (any_slab_objects(s))
4579 		return -EBUSY;
4580 
4581 	s->flags &= ~SLAB_POISON;
4582 	if (buf[0] == '1') {
4583 		s->flags &= ~__CMPXCHG_DOUBLE;
4584 		s->flags |= SLAB_POISON;
4585 	}
4586 	calculate_sizes(s, -1);
4587 	return length;
4588 }
4589 SLAB_ATTR(poison);
4590 
4591 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4592 {
4593 	return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4594 }
4595 
4596 static ssize_t store_user_store(struct kmem_cache *s,
4597 				const char *buf, size_t length)
4598 {
4599 	if (any_slab_objects(s))
4600 		return -EBUSY;
4601 
4602 	s->flags &= ~SLAB_STORE_USER;
4603 	if (buf[0] == '1') {
4604 		s->flags &= ~__CMPXCHG_DOUBLE;
4605 		s->flags |= SLAB_STORE_USER;
4606 	}
4607 	calculate_sizes(s, -1);
4608 	return length;
4609 }
4610 SLAB_ATTR(store_user);
4611 
4612 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4613 {
4614 	return 0;
4615 }
4616 
4617 static ssize_t validate_store(struct kmem_cache *s,
4618 			const char *buf, size_t length)
4619 {
4620 	int ret = -EINVAL;
4621 
4622 	if (buf[0] == '1') {
4623 		ret = validate_slab_cache(s);
4624 		if (ret >= 0)
4625 			ret = length;
4626 	}
4627 	return ret;
4628 }
4629 SLAB_ATTR(validate);
4630 
4631 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4632 {
4633 	if (!(s->flags & SLAB_STORE_USER))
4634 		return -ENOSYS;
4635 	return list_locations(s, buf, TRACK_ALLOC);
4636 }
4637 SLAB_ATTR_RO(alloc_calls);
4638 
4639 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4640 {
4641 	if (!(s->flags & SLAB_STORE_USER))
4642 		return -ENOSYS;
4643 	return list_locations(s, buf, TRACK_FREE);
4644 }
4645 SLAB_ATTR_RO(free_calls);
4646 #endif /* CONFIG_SLUB_DEBUG */
4647 
4648 #ifdef CONFIG_FAILSLAB
4649 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4650 {
4651 	return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4652 }
4653 
4654 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4655 							size_t length)
4656 {
4657 	if (s->refcount > 1)
4658 		return -EINVAL;
4659 
4660 	s->flags &= ~SLAB_FAILSLAB;
4661 	if (buf[0] == '1')
4662 		s->flags |= SLAB_FAILSLAB;
4663 	return length;
4664 }
4665 SLAB_ATTR(failslab);
4666 #endif
4667 
4668 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4669 {
4670 	return 0;
4671 }
4672 
4673 static ssize_t shrink_store(struct kmem_cache *s,
4674 			const char *buf, size_t length)
4675 {
4676 	if (buf[0] == '1') {
4677 		int rc = kmem_cache_shrink(s);
4678 
4679 		if (rc)
4680 			return rc;
4681 	} else
4682 		return -EINVAL;
4683 	return length;
4684 }
4685 SLAB_ATTR(shrink);
4686 
4687 #ifdef CONFIG_NUMA
4688 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4689 {
4690 	return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4691 }
4692 
4693 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4694 				const char *buf, size_t length)
4695 {
4696 	unsigned long ratio;
4697 	int err;
4698 
4699 	err = kstrtoul(buf, 10, &ratio);
4700 	if (err)
4701 		return err;
4702 
4703 	if (ratio <= 100)
4704 		s->remote_node_defrag_ratio = ratio * 10;
4705 
4706 	return length;
4707 }
4708 SLAB_ATTR(remote_node_defrag_ratio);
4709 #endif
4710 
4711 #ifdef CONFIG_SLUB_STATS
4712 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4713 {
4714 	unsigned long sum  = 0;
4715 	int cpu;
4716 	int len;
4717 	int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4718 
4719 	if (!data)
4720 		return -ENOMEM;
4721 
4722 	for_each_online_cpu(cpu) {
4723 		unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4724 
4725 		data[cpu] = x;
4726 		sum += x;
4727 	}
4728 
4729 	len = sprintf(buf, "%lu", sum);
4730 
4731 #ifdef CONFIG_SMP
4732 	for_each_online_cpu(cpu) {
4733 		if (data[cpu] && len < PAGE_SIZE - 20)
4734 			len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4735 	}
4736 #endif
4737 	kfree(data);
4738 	return len + sprintf(buf + len, "\n");
4739 }
4740 
4741 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4742 {
4743 	int cpu;
4744 
4745 	for_each_online_cpu(cpu)
4746 		per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4747 }
4748 
4749 #define STAT_ATTR(si, text) 					\
4750 static ssize_t text##_show(struct kmem_cache *s, char *buf)	\
4751 {								\
4752 	return show_stat(s, buf, si);				\
4753 }								\
4754 static ssize_t text##_store(struct kmem_cache *s,		\
4755 				const char *buf, size_t length)	\
4756 {								\
4757 	if (buf[0] != '0')					\
4758 		return -EINVAL;					\
4759 	clear_stat(s, si);					\
4760 	return length;						\
4761 }								\
4762 SLAB_ATTR(text);						\
4763 
4764 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4765 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4766 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4767 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4768 STAT_ATTR(FREE_FROZEN, free_frozen);
4769 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4770 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4771 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4772 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4773 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4774 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
4775 STAT_ATTR(FREE_SLAB, free_slab);
4776 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4777 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4778 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4779 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4780 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4781 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4782 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
4783 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4784 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4785 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4786 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
4787 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
4788 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
4789 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
4790 #endif
4791 
4792 static struct attribute *slab_attrs[] = {
4793 	&slab_size_attr.attr,
4794 	&object_size_attr.attr,
4795 	&objs_per_slab_attr.attr,
4796 	&order_attr.attr,
4797 	&min_partial_attr.attr,
4798 	&cpu_partial_attr.attr,
4799 	&objects_attr.attr,
4800 	&objects_partial_attr.attr,
4801 	&partial_attr.attr,
4802 	&cpu_slabs_attr.attr,
4803 	&ctor_attr.attr,
4804 	&aliases_attr.attr,
4805 	&align_attr.attr,
4806 	&hwcache_align_attr.attr,
4807 	&reclaim_account_attr.attr,
4808 	&destroy_by_rcu_attr.attr,
4809 	&shrink_attr.attr,
4810 	&reserved_attr.attr,
4811 	&slabs_cpu_partial_attr.attr,
4812 #ifdef CONFIG_SLUB_DEBUG
4813 	&total_objects_attr.attr,
4814 	&slabs_attr.attr,
4815 	&sanity_checks_attr.attr,
4816 	&trace_attr.attr,
4817 	&red_zone_attr.attr,
4818 	&poison_attr.attr,
4819 	&store_user_attr.attr,
4820 	&validate_attr.attr,
4821 	&alloc_calls_attr.attr,
4822 	&free_calls_attr.attr,
4823 #endif
4824 #ifdef CONFIG_ZONE_DMA
4825 	&cache_dma_attr.attr,
4826 #endif
4827 #ifdef CONFIG_NUMA
4828 	&remote_node_defrag_ratio_attr.attr,
4829 #endif
4830 #ifdef CONFIG_SLUB_STATS
4831 	&alloc_fastpath_attr.attr,
4832 	&alloc_slowpath_attr.attr,
4833 	&free_fastpath_attr.attr,
4834 	&free_slowpath_attr.attr,
4835 	&free_frozen_attr.attr,
4836 	&free_add_partial_attr.attr,
4837 	&free_remove_partial_attr.attr,
4838 	&alloc_from_partial_attr.attr,
4839 	&alloc_slab_attr.attr,
4840 	&alloc_refill_attr.attr,
4841 	&alloc_node_mismatch_attr.attr,
4842 	&free_slab_attr.attr,
4843 	&cpuslab_flush_attr.attr,
4844 	&deactivate_full_attr.attr,
4845 	&deactivate_empty_attr.attr,
4846 	&deactivate_to_head_attr.attr,
4847 	&deactivate_to_tail_attr.attr,
4848 	&deactivate_remote_frees_attr.attr,
4849 	&deactivate_bypass_attr.attr,
4850 	&order_fallback_attr.attr,
4851 	&cmpxchg_double_fail_attr.attr,
4852 	&cmpxchg_double_cpu_fail_attr.attr,
4853 	&cpu_partial_alloc_attr.attr,
4854 	&cpu_partial_free_attr.attr,
4855 	&cpu_partial_node_attr.attr,
4856 	&cpu_partial_drain_attr.attr,
4857 #endif
4858 #ifdef CONFIG_FAILSLAB
4859 	&failslab_attr.attr,
4860 #endif
4861 
4862 	NULL
4863 };
4864 
4865 static struct attribute_group slab_attr_group = {
4866 	.attrs = slab_attrs,
4867 };
4868 
4869 static ssize_t slab_attr_show(struct kobject *kobj,
4870 				struct attribute *attr,
4871 				char *buf)
4872 {
4873 	struct slab_attribute *attribute;
4874 	struct kmem_cache *s;
4875 	int err;
4876 
4877 	attribute = to_slab_attr(attr);
4878 	s = to_slab(kobj);
4879 
4880 	if (!attribute->show)
4881 		return -EIO;
4882 
4883 	err = attribute->show(s, buf);
4884 
4885 	return err;
4886 }
4887 
4888 static ssize_t slab_attr_store(struct kobject *kobj,
4889 				struct attribute *attr,
4890 				const char *buf, size_t len)
4891 {
4892 	struct slab_attribute *attribute;
4893 	struct kmem_cache *s;
4894 	int err;
4895 
4896 	attribute = to_slab_attr(attr);
4897 	s = to_slab(kobj);
4898 
4899 	if (!attribute->store)
4900 		return -EIO;
4901 
4902 	err = attribute->store(s, buf, len);
4903 #ifdef CONFIG_MEMCG_KMEM
4904 	if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
4905 		int i;
4906 
4907 		mutex_lock(&slab_mutex);
4908 		if (s->max_attr_size < len)
4909 			s->max_attr_size = len;
4910 
4911 		/*
4912 		 * This is a best effort propagation, so this function's return
4913 		 * value will be determined by the parent cache only. This is
4914 		 * basically because not all attributes will have a well
4915 		 * defined semantics for rollbacks - most of the actions will
4916 		 * have permanent effects.
4917 		 *
4918 		 * Returning the error value of any of the children that fail
4919 		 * is not 100 % defined, in the sense that users seeing the
4920 		 * error code won't be able to know anything about the state of
4921 		 * the cache.
4922 		 *
4923 		 * Only returning the error code for the parent cache at least
4924 		 * has well defined semantics. The cache being written to
4925 		 * directly either failed or succeeded, in which case we loop
4926 		 * through the descendants with best-effort propagation.
4927 		 */
4928 		for_each_memcg_cache_index(i) {
4929 			struct kmem_cache *c = cache_from_memcg_idx(s, i);
4930 			if (c)
4931 				attribute->store(c, buf, len);
4932 		}
4933 		mutex_unlock(&slab_mutex);
4934 	}
4935 #endif
4936 	return err;
4937 }
4938 
4939 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
4940 {
4941 #ifdef CONFIG_MEMCG_KMEM
4942 	int i;
4943 	char *buffer = NULL;
4944 	struct kmem_cache *root_cache;
4945 
4946 	if (is_root_cache(s))
4947 		return;
4948 
4949 	root_cache = s->memcg_params->root_cache;
4950 
4951 	/*
4952 	 * This mean this cache had no attribute written. Therefore, no point
4953 	 * in copying default values around
4954 	 */
4955 	if (!root_cache->max_attr_size)
4956 		return;
4957 
4958 	for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
4959 		char mbuf[64];
4960 		char *buf;
4961 		struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
4962 
4963 		if (!attr || !attr->store || !attr->show)
4964 			continue;
4965 
4966 		/*
4967 		 * It is really bad that we have to allocate here, so we will
4968 		 * do it only as a fallback. If we actually allocate, though,
4969 		 * we can just use the allocated buffer until the end.
4970 		 *
4971 		 * Most of the slub attributes will tend to be very small in
4972 		 * size, but sysfs allows buffers up to a page, so they can
4973 		 * theoretically happen.
4974 		 */
4975 		if (buffer)
4976 			buf = buffer;
4977 		else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
4978 			buf = mbuf;
4979 		else {
4980 			buffer = (char *) get_zeroed_page(GFP_KERNEL);
4981 			if (WARN_ON(!buffer))
4982 				continue;
4983 			buf = buffer;
4984 		}
4985 
4986 		attr->show(root_cache, buf);
4987 		attr->store(s, buf, strlen(buf));
4988 	}
4989 
4990 	if (buffer)
4991 		free_page((unsigned long)buffer);
4992 #endif
4993 }
4994 
4995 static void kmem_cache_release(struct kobject *k)
4996 {
4997 	slab_kmem_cache_release(to_slab(k));
4998 }
4999 
5000 static const struct sysfs_ops slab_sysfs_ops = {
5001 	.show = slab_attr_show,
5002 	.store = slab_attr_store,
5003 };
5004 
5005 static struct kobj_type slab_ktype = {
5006 	.sysfs_ops = &slab_sysfs_ops,
5007 	.release = kmem_cache_release,
5008 };
5009 
5010 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5011 {
5012 	struct kobj_type *ktype = get_ktype(kobj);
5013 
5014 	if (ktype == &slab_ktype)
5015 		return 1;
5016 	return 0;
5017 }
5018 
5019 static const struct kset_uevent_ops slab_uevent_ops = {
5020 	.filter = uevent_filter,
5021 };
5022 
5023 static struct kset *slab_kset;
5024 
5025 static inline struct kset *cache_kset(struct kmem_cache *s)
5026 {
5027 #ifdef CONFIG_MEMCG_KMEM
5028 	if (!is_root_cache(s))
5029 		return s->memcg_params->root_cache->memcg_kset;
5030 #endif
5031 	return slab_kset;
5032 }
5033 
5034 #define ID_STR_LENGTH 64
5035 
5036 /* Create a unique string id for a slab cache:
5037  *
5038  * Format	:[flags-]size
5039  */
5040 static char *create_unique_id(struct kmem_cache *s)
5041 {
5042 	char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5043 	char *p = name;
5044 
5045 	BUG_ON(!name);
5046 
5047 	*p++ = ':';
5048 	/*
5049 	 * First flags affecting slabcache operations. We will only
5050 	 * get here for aliasable slabs so we do not need to support
5051 	 * too many flags. The flags here must cover all flags that
5052 	 * are matched during merging to guarantee that the id is
5053 	 * unique.
5054 	 */
5055 	if (s->flags & SLAB_CACHE_DMA)
5056 		*p++ = 'd';
5057 	if (s->flags & SLAB_RECLAIM_ACCOUNT)
5058 		*p++ = 'a';
5059 	if (s->flags & SLAB_DEBUG_FREE)
5060 		*p++ = 'F';
5061 	if (!(s->flags & SLAB_NOTRACK))
5062 		*p++ = 't';
5063 	if (p != name + 1)
5064 		*p++ = '-';
5065 	p += sprintf(p, "%07d", s->size);
5066 
5067 	BUG_ON(p > name + ID_STR_LENGTH - 1);
5068 	return name;
5069 }
5070 
5071 static int sysfs_slab_add(struct kmem_cache *s)
5072 {
5073 	int err;
5074 	const char *name;
5075 	int unmergeable = slab_unmergeable(s);
5076 
5077 	if (unmergeable) {
5078 		/*
5079 		 * Slabcache can never be merged so we can use the name proper.
5080 		 * This is typically the case for debug situations. In that
5081 		 * case we can catch duplicate names easily.
5082 		 */
5083 		sysfs_remove_link(&slab_kset->kobj, s->name);
5084 		name = s->name;
5085 	} else {
5086 		/*
5087 		 * Create a unique name for the slab as a target
5088 		 * for the symlinks.
5089 		 */
5090 		name = create_unique_id(s);
5091 	}
5092 
5093 	s->kobj.kset = cache_kset(s);
5094 	err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5095 	if (err)
5096 		goto out_put_kobj;
5097 
5098 	err = sysfs_create_group(&s->kobj, &slab_attr_group);
5099 	if (err)
5100 		goto out_del_kobj;
5101 
5102 #ifdef CONFIG_MEMCG_KMEM
5103 	if (is_root_cache(s)) {
5104 		s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5105 		if (!s->memcg_kset) {
5106 			err = -ENOMEM;
5107 			goto out_del_kobj;
5108 		}
5109 	}
5110 #endif
5111 
5112 	kobject_uevent(&s->kobj, KOBJ_ADD);
5113 	if (!unmergeable) {
5114 		/* Setup first alias */
5115 		sysfs_slab_alias(s, s->name);
5116 	}
5117 out:
5118 	if (!unmergeable)
5119 		kfree(name);
5120 	return err;
5121 out_del_kobj:
5122 	kobject_del(&s->kobj);
5123 out_put_kobj:
5124 	kobject_put(&s->kobj);
5125 	goto out;
5126 }
5127 
5128 void sysfs_slab_remove(struct kmem_cache *s)
5129 {
5130 	if (slab_state < FULL)
5131 		/*
5132 		 * Sysfs has not been setup yet so no need to remove the
5133 		 * cache from sysfs.
5134 		 */
5135 		return;
5136 
5137 #ifdef CONFIG_MEMCG_KMEM
5138 	kset_unregister(s->memcg_kset);
5139 #endif
5140 	kobject_uevent(&s->kobj, KOBJ_REMOVE);
5141 	kobject_del(&s->kobj);
5142 	kobject_put(&s->kobj);
5143 }
5144 
5145 /*
5146  * Need to buffer aliases during bootup until sysfs becomes
5147  * available lest we lose that information.
5148  */
5149 struct saved_alias {
5150 	struct kmem_cache *s;
5151 	const char *name;
5152 	struct saved_alias *next;
5153 };
5154 
5155 static struct saved_alias *alias_list;
5156 
5157 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5158 {
5159 	struct saved_alias *al;
5160 
5161 	if (slab_state == FULL) {
5162 		/*
5163 		 * If we have a leftover link then remove it.
5164 		 */
5165 		sysfs_remove_link(&slab_kset->kobj, name);
5166 		return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5167 	}
5168 
5169 	al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5170 	if (!al)
5171 		return -ENOMEM;
5172 
5173 	al->s = s;
5174 	al->name = name;
5175 	al->next = alias_list;
5176 	alias_list = al;
5177 	return 0;
5178 }
5179 
5180 static int __init slab_sysfs_init(void)
5181 {
5182 	struct kmem_cache *s;
5183 	int err;
5184 
5185 	mutex_lock(&slab_mutex);
5186 
5187 	slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5188 	if (!slab_kset) {
5189 		mutex_unlock(&slab_mutex);
5190 		pr_err("Cannot register slab subsystem.\n");
5191 		return -ENOSYS;
5192 	}
5193 
5194 	slab_state = FULL;
5195 
5196 	list_for_each_entry(s, &slab_caches, list) {
5197 		err = sysfs_slab_add(s);
5198 		if (err)
5199 			pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5200 			       s->name);
5201 	}
5202 
5203 	while (alias_list) {
5204 		struct saved_alias *al = alias_list;
5205 
5206 		alias_list = alias_list->next;
5207 		err = sysfs_slab_alias(al->s, al->name);
5208 		if (err)
5209 			pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5210 			       al->name);
5211 		kfree(al);
5212 	}
5213 
5214 	mutex_unlock(&slab_mutex);
5215 	resiliency_test();
5216 	return 0;
5217 }
5218 
5219 __initcall(slab_sysfs_init);
5220 #endif /* CONFIG_SYSFS */
5221 
5222 /*
5223  * The /proc/slabinfo ABI
5224  */
5225 #ifdef CONFIG_SLABINFO
5226 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5227 {
5228 	unsigned long nr_slabs = 0;
5229 	unsigned long nr_objs = 0;
5230 	unsigned long nr_free = 0;
5231 	int node;
5232 	struct kmem_cache_node *n;
5233 
5234 	for_each_kmem_cache_node(s, node, n) {
5235 		nr_slabs += node_nr_slabs(n);
5236 		nr_objs += node_nr_objs(n);
5237 		nr_free += count_partial(n, count_free);
5238 	}
5239 
5240 	sinfo->active_objs = nr_objs - nr_free;
5241 	sinfo->num_objs = nr_objs;
5242 	sinfo->active_slabs = nr_slabs;
5243 	sinfo->num_slabs = nr_slabs;
5244 	sinfo->objects_per_slab = oo_objects(s->oo);
5245 	sinfo->cache_order = oo_order(s->oo);
5246 }
5247 
5248 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5249 {
5250 }
5251 
5252 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5253 		       size_t count, loff_t *ppos)
5254 {
5255 	return -EIO;
5256 }
5257 #endif /* CONFIG_SLABINFO */
5258