xref: /openbmc/linux/mm/slub.c (revision e8f6f3b4)
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 			page->objects, maxobj);
853 		return 0;
854 	}
855 	if (page->inuse > page->objects) {
856 		slab_err(s, page, "inuse %u > max %u",
857 			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 	int 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 struct kmem_cache *slab_pre_alloc_hook(struct kmem_cache *s,
1237 						     gfp_t flags)
1238 {
1239 	flags &= gfp_allowed_mask;
1240 	lockdep_trace_alloc(flags);
1241 	might_sleep_if(flags & __GFP_WAIT);
1242 
1243 	if (should_failslab(s->object_size, flags, s->flags))
1244 		return NULL;
1245 
1246 	return memcg_kmem_get_cache(s, flags);
1247 }
1248 
1249 static inline void slab_post_alloc_hook(struct kmem_cache *s,
1250 					gfp_t flags, void *object)
1251 {
1252 	flags &= gfp_allowed_mask;
1253 	kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
1254 	kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
1255 	memcg_kmem_put_cache(s);
1256 }
1257 
1258 static inline void slab_free_hook(struct kmem_cache *s, void *x)
1259 {
1260 	kmemleak_free_recursive(x, s->flags);
1261 
1262 	/*
1263 	 * Trouble is that we may no longer disable interrupts in the fast path
1264 	 * So in order to make the debug calls that expect irqs to be
1265 	 * disabled we need to disable interrupts temporarily.
1266 	 */
1267 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1268 	{
1269 		unsigned long flags;
1270 
1271 		local_irq_save(flags);
1272 		kmemcheck_slab_free(s, x, s->object_size);
1273 		debug_check_no_locks_freed(x, s->object_size);
1274 		local_irq_restore(flags);
1275 	}
1276 #endif
1277 	if (!(s->flags & SLAB_DEBUG_OBJECTS))
1278 		debug_check_no_obj_freed(x, s->object_size);
1279 }
1280 
1281 /*
1282  * Slab allocation and freeing
1283  */
1284 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1285 		gfp_t flags, int node, struct kmem_cache_order_objects oo)
1286 {
1287 	struct page *page;
1288 	int order = oo_order(oo);
1289 
1290 	flags |= __GFP_NOTRACK;
1291 
1292 	if (memcg_charge_slab(s, flags, order))
1293 		return NULL;
1294 
1295 	if (node == NUMA_NO_NODE)
1296 		page = alloc_pages(flags, order);
1297 	else
1298 		page = alloc_pages_exact_node(node, flags, order);
1299 
1300 	if (!page)
1301 		memcg_uncharge_slab(s, order);
1302 
1303 	return page;
1304 }
1305 
1306 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1307 {
1308 	struct page *page;
1309 	struct kmem_cache_order_objects oo = s->oo;
1310 	gfp_t alloc_gfp;
1311 
1312 	flags &= gfp_allowed_mask;
1313 
1314 	if (flags & __GFP_WAIT)
1315 		local_irq_enable();
1316 
1317 	flags |= s->allocflags;
1318 
1319 	/*
1320 	 * Let the initial higher-order allocation fail under memory pressure
1321 	 * so we fall-back to the minimum order allocation.
1322 	 */
1323 	alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1324 
1325 	page = alloc_slab_page(s, alloc_gfp, node, oo);
1326 	if (unlikely(!page)) {
1327 		oo = s->min;
1328 		alloc_gfp = flags;
1329 		/*
1330 		 * Allocation may have failed due to fragmentation.
1331 		 * Try a lower order alloc if possible
1332 		 */
1333 		page = alloc_slab_page(s, alloc_gfp, node, oo);
1334 
1335 		if (page)
1336 			stat(s, ORDER_FALLBACK);
1337 	}
1338 
1339 	if (kmemcheck_enabled && page
1340 		&& !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1341 		int pages = 1 << oo_order(oo);
1342 
1343 		kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node);
1344 
1345 		/*
1346 		 * Objects from caches that have a constructor don't get
1347 		 * cleared when they're allocated, so we need to do it here.
1348 		 */
1349 		if (s->ctor)
1350 			kmemcheck_mark_uninitialized_pages(page, pages);
1351 		else
1352 			kmemcheck_mark_unallocated_pages(page, pages);
1353 	}
1354 
1355 	if (flags & __GFP_WAIT)
1356 		local_irq_disable();
1357 	if (!page)
1358 		return NULL;
1359 
1360 	page->objects = oo_objects(oo);
1361 	mod_zone_page_state(page_zone(page),
1362 		(s->flags & SLAB_RECLAIM_ACCOUNT) ?
1363 		NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1364 		1 << oo_order(oo));
1365 
1366 	return page;
1367 }
1368 
1369 static void setup_object(struct kmem_cache *s, struct page *page,
1370 				void *object)
1371 {
1372 	setup_object_debug(s, page, object);
1373 	if (unlikely(s->ctor))
1374 		s->ctor(object);
1375 }
1376 
1377 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1378 {
1379 	struct page *page;
1380 	void *start;
1381 	void *p;
1382 	int order;
1383 	int idx;
1384 
1385 	if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1386 		pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
1387 		BUG();
1388 	}
1389 
1390 	page = allocate_slab(s,
1391 		flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1392 	if (!page)
1393 		goto out;
1394 
1395 	order = compound_order(page);
1396 	inc_slabs_node(s, page_to_nid(page), page->objects);
1397 	page->slab_cache = s;
1398 	__SetPageSlab(page);
1399 	if (page->pfmemalloc)
1400 		SetPageSlabPfmemalloc(page);
1401 
1402 	start = page_address(page);
1403 
1404 	if (unlikely(s->flags & SLAB_POISON))
1405 		memset(start, POISON_INUSE, PAGE_SIZE << order);
1406 
1407 	for_each_object_idx(p, idx, s, start, page->objects) {
1408 		setup_object(s, page, p);
1409 		if (likely(idx < page->objects))
1410 			set_freepointer(s, p, p + s->size);
1411 		else
1412 			set_freepointer(s, p, NULL);
1413 	}
1414 
1415 	page->freelist = start;
1416 	page->inuse = page->objects;
1417 	page->frozen = 1;
1418 out:
1419 	return page;
1420 }
1421 
1422 static void __free_slab(struct kmem_cache *s, struct page *page)
1423 {
1424 	int order = compound_order(page);
1425 	int pages = 1 << order;
1426 
1427 	if (kmem_cache_debug(s)) {
1428 		void *p;
1429 
1430 		slab_pad_check(s, page);
1431 		for_each_object(p, s, page_address(page),
1432 						page->objects)
1433 			check_object(s, page, p, SLUB_RED_INACTIVE);
1434 	}
1435 
1436 	kmemcheck_free_shadow(page, compound_order(page));
1437 
1438 	mod_zone_page_state(page_zone(page),
1439 		(s->flags & SLAB_RECLAIM_ACCOUNT) ?
1440 		NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1441 		-pages);
1442 
1443 	__ClearPageSlabPfmemalloc(page);
1444 	__ClearPageSlab(page);
1445 
1446 	page_mapcount_reset(page);
1447 	if (current->reclaim_state)
1448 		current->reclaim_state->reclaimed_slab += pages;
1449 	__free_pages(page, order);
1450 	memcg_uncharge_slab(s, order);
1451 }
1452 
1453 #define need_reserve_slab_rcu						\
1454 	(sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1455 
1456 static void rcu_free_slab(struct rcu_head *h)
1457 {
1458 	struct page *page;
1459 
1460 	if (need_reserve_slab_rcu)
1461 		page = virt_to_head_page(h);
1462 	else
1463 		page = container_of((struct list_head *)h, struct page, lru);
1464 
1465 	__free_slab(page->slab_cache, page);
1466 }
1467 
1468 static void free_slab(struct kmem_cache *s, struct page *page)
1469 {
1470 	if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1471 		struct rcu_head *head;
1472 
1473 		if (need_reserve_slab_rcu) {
1474 			int order = compound_order(page);
1475 			int offset = (PAGE_SIZE << order) - s->reserved;
1476 
1477 			VM_BUG_ON(s->reserved != sizeof(*head));
1478 			head = page_address(page) + offset;
1479 		} else {
1480 			/*
1481 			 * RCU free overloads the RCU head over the LRU
1482 			 */
1483 			head = (void *)&page->lru;
1484 		}
1485 
1486 		call_rcu(head, rcu_free_slab);
1487 	} else
1488 		__free_slab(s, page);
1489 }
1490 
1491 static void discard_slab(struct kmem_cache *s, struct page *page)
1492 {
1493 	dec_slabs_node(s, page_to_nid(page), page->objects);
1494 	free_slab(s, page);
1495 }
1496 
1497 /*
1498  * Management of partially allocated slabs.
1499  */
1500 static inline void
1501 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1502 {
1503 	n->nr_partial++;
1504 	if (tail == DEACTIVATE_TO_TAIL)
1505 		list_add_tail(&page->lru, &n->partial);
1506 	else
1507 		list_add(&page->lru, &n->partial);
1508 }
1509 
1510 static inline void add_partial(struct kmem_cache_node *n,
1511 				struct page *page, int tail)
1512 {
1513 	lockdep_assert_held(&n->list_lock);
1514 	__add_partial(n, page, tail);
1515 }
1516 
1517 static inline void
1518 __remove_partial(struct kmem_cache_node *n, struct page *page)
1519 {
1520 	list_del(&page->lru);
1521 	n->nr_partial--;
1522 }
1523 
1524 static inline void remove_partial(struct kmem_cache_node *n,
1525 					struct page *page)
1526 {
1527 	lockdep_assert_held(&n->list_lock);
1528 	__remove_partial(n, page);
1529 }
1530 
1531 /*
1532  * Remove slab from the partial list, freeze it and
1533  * return the pointer to the freelist.
1534  *
1535  * Returns a list of objects or NULL if it fails.
1536  */
1537 static inline void *acquire_slab(struct kmem_cache *s,
1538 		struct kmem_cache_node *n, struct page *page,
1539 		int mode, int *objects)
1540 {
1541 	void *freelist;
1542 	unsigned long counters;
1543 	struct page new;
1544 
1545 	lockdep_assert_held(&n->list_lock);
1546 
1547 	/*
1548 	 * Zap the freelist and set the frozen bit.
1549 	 * The old freelist is the list of objects for the
1550 	 * per cpu allocation list.
1551 	 */
1552 	freelist = page->freelist;
1553 	counters = page->counters;
1554 	new.counters = counters;
1555 	*objects = new.objects - new.inuse;
1556 	if (mode) {
1557 		new.inuse = page->objects;
1558 		new.freelist = NULL;
1559 	} else {
1560 		new.freelist = freelist;
1561 	}
1562 
1563 	VM_BUG_ON(new.frozen);
1564 	new.frozen = 1;
1565 
1566 	if (!__cmpxchg_double_slab(s, page,
1567 			freelist, counters,
1568 			new.freelist, new.counters,
1569 			"acquire_slab"))
1570 		return NULL;
1571 
1572 	remove_partial(n, page);
1573 	WARN_ON(!freelist);
1574 	return freelist;
1575 }
1576 
1577 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1578 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1579 
1580 /*
1581  * Try to allocate a partial slab from a specific node.
1582  */
1583 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1584 				struct kmem_cache_cpu *c, gfp_t flags)
1585 {
1586 	struct page *page, *page2;
1587 	void *object = NULL;
1588 	int available = 0;
1589 	int objects;
1590 
1591 	/*
1592 	 * Racy check. If we mistakenly see no partial slabs then we
1593 	 * just allocate an empty slab. If we mistakenly try to get a
1594 	 * partial slab and there is none available then get_partials()
1595 	 * will return NULL.
1596 	 */
1597 	if (!n || !n->nr_partial)
1598 		return NULL;
1599 
1600 	spin_lock(&n->list_lock);
1601 	list_for_each_entry_safe(page, page2, &n->partial, lru) {
1602 		void *t;
1603 
1604 		if (!pfmemalloc_match(page, flags))
1605 			continue;
1606 
1607 		t = acquire_slab(s, n, page, object == NULL, &objects);
1608 		if (!t)
1609 			break;
1610 
1611 		available += objects;
1612 		if (!object) {
1613 			c->page = page;
1614 			stat(s, ALLOC_FROM_PARTIAL);
1615 			object = t;
1616 		} else {
1617 			put_cpu_partial(s, page, 0);
1618 			stat(s, CPU_PARTIAL_NODE);
1619 		}
1620 		if (!kmem_cache_has_cpu_partial(s)
1621 			|| available > s->cpu_partial / 2)
1622 			break;
1623 
1624 	}
1625 	spin_unlock(&n->list_lock);
1626 	return object;
1627 }
1628 
1629 /*
1630  * Get a page from somewhere. Search in increasing NUMA distances.
1631  */
1632 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1633 		struct kmem_cache_cpu *c)
1634 {
1635 #ifdef CONFIG_NUMA
1636 	struct zonelist *zonelist;
1637 	struct zoneref *z;
1638 	struct zone *zone;
1639 	enum zone_type high_zoneidx = gfp_zone(flags);
1640 	void *object;
1641 	unsigned int cpuset_mems_cookie;
1642 
1643 	/*
1644 	 * The defrag ratio allows a configuration of the tradeoffs between
1645 	 * inter node defragmentation and node local allocations. A lower
1646 	 * defrag_ratio increases the tendency to do local allocations
1647 	 * instead of attempting to obtain partial slabs from other nodes.
1648 	 *
1649 	 * If the defrag_ratio is set to 0 then kmalloc() always
1650 	 * returns node local objects. If the ratio is higher then kmalloc()
1651 	 * may return off node objects because partial slabs are obtained
1652 	 * from other nodes and filled up.
1653 	 *
1654 	 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1655 	 * defrag_ratio = 1000) then every (well almost) allocation will
1656 	 * first attempt to defrag slab caches on other nodes. This means
1657 	 * scanning over all nodes to look for partial slabs which may be
1658 	 * expensive if we do it every time we are trying to find a slab
1659 	 * with available objects.
1660 	 */
1661 	if (!s->remote_node_defrag_ratio ||
1662 			get_cycles() % 1024 > s->remote_node_defrag_ratio)
1663 		return NULL;
1664 
1665 	do {
1666 		cpuset_mems_cookie = read_mems_allowed_begin();
1667 		zonelist = node_zonelist(mempolicy_slab_node(), flags);
1668 		for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1669 			struct kmem_cache_node *n;
1670 
1671 			n = get_node(s, zone_to_nid(zone));
1672 
1673 			if (n && cpuset_zone_allowed(zone, flags) &&
1674 					n->nr_partial > s->min_partial) {
1675 				object = get_partial_node(s, n, c, flags);
1676 				if (object) {
1677 					/*
1678 					 * Don't check read_mems_allowed_retry()
1679 					 * here - if mems_allowed was updated in
1680 					 * parallel, that was a harmless race
1681 					 * between allocation and the cpuset
1682 					 * update
1683 					 */
1684 					return object;
1685 				}
1686 			}
1687 		}
1688 	} while (read_mems_allowed_retry(cpuset_mems_cookie));
1689 #endif
1690 	return NULL;
1691 }
1692 
1693 /*
1694  * Get a partial page, lock it and return it.
1695  */
1696 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1697 		struct kmem_cache_cpu *c)
1698 {
1699 	void *object;
1700 	int searchnode = node;
1701 
1702 	if (node == NUMA_NO_NODE)
1703 		searchnode = numa_mem_id();
1704 	else if (!node_present_pages(node))
1705 		searchnode = node_to_mem_node(node);
1706 
1707 	object = get_partial_node(s, get_node(s, searchnode), c, flags);
1708 	if (object || node != NUMA_NO_NODE)
1709 		return object;
1710 
1711 	return get_any_partial(s, flags, c);
1712 }
1713 
1714 #ifdef CONFIG_PREEMPT
1715 /*
1716  * Calculate the next globally unique transaction for disambiguiation
1717  * during cmpxchg. The transactions start with the cpu number and are then
1718  * incremented by CONFIG_NR_CPUS.
1719  */
1720 #define TID_STEP  roundup_pow_of_two(CONFIG_NR_CPUS)
1721 #else
1722 /*
1723  * No preemption supported therefore also no need to check for
1724  * different cpus.
1725  */
1726 #define TID_STEP 1
1727 #endif
1728 
1729 static inline unsigned long next_tid(unsigned long tid)
1730 {
1731 	return tid + TID_STEP;
1732 }
1733 
1734 static inline unsigned int tid_to_cpu(unsigned long tid)
1735 {
1736 	return tid % TID_STEP;
1737 }
1738 
1739 static inline unsigned long tid_to_event(unsigned long tid)
1740 {
1741 	return tid / TID_STEP;
1742 }
1743 
1744 static inline unsigned int init_tid(int cpu)
1745 {
1746 	return cpu;
1747 }
1748 
1749 static inline void note_cmpxchg_failure(const char *n,
1750 		const struct kmem_cache *s, unsigned long tid)
1751 {
1752 #ifdef SLUB_DEBUG_CMPXCHG
1753 	unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1754 
1755 	pr_info("%s %s: cmpxchg redo ", n, s->name);
1756 
1757 #ifdef CONFIG_PREEMPT
1758 	if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1759 		pr_warn("due to cpu change %d -> %d\n",
1760 			tid_to_cpu(tid), tid_to_cpu(actual_tid));
1761 	else
1762 #endif
1763 	if (tid_to_event(tid) != tid_to_event(actual_tid))
1764 		pr_warn("due to cpu running other code. Event %ld->%ld\n",
1765 			tid_to_event(tid), tid_to_event(actual_tid));
1766 	else
1767 		pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1768 			actual_tid, tid, next_tid(tid));
1769 #endif
1770 	stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1771 }
1772 
1773 static void init_kmem_cache_cpus(struct kmem_cache *s)
1774 {
1775 	int cpu;
1776 
1777 	for_each_possible_cpu(cpu)
1778 		per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1779 }
1780 
1781 /*
1782  * Remove the cpu slab
1783  */
1784 static void deactivate_slab(struct kmem_cache *s, struct page *page,
1785 				void *freelist)
1786 {
1787 	enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1788 	struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1789 	int lock = 0;
1790 	enum slab_modes l = M_NONE, m = M_NONE;
1791 	void *nextfree;
1792 	int tail = DEACTIVATE_TO_HEAD;
1793 	struct page new;
1794 	struct page old;
1795 
1796 	if (page->freelist) {
1797 		stat(s, DEACTIVATE_REMOTE_FREES);
1798 		tail = DEACTIVATE_TO_TAIL;
1799 	}
1800 
1801 	/*
1802 	 * Stage one: Free all available per cpu objects back
1803 	 * to the page freelist while it is still frozen. Leave the
1804 	 * last one.
1805 	 *
1806 	 * There is no need to take the list->lock because the page
1807 	 * is still frozen.
1808 	 */
1809 	while (freelist && (nextfree = get_freepointer(s, freelist))) {
1810 		void *prior;
1811 		unsigned long counters;
1812 
1813 		do {
1814 			prior = page->freelist;
1815 			counters = page->counters;
1816 			set_freepointer(s, freelist, prior);
1817 			new.counters = counters;
1818 			new.inuse--;
1819 			VM_BUG_ON(!new.frozen);
1820 
1821 		} while (!__cmpxchg_double_slab(s, page,
1822 			prior, counters,
1823 			freelist, new.counters,
1824 			"drain percpu freelist"));
1825 
1826 		freelist = nextfree;
1827 	}
1828 
1829 	/*
1830 	 * Stage two: Ensure that the page is unfrozen while the
1831 	 * list presence reflects the actual number of objects
1832 	 * during unfreeze.
1833 	 *
1834 	 * We setup the list membership and then perform a cmpxchg
1835 	 * with the count. If there is a mismatch then the page
1836 	 * is not unfrozen but the page is on the wrong list.
1837 	 *
1838 	 * Then we restart the process which may have to remove
1839 	 * the page from the list that we just put it on again
1840 	 * because the number of objects in the slab may have
1841 	 * changed.
1842 	 */
1843 redo:
1844 
1845 	old.freelist = page->freelist;
1846 	old.counters = page->counters;
1847 	VM_BUG_ON(!old.frozen);
1848 
1849 	/* Determine target state of the slab */
1850 	new.counters = old.counters;
1851 	if (freelist) {
1852 		new.inuse--;
1853 		set_freepointer(s, freelist, old.freelist);
1854 		new.freelist = freelist;
1855 	} else
1856 		new.freelist = old.freelist;
1857 
1858 	new.frozen = 0;
1859 
1860 	if (!new.inuse && n->nr_partial >= s->min_partial)
1861 		m = M_FREE;
1862 	else if (new.freelist) {
1863 		m = M_PARTIAL;
1864 		if (!lock) {
1865 			lock = 1;
1866 			/*
1867 			 * Taking the spinlock removes the possiblity
1868 			 * that acquire_slab() will see a slab page that
1869 			 * is frozen
1870 			 */
1871 			spin_lock(&n->list_lock);
1872 		}
1873 	} else {
1874 		m = M_FULL;
1875 		if (kmem_cache_debug(s) && !lock) {
1876 			lock = 1;
1877 			/*
1878 			 * This also ensures that the scanning of full
1879 			 * slabs from diagnostic functions will not see
1880 			 * any frozen slabs.
1881 			 */
1882 			spin_lock(&n->list_lock);
1883 		}
1884 	}
1885 
1886 	if (l != m) {
1887 
1888 		if (l == M_PARTIAL)
1889 
1890 			remove_partial(n, page);
1891 
1892 		else if (l == M_FULL)
1893 
1894 			remove_full(s, n, page);
1895 
1896 		if (m == M_PARTIAL) {
1897 
1898 			add_partial(n, page, tail);
1899 			stat(s, tail);
1900 
1901 		} else if (m == M_FULL) {
1902 
1903 			stat(s, DEACTIVATE_FULL);
1904 			add_full(s, n, page);
1905 
1906 		}
1907 	}
1908 
1909 	l = m;
1910 	if (!__cmpxchg_double_slab(s, page,
1911 				old.freelist, old.counters,
1912 				new.freelist, new.counters,
1913 				"unfreezing slab"))
1914 		goto redo;
1915 
1916 	if (lock)
1917 		spin_unlock(&n->list_lock);
1918 
1919 	if (m == M_FREE) {
1920 		stat(s, DEACTIVATE_EMPTY);
1921 		discard_slab(s, page);
1922 		stat(s, FREE_SLAB);
1923 	}
1924 }
1925 
1926 /*
1927  * Unfreeze all the cpu partial slabs.
1928  *
1929  * This function must be called with interrupts disabled
1930  * for the cpu using c (or some other guarantee must be there
1931  * to guarantee no concurrent accesses).
1932  */
1933 static void unfreeze_partials(struct kmem_cache *s,
1934 		struct kmem_cache_cpu *c)
1935 {
1936 #ifdef CONFIG_SLUB_CPU_PARTIAL
1937 	struct kmem_cache_node *n = NULL, *n2 = NULL;
1938 	struct page *page, *discard_page = NULL;
1939 
1940 	while ((page = c->partial)) {
1941 		struct page new;
1942 		struct page old;
1943 
1944 		c->partial = page->next;
1945 
1946 		n2 = get_node(s, page_to_nid(page));
1947 		if (n != n2) {
1948 			if (n)
1949 				spin_unlock(&n->list_lock);
1950 
1951 			n = n2;
1952 			spin_lock(&n->list_lock);
1953 		}
1954 
1955 		do {
1956 
1957 			old.freelist = page->freelist;
1958 			old.counters = page->counters;
1959 			VM_BUG_ON(!old.frozen);
1960 
1961 			new.counters = old.counters;
1962 			new.freelist = old.freelist;
1963 
1964 			new.frozen = 0;
1965 
1966 		} while (!__cmpxchg_double_slab(s, page,
1967 				old.freelist, old.counters,
1968 				new.freelist, new.counters,
1969 				"unfreezing slab"));
1970 
1971 		if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
1972 			page->next = discard_page;
1973 			discard_page = page;
1974 		} else {
1975 			add_partial(n, page, DEACTIVATE_TO_TAIL);
1976 			stat(s, FREE_ADD_PARTIAL);
1977 		}
1978 	}
1979 
1980 	if (n)
1981 		spin_unlock(&n->list_lock);
1982 
1983 	while (discard_page) {
1984 		page = discard_page;
1985 		discard_page = discard_page->next;
1986 
1987 		stat(s, DEACTIVATE_EMPTY);
1988 		discard_slab(s, page);
1989 		stat(s, FREE_SLAB);
1990 	}
1991 #endif
1992 }
1993 
1994 /*
1995  * Put a page that was just frozen (in __slab_free) into a partial page
1996  * slot if available. This is done without interrupts disabled and without
1997  * preemption disabled. The cmpxchg is racy and may put the partial page
1998  * onto a random cpus partial slot.
1999  *
2000  * If we did not find a slot then simply move all the partials to the
2001  * per node partial list.
2002  */
2003 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2004 {
2005 #ifdef CONFIG_SLUB_CPU_PARTIAL
2006 	struct page *oldpage;
2007 	int pages;
2008 	int pobjects;
2009 
2010 	do {
2011 		pages = 0;
2012 		pobjects = 0;
2013 		oldpage = this_cpu_read(s->cpu_slab->partial);
2014 
2015 		if (oldpage) {
2016 			pobjects = oldpage->pobjects;
2017 			pages = oldpage->pages;
2018 			if (drain && pobjects > s->cpu_partial) {
2019 				unsigned long flags;
2020 				/*
2021 				 * partial array is full. Move the existing
2022 				 * set to the per node partial list.
2023 				 */
2024 				local_irq_save(flags);
2025 				unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2026 				local_irq_restore(flags);
2027 				oldpage = NULL;
2028 				pobjects = 0;
2029 				pages = 0;
2030 				stat(s, CPU_PARTIAL_DRAIN);
2031 			}
2032 		}
2033 
2034 		pages++;
2035 		pobjects += page->objects - page->inuse;
2036 
2037 		page->pages = pages;
2038 		page->pobjects = pobjects;
2039 		page->next = oldpage;
2040 
2041 	} while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2042 								!= oldpage);
2043 #endif
2044 }
2045 
2046 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2047 {
2048 	stat(s, CPUSLAB_FLUSH);
2049 	deactivate_slab(s, c->page, c->freelist);
2050 
2051 	c->tid = next_tid(c->tid);
2052 	c->page = NULL;
2053 	c->freelist = NULL;
2054 }
2055 
2056 /*
2057  * Flush cpu slab.
2058  *
2059  * Called from IPI handler with interrupts disabled.
2060  */
2061 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2062 {
2063 	struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2064 
2065 	if (likely(c)) {
2066 		if (c->page)
2067 			flush_slab(s, c);
2068 
2069 		unfreeze_partials(s, c);
2070 	}
2071 }
2072 
2073 static void flush_cpu_slab(void *d)
2074 {
2075 	struct kmem_cache *s = d;
2076 
2077 	__flush_cpu_slab(s, smp_processor_id());
2078 }
2079 
2080 static bool has_cpu_slab(int cpu, void *info)
2081 {
2082 	struct kmem_cache *s = info;
2083 	struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2084 
2085 	return c->page || c->partial;
2086 }
2087 
2088 static void flush_all(struct kmem_cache *s)
2089 {
2090 	on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2091 }
2092 
2093 /*
2094  * Check if the objects in a per cpu structure fit numa
2095  * locality expectations.
2096  */
2097 static inline int node_match(struct page *page, int node)
2098 {
2099 #ifdef CONFIG_NUMA
2100 	if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2101 		return 0;
2102 #endif
2103 	return 1;
2104 }
2105 
2106 #ifdef CONFIG_SLUB_DEBUG
2107 static int count_free(struct page *page)
2108 {
2109 	return page->objects - page->inuse;
2110 }
2111 
2112 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2113 {
2114 	return atomic_long_read(&n->total_objects);
2115 }
2116 #endif /* CONFIG_SLUB_DEBUG */
2117 
2118 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2119 static unsigned long count_partial(struct kmem_cache_node *n,
2120 					int (*get_count)(struct page *))
2121 {
2122 	unsigned long flags;
2123 	unsigned long x = 0;
2124 	struct page *page;
2125 
2126 	spin_lock_irqsave(&n->list_lock, flags);
2127 	list_for_each_entry(page, &n->partial, lru)
2128 		x += get_count(page);
2129 	spin_unlock_irqrestore(&n->list_lock, flags);
2130 	return x;
2131 }
2132 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2133 
2134 static noinline void
2135 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2136 {
2137 #ifdef CONFIG_SLUB_DEBUG
2138 	static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2139 				      DEFAULT_RATELIMIT_BURST);
2140 	int node;
2141 	struct kmem_cache_node *n;
2142 
2143 	if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2144 		return;
2145 
2146 	pr_warn("SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2147 		nid, gfpflags);
2148 	pr_warn("  cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2149 		s->name, s->object_size, s->size, oo_order(s->oo),
2150 		oo_order(s->min));
2151 
2152 	if (oo_order(s->min) > get_order(s->object_size))
2153 		pr_warn("  %s debugging increased min order, use slub_debug=O to disable.\n",
2154 			s->name);
2155 
2156 	for_each_kmem_cache_node(s, node, n) {
2157 		unsigned long nr_slabs;
2158 		unsigned long nr_objs;
2159 		unsigned long nr_free;
2160 
2161 		nr_free  = count_partial(n, count_free);
2162 		nr_slabs = node_nr_slabs(n);
2163 		nr_objs  = node_nr_objs(n);
2164 
2165 		pr_warn("  node %d: slabs: %ld, objs: %ld, free: %ld\n",
2166 			node, nr_slabs, nr_objs, nr_free);
2167 	}
2168 #endif
2169 }
2170 
2171 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2172 			int node, struct kmem_cache_cpu **pc)
2173 {
2174 	void *freelist;
2175 	struct kmem_cache_cpu *c = *pc;
2176 	struct page *page;
2177 
2178 	freelist = get_partial(s, flags, node, c);
2179 
2180 	if (freelist)
2181 		return freelist;
2182 
2183 	page = new_slab(s, flags, node);
2184 	if (page) {
2185 		c = raw_cpu_ptr(s->cpu_slab);
2186 		if (c->page)
2187 			flush_slab(s, c);
2188 
2189 		/*
2190 		 * No other reference to the page yet so we can
2191 		 * muck around with it freely without cmpxchg
2192 		 */
2193 		freelist = page->freelist;
2194 		page->freelist = NULL;
2195 
2196 		stat(s, ALLOC_SLAB);
2197 		c->page = page;
2198 		*pc = c;
2199 	} else
2200 		freelist = NULL;
2201 
2202 	return freelist;
2203 }
2204 
2205 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2206 {
2207 	if (unlikely(PageSlabPfmemalloc(page)))
2208 		return gfp_pfmemalloc_allowed(gfpflags);
2209 
2210 	return true;
2211 }
2212 
2213 /*
2214  * Check the page->freelist of a page and either transfer the freelist to the
2215  * per cpu freelist or deactivate the page.
2216  *
2217  * The page is still frozen if the return value is not NULL.
2218  *
2219  * If this function returns NULL then the page has been unfrozen.
2220  *
2221  * This function must be called with interrupt disabled.
2222  */
2223 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2224 {
2225 	struct page new;
2226 	unsigned long counters;
2227 	void *freelist;
2228 
2229 	do {
2230 		freelist = page->freelist;
2231 		counters = page->counters;
2232 
2233 		new.counters = counters;
2234 		VM_BUG_ON(!new.frozen);
2235 
2236 		new.inuse = page->objects;
2237 		new.frozen = freelist != NULL;
2238 
2239 	} while (!__cmpxchg_double_slab(s, page,
2240 		freelist, counters,
2241 		NULL, new.counters,
2242 		"get_freelist"));
2243 
2244 	return freelist;
2245 }
2246 
2247 /*
2248  * Slow path. The lockless freelist is empty or we need to perform
2249  * debugging duties.
2250  *
2251  * Processing is still very fast if new objects have been freed to the
2252  * regular freelist. In that case we simply take over the regular freelist
2253  * as the lockless freelist and zap the regular freelist.
2254  *
2255  * If that is not working then we fall back to the partial lists. We take the
2256  * first element of the freelist as the object to allocate now and move the
2257  * rest of the freelist to the lockless freelist.
2258  *
2259  * And if we were unable to get a new slab from the partial slab lists then
2260  * we need to allocate a new slab. This is the slowest path since it involves
2261  * a call to the page allocator and the setup of a new slab.
2262  */
2263 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2264 			  unsigned long addr, struct kmem_cache_cpu *c)
2265 {
2266 	void *freelist;
2267 	struct page *page;
2268 	unsigned long flags;
2269 
2270 	local_irq_save(flags);
2271 #ifdef CONFIG_PREEMPT
2272 	/*
2273 	 * We may have been preempted and rescheduled on a different
2274 	 * cpu before disabling interrupts. Need to reload cpu area
2275 	 * pointer.
2276 	 */
2277 	c = this_cpu_ptr(s->cpu_slab);
2278 #endif
2279 
2280 	page = c->page;
2281 	if (!page)
2282 		goto new_slab;
2283 redo:
2284 
2285 	if (unlikely(!node_match(page, node))) {
2286 		int searchnode = node;
2287 
2288 		if (node != NUMA_NO_NODE && !node_present_pages(node))
2289 			searchnode = node_to_mem_node(node);
2290 
2291 		if (unlikely(!node_match(page, searchnode))) {
2292 			stat(s, ALLOC_NODE_MISMATCH);
2293 			deactivate_slab(s, page, c->freelist);
2294 			c->page = NULL;
2295 			c->freelist = NULL;
2296 			goto new_slab;
2297 		}
2298 	}
2299 
2300 	/*
2301 	 * By rights, we should be searching for a slab page that was
2302 	 * PFMEMALLOC but right now, we are losing the pfmemalloc
2303 	 * information when the page leaves the per-cpu allocator
2304 	 */
2305 	if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2306 		deactivate_slab(s, page, c->freelist);
2307 		c->page = NULL;
2308 		c->freelist = NULL;
2309 		goto new_slab;
2310 	}
2311 
2312 	/* must check again c->freelist in case of cpu migration or IRQ */
2313 	freelist = c->freelist;
2314 	if (freelist)
2315 		goto load_freelist;
2316 
2317 	freelist = get_freelist(s, page);
2318 
2319 	if (!freelist) {
2320 		c->page = NULL;
2321 		stat(s, DEACTIVATE_BYPASS);
2322 		goto new_slab;
2323 	}
2324 
2325 	stat(s, ALLOC_REFILL);
2326 
2327 load_freelist:
2328 	/*
2329 	 * freelist is pointing to the list of objects to be used.
2330 	 * page is pointing to the page from which the objects are obtained.
2331 	 * That page must be frozen for per cpu allocations to work.
2332 	 */
2333 	VM_BUG_ON(!c->page->frozen);
2334 	c->freelist = get_freepointer(s, freelist);
2335 	c->tid = next_tid(c->tid);
2336 	local_irq_restore(flags);
2337 	return freelist;
2338 
2339 new_slab:
2340 
2341 	if (c->partial) {
2342 		page = c->page = c->partial;
2343 		c->partial = page->next;
2344 		stat(s, CPU_PARTIAL_ALLOC);
2345 		c->freelist = NULL;
2346 		goto redo;
2347 	}
2348 
2349 	freelist = new_slab_objects(s, gfpflags, node, &c);
2350 
2351 	if (unlikely(!freelist)) {
2352 		slab_out_of_memory(s, gfpflags, node);
2353 		local_irq_restore(flags);
2354 		return NULL;
2355 	}
2356 
2357 	page = c->page;
2358 	if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2359 		goto load_freelist;
2360 
2361 	/* Only entered in the debug case */
2362 	if (kmem_cache_debug(s) &&
2363 			!alloc_debug_processing(s, page, freelist, addr))
2364 		goto new_slab;	/* Slab failed checks. Next slab needed */
2365 
2366 	deactivate_slab(s, page, get_freepointer(s, freelist));
2367 	c->page = NULL;
2368 	c->freelist = NULL;
2369 	local_irq_restore(flags);
2370 	return freelist;
2371 }
2372 
2373 /*
2374  * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2375  * have the fastpath folded into their functions. So no function call
2376  * overhead for requests that can be satisfied on the fastpath.
2377  *
2378  * The fastpath works by first checking if the lockless freelist can be used.
2379  * If not then __slab_alloc is called for slow processing.
2380  *
2381  * Otherwise we can simply pick the next object from the lockless free list.
2382  */
2383 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2384 		gfp_t gfpflags, int node, unsigned long addr)
2385 {
2386 	void **object;
2387 	struct kmem_cache_cpu *c;
2388 	struct page *page;
2389 	unsigned long tid;
2390 
2391 	s = slab_pre_alloc_hook(s, gfpflags);
2392 	if (!s)
2393 		return NULL;
2394 redo:
2395 	/*
2396 	 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2397 	 * enabled. We may switch back and forth between cpus while
2398 	 * reading from one cpu area. That does not matter as long
2399 	 * as we end up on the original cpu again when doing the cmpxchg.
2400 	 *
2401 	 * Preemption is disabled for the retrieval of the tid because that
2402 	 * must occur from the current processor. We cannot allow rescheduling
2403 	 * on a different processor between the determination of the pointer
2404 	 * and the retrieval of the tid.
2405 	 */
2406 	preempt_disable();
2407 	c = this_cpu_ptr(s->cpu_slab);
2408 
2409 	/*
2410 	 * The transaction ids are globally unique per cpu and per operation on
2411 	 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2412 	 * occurs on the right processor and that there was no operation on the
2413 	 * linked list in between.
2414 	 */
2415 	tid = c->tid;
2416 	preempt_enable();
2417 
2418 	object = c->freelist;
2419 	page = c->page;
2420 	if (unlikely(!object || !node_match(page, node))) {
2421 		object = __slab_alloc(s, gfpflags, node, addr, c);
2422 		stat(s, ALLOC_SLOWPATH);
2423 	} else {
2424 		void *next_object = get_freepointer_safe(s, object);
2425 
2426 		/*
2427 		 * The cmpxchg will only match if there was no additional
2428 		 * operation and if we are on the right processor.
2429 		 *
2430 		 * The cmpxchg does the following atomically (without lock
2431 		 * semantics!)
2432 		 * 1. Relocate first pointer to the current per cpu area.
2433 		 * 2. Verify that tid and freelist have not been changed
2434 		 * 3. If they were not changed replace tid and freelist
2435 		 *
2436 		 * Since this is without lock semantics the protection is only
2437 		 * against code executing on this cpu *not* from access by
2438 		 * other cpus.
2439 		 */
2440 		if (unlikely(!this_cpu_cmpxchg_double(
2441 				s->cpu_slab->freelist, s->cpu_slab->tid,
2442 				object, tid,
2443 				next_object, next_tid(tid)))) {
2444 
2445 			note_cmpxchg_failure("slab_alloc", s, tid);
2446 			goto redo;
2447 		}
2448 		prefetch_freepointer(s, next_object);
2449 		stat(s, ALLOC_FASTPATH);
2450 	}
2451 
2452 	if (unlikely(gfpflags & __GFP_ZERO) && object)
2453 		memset(object, 0, s->object_size);
2454 
2455 	slab_post_alloc_hook(s, gfpflags, object);
2456 
2457 	return object;
2458 }
2459 
2460 static __always_inline void *slab_alloc(struct kmem_cache *s,
2461 		gfp_t gfpflags, unsigned long addr)
2462 {
2463 	return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2464 }
2465 
2466 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2467 {
2468 	void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2469 
2470 	trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2471 				s->size, gfpflags);
2472 
2473 	return ret;
2474 }
2475 EXPORT_SYMBOL(kmem_cache_alloc);
2476 
2477 #ifdef CONFIG_TRACING
2478 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2479 {
2480 	void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2481 	trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2482 	return ret;
2483 }
2484 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2485 #endif
2486 
2487 #ifdef CONFIG_NUMA
2488 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2489 {
2490 	void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2491 
2492 	trace_kmem_cache_alloc_node(_RET_IP_, ret,
2493 				    s->object_size, s->size, gfpflags, node);
2494 
2495 	return ret;
2496 }
2497 EXPORT_SYMBOL(kmem_cache_alloc_node);
2498 
2499 #ifdef CONFIG_TRACING
2500 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2501 				    gfp_t gfpflags,
2502 				    int node, size_t size)
2503 {
2504 	void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2505 
2506 	trace_kmalloc_node(_RET_IP_, ret,
2507 			   size, s->size, gfpflags, node);
2508 	return ret;
2509 }
2510 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2511 #endif
2512 #endif
2513 
2514 /*
2515  * Slow patch handling. This may still be called frequently since objects
2516  * have a longer lifetime than the cpu slabs in most processing loads.
2517  *
2518  * So we still attempt to reduce cache line usage. Just take the slab
2519  * lock and free the item. If there is no additional partial page
2520  * handling required then we can return immediately.
2521  */
2522 static void __slab_free(struct kmem_cache *s, struct page *page,
2523 			void *x, unsigned long addr)
2524 {
2525 	void *prior;
2526 	void **object = (void *)x;
2527 	int was_frozen;
2528 	struct page new;
2529 	unsigned long counters;
2530 	struct kmem_cache_node *n = NULL;
2531 	unsigned long uninitialized_var(flags);
2532 
2533 	stat(s, FREE_SLOWPATH);
2534 
2535 	if (kmem_cache_debug(s) &&
2536 		!(n = free_debug_processing(s, page, x, addr, &flags)))
2537 		return;
2538 
2539 	do {
2540 		if (unlikely(n)) {
2541 			spin_unlock_irqrestore(&n->list_lock, flags);
2542 			n = NULL;
2543 		}
2544 		prior = page->freelist;
2545 		counters = page->counters;
2546 		set_freepointer(s, object, prior);
2547 		new.counters = counters;
2548 		was_frozen = new.frozen;
2549 		new.inuse--;
2550 		if ((!new.inuse || !prior) && !was_frozen) {
2551 
2552 			if (kmem_cache_has_cpu_partial(s) && !prior) {
2553 
2554 				/*
2555 				 * Slab was on no list before and will be
2556 				 * partially empty
2557 				 * We can defer the list move and instead
2558 				 * freeze it.
2559 				 */
2560 				new.frozen = 1;
2561 
2562 			} else { /* Needs to be taken off a list */
2563 
2564 				n = get_node(s, page_to_nid(page));
2565 				/*
2566 				 * Speculatively acquire the list_lock.
2567 				 * If the cmpxchg does not succeed then we may
2568 				 * drop the list_lock without any processing.
2569 				 *
2570 				 * Otherwise the list_lock will synchronize with
2571 				 * other processors updating the list of slabs.
2572 				 */
2573 				spin_lock_irqsave(&n->list_lock, flags);
2574 
2575 			}
2576 		}
2577 
2578 	} while (!cmpxchg_double_slab(s, page,
2579 		prior, counters,
2580 		object, new.counters,
2581 		"__slab_free"));
2582 
2583 	if (likely(!n)) {
2584 
2585 		/*
2586 		 * If we just froze the page then put it onto the
2587 		 * per cpu partial list.
2588 		 */
2589 		if (new.frozen && !was_frozen) {
2590 			put_cpu_partial(s, page, 1);
2591 			stat(s, CPU_PARTIAL_FREE);
2592 		}
2593 		/*
2594 		 * The list lock was not taken therefore no list
2595 		 * activity can be necessary.
2596 		 */
2597 		if (was_frozen)
2598 			stat(s, FREE_FROZEN);
2599 		return;
2600 	}
2601 
2602 	if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2603 		goto slab_empty;
2604 
2605 	/*
2606 	 * Objects left in the slab. If it was not on the partial list before
2607 	 * then add it.
2608 	 */
2609 	if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2610 		if (kmem_cache_debug(s))
2611 			remove_full(s, n, page);
2612 		add_partial(n, page, DEACTIVATE_TO_TAIL);
2613 		stat(s, FREE_ADD_PARTIAL);
2614 	}
2615 	spin_unlock_irqrestore(&n->list_lock, flags);
2616 	return;
2617 
2618 slab_empty:
2619 	if (prior) {
2620 		/*
2621 		 * Slab on the partial list.
2622 		 */
2623 		remove_partial(n, page);
2624 		stat(s, FREE_REMOVE_PARTIAL);
2625 	} else {
2626 		/* Slab must be on the full list */
2627 		remove_full(s, n, page);
2628 	}
2629 
2630 	spin_unlock_irqrestore(&n->list_lock, flags);
2631 	stat(s, FREE_SLAB);
2632 	discard_slab(s, page);
2633 }
2634 
2635 /*
2636  * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2637  * can perform fastpath freeing without additional function calls.
2638  *
2639  * The fastpath is only possible if we are freeing to the current cpu slab
2640  * of this processor. This typically the case if we have just allocated
2641  * the item before.
2642  *
2643  * If fastpath is not possible then fall back to __slab_free where we deal
2644  * with all sorts of special processing.
2645  */
2646 static __always_inline void slab_free(struct kmem_cache *s,
2647 			struct page *page, void *x, unsigned long addr)
2648 {
2649 	void **object = (void *)x;
2650 	struct kmem_cache_cpu *c;
2651 	unsigned long tid;
2652 
2653 	slab_free_hook(s, x);
2654 
2655 redo:
2656 	/*
2657 	 * Determine the currently cpus per cpu slab.
2658 	 * The cpu may change afterward. However that does not matter since
2659 	 * data is retrieved via this pointer. If we are on the same cpu
2660 	 * during the cmpxchg then the free will succedd.
2661 	 */
2662 	preempt_disable();
2663 	c = this_cpu_ptr(s->cpu_slab);
2664 
2665 	tid = c->tid;
2666 	preempt_enable();
2667 
2668 	if (likely(page == c->page)) {
2669 		set_freepointer(s, object, c->freelist);
2670 
2671 		if (unlikely(!this_cpu_cmpxchg_double(
2672 				s->cpu_slab->freelist, s->cpu_slab->tid,
2673 				c->freelist, tid,
2674 				object, next_tid(tid)))) {
2675 
2676 			note_cmpxchg_failure("slab_free", s, tid);
2677 			goto redo;
2678 		}
2679 		stat(s, FREE_FASTPATH);
2680 	} else
2681 		__slab_free(s, page, x, addr);
2682 
2683 }
2684 
2685 void kmem_cache_free(struct kmem_cache *s, void *x)
2686 {
2687 	s = cache_from_obj(s, x);
2688 	if (!s)
2689 		return;
2690 	slab_free(s, virt_to_head_page(x), x, _RET_IP_);
2691 	trace_kmem_cache_free(_RET_IP_, x);
2692 }
2693 EXPORT_SYMBOL(kmem_cache_free);
2694 
2695 /*
2696  * Object placement in a slab is made very easy because we always start at
2697  * offset 0. If we tune the size of the object to the alignment then we can
2698  * get the required alignment by putting one properly sized object after
2699  * another.
2700  *
2701  * Notice that the allocation order determines the sizes of the per cpu
2702  * caches. Each processor has always one slab available for allocations.
2703  * Increasing the allocation order reduces the number of times that slabs
2704  * must be moved on and off the partial lists and is therefore a factor in
2705  * locking overhead.
2706  */
2707 
2708 /*
2709  * Mininum / Maximum order of slab pages. This influences locking overhead
2710  * and slab fragmentation. A higher order reduces the number of partial slabs
2711  * and increases the number of allocations possible without having to
2712  * take the list_lock.
2713  */
2714 static int slub_min_order;
2715 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2716 static int slub_min_objects;
2717 
2718 /*
2719  * Calculate the order of allocation given an slab object size.
2720  *
2721  * The order of allocation has significant impact on performance and other
2722  * system components. Generally order 0 allocations should be preferred since
2723  * order 0 does not cause fragmentation in the page allocator. Larger objects
2724  * be problematic to put into order 0 slabs because there may be too much
2725  * unused space left. We go to a higher order if more than 1/16th of the slab
2726  * would be wasted.
2727  *
2728  * In order to reach satisfactory performance we must ensure that a minimum
2729  * number of objects is in one slab. Otherwise we may generate too much
2730  * activity on the partial lists which requires taking the list_lock. This is
2731  * less a concern for large slabs though which are rarely used.
2732  *
2733  * slub_max_order specifies the order where we begin to stop considering the
2734  * number of objects in a slab as critical. If we reach slub_max_order then
2735  * we try to keep the page order as low as possible. So we accept more waste
2736  * of space in favor of a small page order.
2737  *
2738  * Higher order allocations also allow the placement of more objects in a
2739  * slab and thereby reduce object handling overhead. If the user has
2740  * requested a higher mininum order then we start with that one instead of
2741  * the smallest order which will fit the object.
2742  */
2743 static inline int slab_order(int size, int min_objects,
2744 				int max_order, int fract_leftover, int reserved)
2745 {
2746 	int order;
2747 	int rem;
2748 	int min_order = slub_min_order;
2749 
2750 	if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2751 		return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2752 
2753 	for (order = max(min_order,
2754 				fls(min_objects * size - 1) - PAGE_SHIFT);
2755 			order <= max_order; order++) {
2756 
2757 		unsigned long slab_size = PAGE_SIZE << order;
2758 
2759 		if (slab_size < min_objects * size + reserved)
2760 			continue;
2761 
2762 		rem = (slab_size - reserved) % size;
2763 
2764 		if (rem <= slab_size / fract_leftover)
2765 			break;
2766 
2767 	}
2768 
2769 	return order;
2770 }
2771 
2772 static inline int calculate_order(int size, int reserved)
2773 {
2774 	int order;
2775 	int min_objects;
2776 	int fraction;
2777 	int max_objects;
2778 
2779 	/*
2780 	 * Attempt to find best configuration for a slab. This
2781 	 * works by first attempting to generate a layout with
2782 	 * the best configuration and backing off gradually.
2783 	 *
2784 	 * First we reduce the acceptable waste in a slab. Then
2785 	 * we reduce the minimum objects required in a slab.
2786 	 */
2787 	min_objects = slub_min_objects;
2788 	if (!min_objects)
2789 		min_objects = 4 * (fls(nr_cpu_ids) + 1);
2790 	max_objects = order_objects(slub_max_order, size, reserved);
2791 	min_objects = min(min_objects, max_objects);
2792 
2793 	while (min_objects > 1) {
2794 		fraction = 16;
2795 		while (fraction >= 4) {
2796 			order = slab_order(size, min_objects,
2797 					slub_max_order, fraction, reserved);
2798 			if (order <= slub_max_order)
2799 				return order;
2800 			fraction /= 2;
2801 		}
2802 		min_objects--;
2803 	}
2804 
2805 	/*
2806 	 * We were unable to place multiple objects in a slab. Now
2807 	 * lets see if we can place a single object there.
2808 	 */
2809 	order = slab_order(size, 1, slub_max_order, 1, reserved);
2810 	if (order <= slub_max_order)
2811 		return order;
2812 
2813 	/*
2814 	 * Doh this slab cannot be placed using slub_max_order.
2815 	 */
2816 	order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2817 	if (order < MAX_ORDER)
2818 		return order;
2819 	return -ENOSYS;
2820 }
2821 
2822 static void
2823 init_kmem_cache_node(struct kmem_cache_node *n)
2824 {
2825 	n->nr_partial = 0;
2826 	spin_lock_init(&n->list_lock);
2827 	INIT_LIST_HEAD(&n->partial);
2828 #ifdef CONFIG_SLUB_DEBUG
2829 	atomic_long_set(&n->nr_slabs, 0);
2830 	atomic_long_set(&n->total_objects, 0);
2831 	INIT_LIST_HEAD(&n->full);
2832 #endif
2833 }
2834 
2835 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2836 {
2837 	BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2838 			KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
2839 
2840 	/*
2841 	 * Must align to double word boundary for the double cmpxchg
2842 	 * instructions to work; see __pcpu_double_call_return_bool().
2843 	 */
2844 	s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2845 				     2 * sizeof(void *));
2846 
2847 	if (!s->cpu_slab)
2848 		return 0;
2849 
2850 	init_kmem_cache_cpus(s);
2851 
2852 	return 1;
2853 }
2854 
2855 static struct kmem_cache *kmem_cache_node;
2856 
2857 /*
2858  * No kmalloc_node yet so do it by hand. We know that this is the first
2859  * slab on the node for this slabcache. There are no concurrent accesses
2860  * possible.
2861  *
2862  * Note that this function only works on the kmem_cache_node
2863  * when allocating for the kmem_cache_node. This is used for bootstrapping
2864  * memory on a fresh node that has no slab structures yet.
2865  */
2866 static void early_kmem_cache_node_alloc(int node)
2867 {
2868 	struct page *page;
2869 	struct kmem_cache_node *n;
2870 
2871 	BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2872 
2873 	page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2874 
2875 	BUG_ON(!page);
2876 	if (page_to_nid(page) != node) {
2877 		pr_err("SLUB: Unable to allocate memory from node %d\n", node);
2878 		pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
2879 	}
2880 
2881 	n = page->freelist;
2882 	BUG_ON(!n);
2883 	page->freelist = get_freepointer(kmem_cache_node, n);
2884 	page->inuse = 1;
2885 	page->frozen = 0;
2886 	kmem_cache_node->node[node] = n;
2887 #ifdef CONFIG_SLUB_DEBUG
2888 	init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2889 	init_tracking(kmem_cache_node, n);
2890 #endif
2891 	init_kmem_cache_node(n);
2892 	inc_slabs_node(kmem_cache_node, node, page->objects);
2893 
2894 	/*
2895 	 * No locks need to be taken here as it has just been
2896 	 * initialized and there is no concurrent access.
2897 	 */
2898 	__add_partial(n, page, DEACTIVATE_TO_HEAD);
2899 }
2900 
2901 static void free_kmem_cache_nodes(struct kmem_cache *s)
2902 {
2903 	int node;
2904 	struct kmem_cache_node *n;
2905 
2906 	for_each_kmem_cache_node(s, node, n) {
2907 		kmem_cache_free(kmem_cache_node, n);
2908 		s->node[node] = NULL;
2909 	}
2910 }
2911 
2912 static int init_kmem_cache_nodes(struct kmem_cache *s)
2913 {
2914 	int node;
2915 
2916 	for_each_node_state(node, N_NORMAL_MEMORY) {
2917 		struct kmem_cache_node *n;
2918 
2919 		if (slab_state == DOWN) {
2920 			early_kmem_cache_node_alloc(node);
2921 			continue;
2922 		}
2923 		n = kmem_cache_alloc_node(kmem_cache_node,
2924 						GFP_KERNEL, node);
2925 
2926 		if (!n) {
2927 			free_kmem_cache_nodes(s);
2928 			return 0;
2929 		}
2930 
2931 		s->node[node] = n;
2932 		init_kmem_cache_node(n);
2933 	}
2934 	return 1;
2935 }
2936 
2937 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2938 {
2939 	if (min < MIN_PARTIAL)
2940 		min = MIN_PARTIAL;
2941 	else if (min > MAX_PARTIAL)
2942 		min = MAX_PARTIAL;
2943 	s->min_partial = min;
2944 }
2945 
2946 /*
2947  * calculate_sizes() determines the order and the distribution of data within
2948  * a slab object.
2949  */
2950 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2951 {
2952 	unsigned long flags = s->flags;
2953 	unsigned long size = s->object_size;
2954 	int order;
2955 
2956 	/*
2957 	 * Round up object size to the next word boundary. We can only
2958 	 * place the free pointer at word boundaries and this determines
2959 	 * the possible location of the free pointer.
2960 	 */
2961 	size = ALIGN(size, sizeof(void *));
2962 
2963 #ifdef CONFIG_SLUB_DEBUG
2964 	/*
2965 	 * Determine if we can poison the object itself. If the user of
2966 	 * the slab may touch the object after free or before allocation
2967 	 * then we should never poison the object itself.
2968 	 */
2969 	if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2970 			!s->ctor)
2971 		s->flags |= __OBJECT_POISON;
2972 	else
2973 		s->flags &= ~__OBJECT_POISON;
2974 
2975 
2976 	/*
2977 	 * If we are Redzoning then check if there is some space between the
2978 	 * end of the object and the free pointer. If not then add an
2979 	 * additional word to have some bytes to store Redzone information.
2980 	 */
2981 	if ((flags & SLAB_RED_ZONE) && size == s->object_size)
2982 		size += sizeof(void *);
2983 #endif
2984 
2985 	/*
2986 	 * With that we have determined the number of bytes in actual use
2987 	 * by the object. This is the potential offset to the free pointer.
2988 	 */
2989 	s->inuse = size;
2990 
2991 	if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2992 		s->ctor)) {
2993 		/*
2994 		 * Relocate free pointer after the object if it is not
2995 		 * permitted to overwrite the first word of the object on
2996 		 * kmem_cache_free.
2997 		 *
2998 		 * This is the case if we do RCU, have a constructor or
2999 		 * destructor or are poisoning the objects.
3000 		 */
3001 		s->offset = size;
3002 		size += sizeof(void *);
3003 	}
3004 
3005 #ifdef CONFIG_SLUB_DEBUG
3006 	if (flags & SLAB_STORE_USER)
3007 		/*
3008 		 * Need to store information about allocs and frees after
3009 		 * the object.
3010 		 */
3011 		size += 2 * sizeof(struct track);
3012 
3013 	if (flags & SLAB_RED_ZONE)
3014 		/*
3015 		 * Add some empty padding so that we can catch
3016 		 * overwrites from earlier objects rather than let
3017 		 * tracking information or the free pointer be
3018 		 * corrupted if a user writes before the start
3019 		 * of the object.
3020 		 */
3021 		size += sizeof(void *);
3022 #endif
3023 
3024 	/*
3025 	 * SLUB stores one object immediately after another beginning from
3026 	 * offset 0. In order to align the objects we have to simply size
3027 	 * each object to conform to the alignment.
3028 	 */
3029 	size = ALIGN(size, s->align);
3030 	s->size = size;
3031 	if (forced_order >= 0)
3032 		order = forced_order;
3033 	else
3034 		order = calculate_order(size, s->reserved);
3035 
3036 	if (order < 0)
3037 		return 0;
3038 
3039 	s->allocflags = 0;
3040 	if (order)
3041 		s->allocflags |= __GFP_COMP;
3042 
3043 	if (s->flags & SLAB_CACHE_DMA)
3044 		s->allocflags |= GFP_DMA;
3045 
3046 	if (s->flags & SLAB_RECLAIM_ACCOUNT)
3047 		s->allocflags |= __GFP_RECLAIMABLE;
3048 
3049 	/*
3050 	 * Determine the number of objects per slab
3051 	 */
3052 	s->oo = oo_make(order, size, s->reserved);
3053 	s->min = oo_make(get_order(size), size, s->reserved);
3054 	if (oo_objects(s->oo) > oo_objects(s->max))
3055 		s->max = s->oo;
3056 
3057 	return !!oo_objects(s->oo);
3058 }
3059 
3060 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3061 {
3062 	s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3063 	s->reserved = 0;
3064 
3065 	if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3066 		s->reserved = sizeof(struct rcu_head);
3067 
3068 	if (!calculate_sizes(s, -1))
3069 		goto error;
3070 	if (disable_higher_order_debug) {
3071 		/*
3072 		 * Disable debugging flags that store metadata if the min slab
3073 		 * order increased.
3074 		 */
3075 		if (get_order(s->size) > get_order(s->object_size)) {
3076 			s->flags &= ~DEBUG_METADATA_FLAGS;
3077 			s->offset = 0;
3078 			if (!calculate_sizes(s, -1))
3079 				goto error;
3080 		}
3081 	}
3082 
3083 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3084     defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3085 	if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3086 		/* Enable fast mode */
3087 		s->flags |= __CMPXCHG_DOUBLE;
3088 #endif
3089 
3090 	/*
3091 	 * The larger the object size is, the more pages we want on the partial
3092 	 * list to avoid pounding the page allocator excessively.
3093 	 */
3094 	set_min_partial(s, ilog2(s->size) / 2);
3095 
3096 	/*
3097 	 * cpu_partial determined the maximum number of objects kept in the
3098 	 * per cpu partial lists of a processor.
3099 	 *
3100 	 * Per cpu partial lists mainly contain slabs that just have one
3101 	 * object freed. If they are used for allocation then they can be
3102 	 * filled up again with minimal effort. The slab will never hit the
3103 	 * per node partial lists and therefore no locking will be required.
3104 	 *
3105 	 * This setting also determines
3106 	 *
3107 	 * A) The number of objects from per cpu partial slabs dumped to the
3108 	 *    per node list when we reach the limit.
3109 	 * B) The number of objects in cpu partial slabs to extract from the
3110 	 *    per node list when we run out of per cpu objects. We only fetch
3111 	 *    50% to keep some capacity around for frees.
3112 	 */
3113 	if (!kmem_cache_has_cpu_partial(s))
3114 		s->cpu_partial = 0;
3115 	else if (s->size >= PAGE_SIZE)
3116 		s->cpu_partial = 2;
3117 	else if (s->size >= 1024)
3118 		s->cpu_partial = 6;
3119 	else if (s->size >= 256)
3120 		s->cpu_partial = 13;
3121 	else
3122 		s->cpu_partial = 30;
3123 
3124 #ifdef CONFIG_NUMA
3125 	s->remote_node_defrag_ratio = 1000;
3126 #endif
3127 	if (!init_kmem_cache_nodes(s))
3128 		goto error;
3129 
3130 	if (alloc_kmem_cache_cpus(s))
3131 		return 0;
3132 
3133 	free_kmem_cache_nodes(s);
3134 error:
3135 	if (flags & SLAB_PANIC)
3136 		panic("Cannot create slab %s size=%lu realsize=%u "
3137 			"order=%u offset=%u flags=%lx\n",
3138 			s->name, (unsigned long)s->size, s->size,
3139 			oo_order(s->oo), s->offset, flags);
3140 	return -EINVAL;
3141 }
3142 
3143 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3144 							const char *text)
3145 {
3146 #ifdef CONFIG_SLUB_DEBUG
3147 	void *addr = page_address(page);
3148 	void *p;
3149 	unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3150 				     sizeof(long), GFP_ATOMIC);
3151 	if (!map)
3152 		return;
3153 	slab_err(s, page, text, s->name);
3154 	slab_lock(page);
3155 
3156 	get_map(s, page, map);
3157 	for_each_object(p, s, addr, page->objects) {
3158 
3159 		if (!test_bit(slab_index(p, s, addr), map)) {
3160 			pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3161 			print_tracking(s, p);
3162 		}
3163 	}
3164 	slab_unlock(page);
3165 	kfree(map);
3166 #endif
3167 }
3168 
3169 /*
3170  * Attempt to free all partial slabs on a node.
3171  * This is called from kmem_cache_close(). We must be the last thread
3172  * using the cache and therefore we do not need to lock anymore.
3173  */
3174 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3175 {
3176 	struct page *page, *h;
3177 
3178 	list_for_each_entry_safe(page, h, &n->partial, lru) {
3179 		if (!page->inuse) {
3180 			__remove_partial(n, page);
3181 			discard_slab(s, page);
3182 		} else {
3183 			list_slab_objects(s, page,
3184 			"Objects remaining in %s on kmem_cache_close()");
3185 		}
3186 	}
3187 }
3188 
3189 /*
3190  * Release all resources used by a slab cache.
3191  */
3192 static inline int kmem_cache_close(struct kmem_cache *s)
3193 {
3194 	int node;
3195 	struct kmem_cache_node *n;
3196 
3197 	flush_all(s);
3198 	/* Attempt to free all objects */
3199 	for_each_kmem_cache_node(s, node, n) {
3200 		free_partial(s, n);
3201 		if (n->nr_partial || slabs_node(s, node))
3202 			return 1;
3203 	}
3204 	free_percpu(s->cpu_slab);
3205 	free_kmem_cache_nodes(s);
3206 	return 0;
3207 }
3208 
3209 int __kmem_cache_shutdown(struct kmem_cache *s)
3210 {
3211 	return kmem_cache_close(s);
3212 }
3213 
3214 /********************************************************************
3215  *		Kmalloc subsystem
3216  *******************************************************************/
3217 
3218 static int __init setup_slub_min_order(char *str)
3219 {
3220 	get_option(&str, &slub_min_order);
3221 
3222 	return 1;
3223 }
3224 
3225 __setup("slub_min_order=", setup_slub_min_order);
3226 
3227 static int __init setup_slub_max_order(char *str)
3228 {
3229 	get_option(&str, &slub_max_order);
3230 	slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3231 
3232 	return 1;
3233 }
3234 
3235 __setup("slub_max_order=", setup_slub_max_order);
3236 
3237 static int __init setup_slub_min_objects(char *str)
3238 {
3239 	get_option(&str, &slub_min_objects);
3240 
3241 	return 1;
3242 }
3243 
3244 __setup("slub_min_objects=", setup_slub_min_objects);
3245 
3246 void *__kmalloc(size_t size, gfp_t flags)
3247 {
3248 	struct kmem_cache *s;
3249 	void *ret;
3250 
3251 	if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3252 		return kmalloc_large(size, flags);
3253 
3254 	s = kmalloc_slab(size, flags);
3255 
3256 	if (unlikely(ZERO_OR_NULL_PTR(s)))
3257 		return s;
3258 
3259 	ret = slab_alloc(s, flags, _RET_IP_);
3260 
3261 	trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3262 
3263 	return ret;
3264 }
3265 EXPORT_SYMBOL(__kmalloc);
3266 
3267 #ifdef CONFIG_NUMA
3268 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3269 {
3270 	struct page *page;
3271 	void *ptr = NULL;
3272 
3273 	flags |= __GFP_COMP | __GFP_NOTRACK;
3274 	page = alloc_kmem_pages_node(node, flags, get_order(size));
3275 	if (page)
3276 		ptr = page_address(page);
3277 
3278 	kmalloc_large_node_hook(ptr, size, flags);
3279 	return ptr;
3280 }
3281 
3282 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3283 {
3284 	struct kmem_cache *s;
3285 	void *ret;
3286 
3287 	if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3288 		ret = kmalloc_large_node(size, flags, node);
3289 
3290 		trace_kmalloc_node(_RET_IP_, ret,
3291 				   size, PAGE_SIZE << get_order(size),
3292 				   flags, node);
3293 
3294 		return ret;
3295 	}
3296 
3297 	s = kmalloc_slab(size, flags);
3298 
3299 	if (unlikely(ZERO_OR_NULL_PTR(s)))
3300 		return s;
3301 
3302 	ret = slab_alloc_node(s, flags, node, _RET_IP_);
3303 
3304 	trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3305 
3306 	return ret;
3307 }
3308 EXPORT_SYMBOL(__kmalloc_node);
3309 #endif
3310 
3311 size_t ksize(const void *object)
3312 {
3313 	struct page *page;
3314 
3315 	if (unlikely(object == ZERO_SIZE_PTR))
3316 		return 0;
3317 
3318 	page = virt_to_head_page(object);
3319 
3320 	if (unlikely(!PageSlab(page))) {
3321 		WARN_ON(!PageCompound(page));
3322 		return PAGE_SIZE << compound_order(page);
3323 	}
3324 
3325 	return slab_ksize(page->slab_cache);
3326 }
3327 EXPORT_SYMBOL(ksize);
3328 
3329 void kfree(const void *x)
3330 {
3331 	struct page *page;
3332 	void *object = (void *)x;
3333 
3334 	trace_kfree(_RET_IP_, x);
3335 
3336 	if (unlikely(ZERO_OR_NULL_PTR(x)))
3337 		return;
3338 
3339 	page = virt_to_head_page(x);
3340 	if (unlikely(!PageSlab(page))) {
3341 		BUG_ON(!PageCompound(page));
3342 		kfree_hook(x);
3343 		__free_kmem_pages(page, compound_order(page));
3344 		return;
3345 	}
3346 	slab_free(page->slab_cache, page, object, _RET_IP_);
3347 }
3348 EXPORT_SYMBOL(kfree);
3349 
3350 /*
3351  * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3352  * the remaining slabs by the number of items in use. The slabs with the
3353  * most items in use come first. New allocations will then fill those up
3354  * and thus they can be removed from the partial lists.
3355  *
3356  * The slabs with the least items are placed last. This results in them
3357  * being allocated from last increasing the chance that the last objects
3358  * are freed in them.
3359  */
3360 int __kmem_cache_shrink(struct kmem_cache *s)
3361 {
3362 	int node;
3363 	int i;
3364 	struct kmem_cache_node *n;
3365 	struct page *page;
3366 	struct page *t;
3367 	int objects = oo_objects(s->max);
3368 	struct list_head *slabs_by_inuse =
3369 		kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3370 	unsigned long flags;
3371 
3372 	if (!slabs_by_inuse)
3373 		return -ENOMEM;
3374 
3375 	flush_all(s);
3376 	for_each_kmem_cache_node(s, node, n) {
3377 		if (!n->nr_partial)
3378 			continue;
3379 
3380 		for (i = 0; i < objects; i++)
3381 			INIT_LIST_HEAD(slabs_by_inuse + i);
3382 
3383 		spin_lock_irqsave(&n->list_lock, flags);
3384 
3385 		/*
3386 		 * Build lists indexed by the items in use in each slab.
3387 		 *
3388 		 * Note that concurrent frees may occur while we hold the
3389 		 * list_lock. page->inuse here is the upper limit.
3390 		 */
3391 		list_for_each_entry_safe(page, t, &n->partial, lru) {
3392 			list_move(&page->lru, slabs_by_inuse + page->inuse);
3393 			if (!page->inuse)
3394 				n->nr_partial--;
3395 		}
3396 
3397 		/*
3398 		 * Rebuild the partial list with the slabs filled up most
3399 		 * first and the least used slabs at the end.
3400 		 */
3401 		for (i = objects - 1; i > 0; i--)
3402 			list_splice(slabs_by_inuse + i, n->partial.prev);
3403 
3404 		spin_unlock_irqrestore(&n->list_lock, flags);
3405 
3406 		/* Release empty slabs */
3407 		list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3408 			discard_slab(s, page);
3409 	}
3410 
3411 	kfree(slabs_by_inuse);
3412 	return 0;
3413 }
3414 
3415 static int slab_mem_going_offline_callback(void *arg)
3416 {
3417 	struct kmem_cache *s;
3418 
3419 	mutex_lock(&slab_mutex);
3420 	list_for_each_entry(s, &slab_caches, list)
3421 		__kmem_cache_shrink(s);
3422 	mutex_unlock(&slab_mutex);
3423 
3424 	return 0;
3425 }
3426 
3427 static void slab_mem_offline_callback(void *arg)
3428 {
3429 	struct kmem_cache_node *n;
3430 	struct kmem_cache *s;
3431 	struct memory_notify *marg = arg;
3432 	int offline_node;
3433 
3434 	offline_node = marg->status_change_nid_normal;
3435 
3436 	/*
3437 	 * If the node still has available memory. we need kmem_cache_node
3438 	 * for it yet.
3439 	 */
3440 	if (offline_node < 0)
3441 		return;
3442 
3443 	mutex_lock(&slab_mutex);
3444 	list_for_each_entry(s, &slab_caches, list) {
3445 		n = get_node(s, offline_node);
3446 		if (n) {
3447 			/*
3448 			 * if n->nr_slabs > 0, slabs still exist on the node
3449 			 * that is going down. We were unable to free them,
3450 			 * and offline_pages() function shouldn't call this
3451 			 * callback. So, we must fail.
3452 			 */
3453 			BUG_ON(slabs_node(s, offline_node));
3454 
3455 			s->node[offline_node] = NULL;
3456 			kmem_cache_free(kmem_cache_node, n);
3457 		}
3458 	}
3459 	mutex_unlock(&slab_mutex);
3460 }
3461 
3462 static int slab_mem_going_online_callback(void *arg)
3463 {
3464 	struct kmem_cache_node *n;
3465 	struct kmem_cache *s;
3466 	struct memory_notify *marg = arg;
3467 	int nid = marg->status_change_nid_normal;
3468 	int ret = 0;
3469 
3470 	/*
3471 	 * If the node's memory is already available, then kmem_cache_node is
3472 	 * already created. Nothing to do.
3473 	 */
3474 	if (nid < 0)
3475 		return 0;
3476 
3477 	/*
3478 	 * We are bringing a node online. No memory is available yet. We must
3479 	 * allocate a kmem_cache_node structure in order to bring the node
3480 	 * online.
3481 	 */
3482 	mutex_lock(&slab_mutex);
3483 	list_for_each_entry(s, &slab_caches, list) {
3484 		/*
3485 		 * XXX: kmem_cache_alloc_node will fallback to other nodes
3486 		 *      since memory is not yet available from the node that
3487 		 *      is brought up.
3488 		 */
3489 		n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3490 		if (!n) {
3491 			ret = -ENOMEM;
3492 			goto out;
3493 		}
3494 		init_kmem_cache_node(n);
3495 		s->node[nid] = n;
3496 	}
3497 out:
3498 	mutex_unlock(&slab_mutex);
3499 	return ret;
3500 }
3501 
3502 static int slab_memory_callback(struct notifier_block *self,
3503 				unsigned long action, void *arg)
3504 {
3505 	int ret = 0;
3506 
3507 	switch (action) {
3508 	case MEM_GOING_ONLINE:
3509 		ret = slab_mem_going_online_callback(arg);
3510 		break;
3511 	case MEM_GOING_OFFLINE:
3512 		ret = slab_mem_going_offline_callback(arg);
3513 		break;
3514 	case MEM_OFFLINE:
3515 	case MEM_CANCEL_ONLINE:
3516 		slab_mem_offline_callback(arg);
3517 		break;
3518 	case MEM_ONLINE:
3519 	case MEM_CANCEL_OFFLINE:
3520 		break;
3521 	}
3522 	if (ret)
3523 		ret = notifier_from_errno(ret);
3524 	else
3525 		ret = NOTIFY_OK;
3526 	return ret;
3527 }
3528 
3529 static struct notifier_block slab_memory_callback_nb = {
3530 	.notifier_call = slab_memory_callback,
3531 	.priority = SLAB_CALLBACK_PRI,
3532 };
3533 
3534 /********************************************************************
3535  *			Basic setup of slabs
3536  *******************************************************************/
3537 
3538 /*
3539  * Used for early kmem_cache structures that were allocated using
3540  * the page allocator. Allocate them properly then fix up the pointers
3541  * that may be pointing to the wrong kmem_cache structure.
3542  */
3543 
3544 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3545 {
3546 	int node;
3547 	struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3548 	struct kmem_cache_node *n;
3549 
3550 	memcpy(s, static_cache, kmem_cache->object_size);
3551 
3552 	/*
3553 	 * This runs very early, and only the boot processor is supposed to be
3554 	 * up.  Even if it weren't true, IRQs are not up so we couldn't fire
3555 	 * IPIs around.
3556 	 */
3557 	__flush_cpu_slab(s, smp_processor_id());
3558 	for_each_kmem_cache_node(s, node, n) {
3559 		struct page *p;
3560 
3561 		list_for_each_entry(p, &n->partial, lru)
3562 			p->slab_cache = s;
3563 
3564 #ifdef CONFIG_SLUB_DEBUG
3565 		list_for_each_entry(p, &n->full, lru)
3566 			p->slab_cache = s;
3567 #endif
3568 	}
3569 	list_add(&s->list, &slab_caches);
3570 	return s;
3571 }
3572 
3573 void __init kmem_cache_init(void)
3574 {
3575 	static __initdata struct kmem_cache boot_kmem_cache,
3576 		boot_kmem_cache_node;
3577 
3578 	if (debug_guardpage_minorder())
3579 		slub_max_order = 0;
3580 
3581 	kmem_cache_node = &boot_kmem_cache_node;
3582 	kmem_cache = &boot_kmem_cache;
3583 
3584 	create_boot_cache(kmem_cache_node, "kmem_cache_node",
3585 		sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3586 
3587 	register_hotmemory_notifier(&slab_memory_callback_nb);
3588 
3589 	/* Able to allocate the per node structures */
3590 	slab_state = PARTIAL;
3591 
3592 	create_boot_cache(kmem_cache, "kmem_cache",
3593 			offsetof(struct kmem_cache, node) +
3594 				nr_node_ids * sizeof(struct kmem_cache_node *),
3595 		       SLAB_HWCACHE_ALIGN);
3596 
3597 	kmem_cache = bootstrap(&boot_kmem_cache);
3598 
3599 	/*
3600 	 * Allocate kmem_cache_node properly from the kmem_cache slab.
3601 	 * kmem_cache_node is separately allocated so no need to
3602 	 * update any list pointers.
3603 	 */
3604 	kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3605 
3606 	/* Now we can use the kmem_cache to allocate kmalloc slabs */
3607 	create_kmalloc_caches(0);
3608 
3609 #ifdef CONFIG_SMP
3610 	register_cpu_notifier(&slab_notifier);
3611 #endif
3612 
3613 	pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
3614 		cache_line_size(),
3615 		slub_min_order, slub_max_order, slub_min_objects,
3616 		nr_cpu_ids, nr_node_ids);
3617 }
3618 
3619 void __init kmem_cache_init_late(void)
3620 {
3621 }
3622 
3623 struct kmem_cache *
3624 __kmem_cache_alias(const char *name, size_t size, size_t align,
3625 		   unsigned long flags, void (*ctor)(void *))
3626 {
3627 	struct kmem_cache *s;
3628 
3629 	s = find_mergeable(size, align, flags, name, ctor);
3630 	if (s) {
3631 		int i;
3632 		struct kmem_cache *c;
3633 
3634 		s->refcount++;
3635 
3636 		/*
3637 		 * Adjust the object sizes so that we clear
3638 		 * the complete object on kzalloc.
3639 		 */
3640 		s->object_size = max(s->object_size, (int)size);
3641 		s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3642 
3643 		for_each_memcg_cache_index(i) {
3644 			c = cache_from_memcg_idx(s, i);
3645 			if (!c)
3646 				continue;
3647 			c->object_size = s->object_size;
3648 			c->inuse = max_t(int, c->inuse,
3649 					 ALIGN(size, sizeof(void *)));
3650 		}
3651 
3652 		if (sysfs_slab_alias(s, name)) {
3653 			s->refcount--;
3654 			s = NULL;
3655 		}
3656 	}
3657 
3658 	return s;
3659 }
3660 
3661 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3662 {
3663 	int err;
3664 
3665 	err = kmem_cache_open(s, flags);
3666 	if (err)
3667 		return err;
3668 
3669 	/* Mutex is not taken during early boot */
3670 	if (slab_state <= UP)
3671 		return 0;
3672 
3673 	memcg_propagate_slab_attrs(s);
3674 	err = sysfs_slab_add(s);
3675 	if (err)
3676 		kmem_cache_close(s);
3677 
3678 	return err;
3679 }
3680 
3681 #ifdef CONFIG_SMP
3682 /*
3683  * Use the cpu notifier to insure that the cpu slabs are flushed when
3684  * necessary.
3685  */
3686 static int slab_cpuup_callback(struct notifier_block *nfb,
3687 		unsigned long action, void *hcpu)
3688 {
3689 	long cpu = (long)hcpu;
3690 	struct kmem_cache *s;
3691 	unsigned long flags;
3692 
3693 	switch (action) {
3694 	case CPU_UP_CANCELED:
3695 	case CPU_UP_CANCELED_FROZEN:
3696 	case CPU_DEAD:
3697 	case CPU_DEAD_FROZEN:
3698 		mutex_lock(&slab_mutex);
3699 		list_for_each_entry(s, &slab_caches, list) {
3700 			local_irq_save(flags);
3701 			__flush_cpu_slab(s, cpu);
3702 			local_irq_restore(flags);
3703 		}
3704 		mutex_unlock(&slab_mutex);
3705 		break;
3706 	default:
3707 		break;
3708 	}
3709 	return NOTIFY_OK;
3710 }
3711 
3712 static struct notifier_block slab_notifier = {
3713 	.notifier_call = slab_cpuup_callback
3714 };
3715 
3716 #endif
3717 
3718 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3719 {
3720 	struct kmem_cache *s;
3721 	void *ret;
3722 
3723 	if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3724 		return kmalloc_large(size, gfpflags);
3725 
3726 	s = kmalloc_slab(size, gfpflags);
3727 
3728 	if (unlikely(ZERO_OR_NULL_PTR(s)))
3729 		return s;
3730 
3731 	ret = slab_alloc(s, gfpflags, caller);
3732 
3733 	/* Honor the call site pointer we received. */
3734 	trace_kmalloc(caller, ret, size, s->size, gfpflags);
3735 
3736 	return ret;
3737 }
3738 
3739 #ifdef CONFIG_NUMA
3740 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3741 					int node, unsigned long caller)
3742 {
3743 	struct kmem_cache *s;
3744 	void *ret;
3745 
3746 	if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3747 		ret = kmalloc_large_node(size, gfpflags, node);
3748 
3749 		trace_kmalloc_node(caller, ret,
3750 				   size, PAGE_SIZE << get_order(size),
3751 				   gfpflags, node);
3752 
3753 		return ret;
3754 	}
3755 
3756 	s = kmalloc_slab(size, gfpflags);
3757 
3758 	if (unlikely(ZERO_OR_NULL_PTR(s)))
3759 		return s;
3760 
3761 	ret = slab_alloc_node(s, gfpflags, node, caller);
3762 
3763 	/* Honor the call site pointer we received. */
3764 	trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3765 
3766 	return ret;
3767 }
3768 #endif
3769 
3770 #ifdef CONFIG_SYSFS
3771 static int count_inuse(struct page *page)
3772 {
3773 	return page->inuse;
3774 }
3775 
3776 static int count_total(struct page *page)
3777 {
3778 	return page->objects;
3779 }
3780 #endif
3781 
3782 #ifdef CONFIG_SLUB_DEBUG
3783 static int validate_slab(struct kmem_cache *s, struct page *page,
3784 						unsigned long *map)
3785 {
3786 	void *p;
3787 	void *addr = page_address(page);
3788 
3789 	if (!check_slab(s, page) ||
3790 			!on_freelist(s, page, NULL))
3791 		return 0;
3792 
3793 	/* Now we know that a valid freelist exists */
3794 	bitmap_zero(map, page->objects);
3795 
3796 	get_map(s, page, map);
3797 	for_each_object(p, s, addr, page->objects) {
3798 		if (test_bit(slab_index(p, s, addr), map))
3799 			if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3800 				return 0;
3801 	}
3802 
3803 	for_each_object(p, s, addr, page->objects)
3804 		if (!test_bit(slab_index(p, s, addr), map))
3805 			if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3806 				return 0;
3807 	return 1;
3808 }
3809 
3810 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3811 						unsigned long *map)
3812 {
3813 	slab_lock(page);
3814 	validate_slab(s, page, map);
3815 	slab_unlock(page);
3816 }
3817 
3818 static int validate_slab_node(struct kmem_cache *s,
3819 		struct kmem_cache_node *n, unsigned long *map)
3820 {
3821 	unsigned long count = 0;
3822 	struct page *page;
3823 	unsigned long flags;
3824 
3825 	spin_lock_irqsave(&n->list_lock, flags);
3826 
3827 	list_for_each_entry(page, &n->partial, lru) {
3828 		validate_slab_slab(s, page, map);
3829 		count++;
3830 	}
3831 	if (count != n->nr_partial)
3832 		pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
3833 		       s->name, count, n->nr_partial);
3834 
3835 	if (!(s->flags & SLAB_STORE_USER))
3836 		goto out;
3837 
3838 	list_for_each_entry(page, &n->full, lru) {
3839 		validate_slab_slab(s, page, map);
3840 		count++;
3841 	}
3842 	if (count != atomic_long_read(&n->nr_slabs))
3843 		pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
3844 		       s->name, count, atomic_long_read(&n->nr_slabs));
3845 
3846 out:
3847 	spin_unlock_irqrestore(&n->list_lock, flags);
3848 	return count;
3849 }
3850 
3851 static long validate_slab_cache(struct kmem_cache *s)
3852 {
3853 	int node;
3854 	unsigned long count = 0;
3855 	unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3856 				sizeof(unsigned long), GFP_KERNEL);
3857 	struct kmem_cache_node *n;
3858 
3859 	if (!map)
3860 		return -ENOMEM;
3861 
3862 	flush_all(s);
3863 	for_each_kmem_cache_node(s, node, n)
3864 		count += validate_slab_node(s, n, map);
3865 	kfree(map);
3866 	return count;
3867 }
3868 /*
3869  * Generate lists of code addresses where slabcache objects are allocated
3870  * and freed.
3871  */
3872 
3873 struct location {
3874 	unsigned long count;
3875 	unsigned long addr;
3876 	long long sum_time;
3877 	long min_time;
3878 	long max_time;
3879 	long min_pid;
3880 	long max_pid;
3881 	DECLARE_BITMAP(cpus, NR_CPUS);
3882 	nodemask_t nodes;
3883 };
3884 
3885 struct loc_track {
3886 	unsigned long max;
3887 	unsigned long count;
3888 	struct location *loc;
3889 };
3890 
3891 static void free_loc_track(struct loc_track *t)
3892 {
3893 	if (t->max)
3894 		free_pages((unsigned long)t->loc,
3895 			get_order(sizeof(struct location) * t->max));
3896 }
3897 
3898 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3899 {
3900 	struct location *l;
3901 	int order;
3902 
3903 	order = get_order(sizeof(struct location) * max);
3904 
3905 	l = (void *)__get_free_pages(flags, order);
3906 	if (!l)
3907 		return 0;
3908 
3909 	if (t->count) {
3910 		memcpy(l, t->loc, sizeof(struct location) * t->count);
3911 		free_loc_track(t);
3912 	}
3913 	t->max = max;
3914 	t->loc = l;
3915 	return 1;
3916 }
3917 
3918 static int add_location(struct loc_track *t, struct kmem_cache *s,
3919 				const struct track *track)
3920 {
3921 	long start, end, pos;
3922 	struct location *l;
3923 	unsigned long caddr;
3924 	unsigned long age = jiffies - track->when;
3925 
3926 	start = -1;
3927 	end = t->count;
3928 
3929 	for ( ; ; ) {
3930 		pos = start + (end - start + 1) / 2;
3931 
3932 		/*
3933 		 * There is nothing at "end". If we end up there
3934 		 * we need to add something to before end.
3935 		 */
3936 		if (pos == end)
3937 			break;
3938 
3939 		caddr = t->loc[pos].addr;
3940 		if (track->addr == caddr) {
3941 
3942 			l = &t->loc[pos];
3943 			l->count++;
3944 			if (track->when) {
3945 				l->sum_time += age;
3946 				if (age < l->min_time)
3947 					l->min_time = age;
3948 				if (age > l->max_time)
3949 					l->max_time = age;
3950 
3951 				if (track->pid < l->min_pid)
3952 					l->min_pid = track->pid;
3953 				if (track->pid > l->max_pid)
3954 					l->max_pid = track->pid;
3955 
3956 				cpumask_set_cpu(track->cpu,
3957 						to_cpumask(l->cpus));
3958 			}
3959 			node_set(page_to_nid(virt_to_page(track)), l->nodes);
3960 			return 1;
3961 		}
3962 
3963 		if (track->addr < caddr)
3964 			end = pos;
3965 		else
3966 			start = pos;
3967 	}
3968 
3969 	/*
3970 	 * Not found. Insert new tracking element.
3971 	 */
3972 	if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3973 		return 0;
3974 
3975 	l = t->loc + pos;
3976 	if (pos < t->count)
3977 		memmove(l + 1, l,
3978 			(t->count - pos) * sizeof(struct location));
3979 	t->count++;
3980 	l->count = 1;
3981 	l->addr = track->addr;
3982 	l->sum_time = age;
3983 	l->min_time = age;
3984 	l->max_time = age;
3985 	l->min_pid = track->pid;
3986 	l->max_pid = track->pid;
3987 	cpumask_clear(to_cpumask(l->cpus));
3988 	cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
3989 	nodes_clear(l->nodes);
3990 	node_set(page_to_nid(virt_to_page(track)), l->nodes);
3991 	return 1;
3992 }
3993 
3994 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3995 		struct page *page, enum track_item alloc,
3996 		unsigned long *map)
3997 {
3998 	void *addr = page_address(page);
3999 	void *p;
4000 
4001 	bitmap_zero(map, page->objects);
4002 	get_map(s, page, map);
4003 
4004 	for_each_object(p, s, addr, page->objects)
4005 		if (!test_bit(slab_index(p, s, addr), map))
4006 			add_location(t, s, get_track(s, p, alloc));
4007 }
4008 
4009 static int list_locations(struct kmem_cache *s, char *buf,
4010 					enum track_item alloc)
4011 {
4012 	int len = 0;
4013 	unsigned long i;
4014 	struct loc_track t = { 0, 0, NULL };
4015 	int node;
4016 	unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4017 				     sizeof(unsigned long), GFP_KERNEL);
4018 	struct kmem_cache_node *n;
4019 
4020 	if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4021 				     GFP_TEMPORARY)) {
4022 		kfree(map);
4023 		return sprintf(buf, "Out of memory\n");
4024 	}
4025 	/* Push back cpu slabs */
4026 	flush_all(s);
4027 
4028 	for_each_kmem_cache_node(s, node, n) {
4029 		unsigned long flags;
4030 		struct page *page;
4031 
4032 		if (!atomic_long_read(&n->nr_slabs))
4033 			continue;
4034 
4035 		spin_lock_irqsave(&n->list_lock, flags);
4036 		list_for_each_entry(page, &n->partial, lru)
4037 			process_slab(&t, s, page, alloc, map);
4038 		list_for_each_entry(page, &n->full, lru)
4039 			process_slab(&t, s, page, alloc, map);
4040 		spin_unlock_irqrestore(&n->list_lock, flags);
4041 	}
4042 
4043 	for (i = 0; i < t.count; i++) {
4044 		struct location *l = &t.loc[i];
4045 
4046 		if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4047 			break;
4048 		len += sprintf(buf + len, "%7ld ", l->count);
4049 
4050 		if (l->addr)
4051 			len += sprintf(buf + len, "%pS", (void *)l->addr);
4052 		else
4053 			len += sprintf(buf + len, "<not-available>");
4054 
4055 		if (l->sum_time != l->min_time) {
4056 			len += sprintf(buf + len, " age=%ld/%ld/%ld",
4057 				l->min_time,
4058 				(long)div_u64(l->sum_time, l->count),
4059 				l->max_time);
4060 		} else
4061 			len += sprintf(buf + len, " age=%ld",
4062 				l->min_time);
4063 
4064 		if (l->min_pid != l->max_pid)
4065 			len += sprintf(buf + len, " pid=%ld-%ld",
4066 				l->min_pid, l->max_pid);
4067 		else
4068 			len += sprintf(buf + len, " pid=%ld",
4069 				l->min_pid);
4070 
4071 		if (num_online_cpus() > 1 &&
4072 				!cpumask_empty(to_cpumask(l->cpus)) &&
4073 				len < PAGE_SIZE - 60) {
4074 			len += sprintf(buf + len, " cpus=");
4075 			len += cpulist_scnprintf(buf + len,
4076 						 PAGE_SIZE - len - 50,
4077 						 to_cpumask(l->cpus));
4078 		}
4079 
4080 		if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4081 				len < PAGE_SIZE - 60) {
4082 			len += sprintf(buf + len, " nodes=");
4083 			len += nodelist_scnprintf(buf + len,
4084 						  PAGE_SIZE - len - 50,
4085 						  l->nodes);
4086 		}
4087 
4088 		len += sprintf(buf + len, "\n");
4089 	}
4090 
4091 	free_loc_track(&t);
4092 	kfree(map);
4093 	if (!t.count)
4094 		len += sprintf(buf, "No data\n");
4095 	return len;
4096 }
4097 #endif
4098 
4099 #ifdef SLUB_RESILIENCY_TEST
4100 static void __init resiliency_test(void)
4101 {
4102 	u8 *p;
4103 
4104 	BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4105 
4106 	pr_err("SLUB resiliency testing\n");
4107 	pr_err("-----------------------\n");
4108 	pr_err("A. Corruption after allocation\n");
4109 
4110 	p = kzalloc(16, GFP_KERNEL);
4111 	p[16] = 0x12;
4112 	pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4113 	       p + 16);
4114 
4115 	validate_slab_cache(kmalloc_caches[4]);
4116 
4117 	/* Hmmm... The next two are dangerous */
4118 	p = kzalloc(32, GFP_KERNEL);
4119 	p[32 + sizeof(void *)] = 0x34;
4120 	pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4121 	       p);
4122 	pr_err("If allocated object is overwritten then not detectable\n\n");
4123 
4124 	validate_slab_cache(kmalloc_caches[5]);
4125 	p = kzalloc(64, GFP_KERNEL);
4126 	p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4127 	*p = 0x56;
4128 	pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4129 	       p);
4130 	pr_err("If allocated object is overwritten then not detectable\n\n");
4131 	validate_slab_cache(kmalloc_caches[6]);
4132 
4133 	pr_err("\nB. Corruption after free\n");
4134 	p = kzalloc(128, GFP_KERNEL);
4135 	kfree(p);
4136 	*p = 0x78;
4137 	pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4138 	validate_slab_cache(kmalloc_caches[7]);
4139 
4140 	p = kzalloc(256, GFP_KERNEL);
4141 	kfree(p);
4142 	p[50] = 0x9a;
4143 	pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4144 	validate_slab_cache(kmalloc_caches[8]);
4145 
4146 	p = kzalloc(512, GFP_KERNEL);
4147 	kfree(p);
4148 	p[512] = 0xab;
4149 	pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4150 	validate_slab_cache(kmalloc_caches[9]);
4151 }
4152 #else
4153 #ifdef CONFIG_SYSFS
4154 static void resiliency_test(void) {};
4155 #endif
4156 #endif
4157 
4158 #ifdef CONFIG_SYSFS
4159 enum slab_stat_type {
4160 	SL_ALL,			/* All slabs */
4161 	SL_PARTIAL,		/* Only partially allocated slabs */
4162 	SL_CPU,			/* Only slabs used for cpu caches */
4163 	SL_OBJECTS,		/* Determine allocated objects not slabs */
4164 	SL_TOTAL		/* Determine object capacity not slabs */
4165 };
4166 
4167 #define SO_ALL		(1 << SL_ALL)
4168 #define SO_PARTIAL	(1 << SL_PARTIAL)
4169 #define SO_CPU		(1 << SL_CPU)
4170 #define SO_OBJECTS	(1 << SL_OBJECTS)
4171 #define SO_TOTAL	(1 << SL_TOTAL)
4172 
4173 static ssize_t show_slab_objects(struct kmem_cache *s,
4174 			    char *buf, unsigned long flags)
4175 {
4176 	unsigned long total = 0;
4177 	int node;
4178 	int x;
4179 	unsigned long *nodes;
4180 
4181 	nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4182 	if (!nodes)
4183 		return -ENOMEM;
4184 
4185 	if (flags & SO_CPU) {
4186 		int cpu;
4187 
4188 		for_each_possible_cpu(cpu) {
4189 			struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4190 							       cpu);
4191 			int node;
4192 			struct page *page;
4193 
4194 			page = ACCESS_ONCE(c->page);
4195 			if (!page)
4196 				continue;
4197 
4198 			node = page_to_nid(page);
4199 			if (flags & SO_TOTAL)
4200 				x = page->objects;
4201 			else if (flags & SO_OBJECTS)
4202 				x = page->inuse;
4203 			else
4204 				x = 1;
4205 
4206 			total += x;
4207 			nodes[node] += x;
4208 
4209 			page = ACCESS_ONCE(c->partial);
4210 			if (page) {
4211 				node = page_to_nid(page);
4212 				if (flags & SO_TOTAL)
4213 					WARN_ON_ONCE(1);
4214 				else if (flags & SO_OBJECTS)
4215 					WARN_ON_ONCE(1);
4216 				else
4217 					x = page->pages;
4218 				total += x;
4219 				nodes[node] += x;
4220 			}
4221 		}
4222 	}
4223 
4224 	get_online_mems();
4225 #ifdef CONFIG_SLUB_DEBUG
4226 	if (flags & SO_ALL) {
4227 		struct kmem_cache_node *n;
4228 
4229 		for_each_kmem_cache_node(s, node, n) {
4230 
4231 			if (flags & SO_TOTAL)
4232 				x = atomic_long_read(&n->total_objects);
4233 			else if (flags & SO_OBJECTS)
4234 				x = atomic_long_read(&n->total_objects) -
4235 					count_partial(n, count_free);
4236 			else
4237 				x = atomic_long_read(&n->nr_slabs);
4238 			total += x;
4239 			nodes[node] += x;
4240 		}
4241 
4242 	} else
4243 #endif
4244 	if (flags & SO_PARTIAL) {
4245 		struct kmem_cache_node *n;
4246 
4247 		for_each_kmem_cache_node(s, node, n) {
4248 			if (flags & SO_TOTAL)
4249 				x = count_partial(n, count_total);
4250 			else if (flags & SO_OBJECTS)
4251 				x = count_partial(n, count_inuse);
4252 			else
4253 				x = n->nr_partial;
4254 			total += x;
4255 			nodes[node] += x;
4256 		}
4257 	}
4258 	x = sprintf(buf, "%lu", total);
4259 #ifdef CONFIG_NUMA
4260 	for (node = 0; node < nr_node_ids; node++)
4261 		if (nodes[node])
4262 			x += sprintf(buf + x, " N%d=%lu",
4263 					node, nodes[node]);
4264 #endif
4265 	put_online_mems();
4266 	kfree(nodes);
4267 	return x + sprintf(buf + x, "\n");
4268 }
4269 
4270 #ifdef CONFIG_SLUB_DEBUG
4271 static int any_slab_objects(struct kmem_cache *s)
4272 {
4273 	int node;
4274 	struct kmem_cache_node *n;
4275 
4276 	for_each_kmem_cache_node(s, node, n)
4277 		if (atomic_long_read(&n->total_objects))
4278 			return 1;
4279 
4280 	return 0;
4281 }
4282 #endif
4283 
4284 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4285 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4286 
4287 struct slab_attribute {
4288 	struct attribute attr;
4289 	ssize_t (*show)(struct kmem_cache *s, char *buf);
4290 	ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4291 };
4292 
4293 #define SLAB_ATTR_RO(_name) \
4294 	static struct slab_attribute _name##_attr = \
4295 	__ATTR(_name, 0400, _name##_show, NULL)
4296 
4297 #define SLAB_ATTR(_name) \
4298 	static struct slab_attribute _name##_attr =  \
4299 	__ATTR(_name, 0600, _name##_show, _name##_store)
4300 
4301 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4302 {
4303 	return sprintf(buf, "%d\n", s->size);
4304 }
4305 SLAB_ATTR_RO(slab_size);
4306 
4307 static ssize_t align_show(struct kmem_cache *s, char *buf)
4308 {
4309 	return sprintf(buf, "%d\n", s->align);
4310 }
4311 SLAB_ATTR_RO(align);
4312 
4313 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4314 {
4315 	return sprintf(buf, "%d\n", s->object_size);
4316 }
4317 SLAB_ATTR_RO(object_size);
4318 
4319 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4320 {
4321 	return sprintf(buf, "%d\n", oo_objects(s->oo));
4322 }
4323 SLAB_ATTR_RO(objs_per_slab);
4324 
4325 static ssize_t order_store(struct kmem_cache *s,
4326 				const char *buf, size_t length)
4327 {
4328 	unsigned long order;
4329 	int err;
4330 
4331 	err = kstrtoul(buf, 10, &order);
4332 	if (err)
4333 		return err;
4334 
4335 	if (order > slub_max_order || order < slub_min_order)
4336 		return -EINVAL;
4337 
4338 	calculate_sizes(s, order);
4339 	return length;
4340 }
4341 
4342 static ssize_t order_show(struct kmem_cache *s, char *buf)
4343 {
4344 	return sprintf(buf, "%d\n", oo_order(s->oo));
4345 }
4346 SLAB_ATTR(order);
4347 
4348 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4349 {
4350 	return sprintf(buf, "%lu\n", s->min_partial);
4351 }
4352 
4353 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4354 				 size_t length)
4355 {
4356 	unsigned long min;
4357 	int err;
4358 
4359 	err = kstrtoul(buf, 10, &min);
4360 	if (err)
4361 		return err;
4362 
4363 	set_min_partial(s, min);
4364 	return length;
4365 }
4366 SLAB_ATTR(min_partial);
4367 
4368 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4369 {
4370 	return sprintf(buf, "%u\n", s->cpu_partial);
4371 }
4372 
4373 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4374 				 size_t length)
4375 {
4376 	unsigned long objects;
4377 	int err;
4378 
4379 	err = kstrtoul(buf, 10, &objects);
4380 	if (err)
4381 		return err;
4382 	if (objects && !kmem_cache_has_cpu_partial(s))
4383 		return -EINVAL;
4384 
4385 	s->cpu_partial = objects;
4386 	flush_all(s);
4387 	return length;
4388 }
4389 SLAB_ATTR(cpu_partial);
4390 
4391 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4392 {
4393 	if (!s->ctor)
4394 		return 0;
4395 	return sprintf(buf, "%pS\n", s->ctor);
4396 }
4397 SLAB_ATTR_RO(ctor);
4398 
4399 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4400 {
4401 	return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
4402 }
4403 SLAB_ATTR_RO(aliases);
4404 
4405 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4406 {
4407 	return show_slab_objects(s, buf, SO_PARTIAL);
4408 }
4409 SLAB_ATTR_RO(partial);
4410 
4411 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4412 {
4413 	return show_slab_objects(s, buf, SO_CPU);
4414 }
4415 SLAB_ATTR_RO(cpu_slabs);
4416 
4417 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4418 {
4419 	return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4420 }
4421 SLAB_ATTR_RO(objects);
4422 
4423 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4424 {
4425 	return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4426 }
4427 SLAB_ATTR_RO(objects_partial);
4428 
4429 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4430 {
4431 	int objects = 0;
4432 	int pages = 0;
4433 	int cpu;
4434 	int len;
4435 
4436 	for_each_online_cpu(cpu) {
4437 		struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4438 
4439 		if (page) {
4440 			pages += page->pages;
4441 			objects += page->pobjects;
4442 		}
4443 	}
4444 
4445 	len = sprintf(buf, "%d(%d)", objects, pages);
4446 
4447 #ifdef CONFIG_SMP
4448 	for_each_online_cpu(cpu) {
4449 		struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4450 
4451 		if (page && len < PAGE_SIZE - 20)
4452 			len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4453 				page->pobjects, page->pages);
4454 	}
4455 #endif
4456 	return len + sprintf(buf + len, "\n");
4457 }
4458 SLAB_ATTR_RO(slabs_cpu_partial);
4459 
4460 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4461 {
4462 	return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4463 }
4464 
4465 static ssize_t reclaim_account_store(struct kmem_cache *s,
4466 				const char *buf, size_t length)
4467 {
4468 	s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4469 	if (buf[0] == '1')
4470 		s->flags |= SLAB_RECLAIM_ACCOUNT;
4471 	return length;
4472 }
4473 SLAB_ATTR(reclaim_account);
4474 
4475 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4476 {
4477 	return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4478 }
4479 SLAB_ATTR_RO(hwcache_align);
4480 
4481 #ifdef CONFIG_ZONE_DMA
4482 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4483 {
4484 	return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4485 }
4486 SLAB_ATTR_RO(cache_dma);
4487 #endif
4488 
4489 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4490 {
4491 	return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4492 }
4493 SLAB_ATTR_RO(destroy_by_rcu);
4494 
4495 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4496 {
4497 	return sprintf(buf, "%d\n", s->reserved);
4498 }
4499 SLAB_ATTR_RO(reserved);
4500 
4501 #ifdef CONFIG_SLUB_DEBUG
4502 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4503 {
4504 	return show_slab_objects(s, buf, SO_ALL);
4505 }
4506 SLAB_ATTR_RO(slabs);
4507 
4508 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4509 {
4510 	return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4511 }
4512 SLAB_ATTR_RO(total_objects);
4513 
4514 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4515 {
4516 	return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4517 }
4518 
4519 static ssize_t sanity_checks_store(struct kmem_cache *s,
4520 				const char *buf, size_t length)
4521 {
4522 	s->flags &= ~SLAB_DEBUG_FREE;
4523 	if (buf[0] == '1') {
4524 		s->flags &= ~__CMPXCHG_DOUBLE;
4525 		s->flags |= SLAB_DEBUG_FREE;
4526 	}
4527 	return length;
4528 }
4529 SLAB_ATTR(sanity_checks);
4530 
4531 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4532 {
4533 	return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4534 }
4535 
4536 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4537 							size_t length)
4538 {
4539 	/*
4540 	 * Tracing a merged cache is going to give confusing results
4541 	 * as well as cause other issues like converting a mergeable
4542 	 * cache into an umergeable one.
4543 	 */
4544 	if (s->refcount > 1)
4545 		return -EINVAL;
4546 
4547 	s->flags &= ~SLAB_TRACE;
4548 	if (buf[0] == '1') {
4549 		s->flags &= ~__CMPXCHG_DOUBLE;
4550 		s->flags |= SLAB_TRACE;
4551 	}
4552 	return length;
4553 }
4554 SLAB_ATTR(trace);
4555 
4556 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4557 {
4558 	return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4559 }
4560 
4561 static ssize_t red_zone_store(struct kmem_cache *s,
4562 				const char *buf, size_t length)
4563 {
4564 	if (any_slab_objects(s))
4565 		return -EBUSY;
4566 
4567 	s->flags &= ~SLAB_RED_ZONE;
4568 	if (buf[0] == '1') {
4569 		s->flags &= ~__CMPXCHG_DOUBLE;
4570 		s->flags |= SLAB_RED_ZONE;
4571 	}
4572 	calculate_sizes(s, -1);
4573 	return length;
4574 }
4575 SLAB_ATTR(red_zone);
4576 
4577 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4578 {
4579 	return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4580 }
4581 
4582 static ssize_t poison_store(struct kmem_cache *s,
4583 				const char *buf, size_t length)
4584 {
4585 	if (any_slab_objects(s))
4586 		return -EBUSY;
4587 
4588 	s->flags &= ~SLAB_POISON;
4589 	if (buf[0] == '1') {
4590 		s->flags &= ~__CMPXCHG_DOUBLE;
4591 		s->flags |= SLAB_POISON;
4592 	}
4593 	calculate_sizes(s, -1);
4594 	return length;
4595 }
4596 SLAB_ATTR(poison);
4597 
4598 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4599 {
4600 	return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4601 }
4602 
4603 static ssize_t store_user_store(struct kmem_cache *s,
4604 				const char *buf, size_t length)
4605 {
4606 	if (any_slab_objects(s))
4607 		return -EBUSY;
4608 
4609 	s->flags &= ~SLAB_STORE_USER;
4610 	if (buf[0] == '1') {
4611 		s->flags &= ~__CMPXCHG_DOUBLE;
4612 		s->flags |= SLAB_STORE_USER;
4613 	}
4614 	calculate_sizes(s, -1);
4615 	return length;
4616 }
4617 SLAB_ATTR(store_user);
4618 
4619 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4620 {
4621 	return 0;
4622 }
4623 
4624 static ssize_t validate_store(struct kmem_cache *s,
4625 			const char *buf, size_t length)
4626 {
4627 	int ret = -EINVAL;
4628 
4629 	if (buf[0] == '1') {
4630 		ret = validate_slab_cache(s);
4631 		if (ret >= 0)
4632 			ret = length;
4633 	}
4634 	return ret;
4635 }
4636 SLAB_ATTR(validate);
4637 
4638 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4639 {
4640 	if (!(s->flags & SLAB_STORE_USER))
4641 		return -ENOSYS;
4642 	return list_locations(s, buf, TRACK_ALLOC);
4643 }
4644 SLAB_ATTR_RO(alloc_calls);
4645 
4646 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4647 {
4648 	if (!(s->flags & SLAB_STORE_USER))
4649 		return -ENOSYS;
4650 	return list_locations(s, buf, TRACK_FREE);
4651 }
4652 SLAB_ATTR_RO(free_calls);
4653 #endif /* CONFIG_SLUB_DEBUG */
4654 
4655 #ifdef CONFIG_FAILSLAB
4656 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4657 {
4658 	return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4659 }
4660 
4661 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4662 							size_t length)
4663 {
4664 	if (s->refcount > 1)
4665 		return -EINVAL;
4666 
4667 	s->flags &= ~SLAB_FAILSLAB;
4668 	if (buf[0] == '1')
4669 		s->flags |= SLAB_FAILSLAB;
4670 	return length;
4671 }
4672 SLAB_ATTR(failslab);
4673 #endif
4674 
4675 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4676 {
4677 	return 0;
4678 }
4679 
4680 static ssize_t shrink_store(struct kmem_cache *s,
4681 			const char *buf, size_t length)
4682 {
4683 	if (buf[0] == '1') {
4684 		int rc = kmem_cache_shrink(s);
4685 
4686 		if (rc)
4687 			return rc;
4688 	} else
4689 		return -EINVAL;
4690 	return length;
4691 }
4692 SLAB_ATTR(shrink);
4693 
4694 #ifdef CONFIG_NUMA
4695 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4696 {
4697 	return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4698 }
4699 
4700 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4701 				const char *buf, size_t length)
4702 {
4703 	unsigned long ratio;
4704 	int err;
4705 
4706 	err = kstrtoul(buf, 10, &ratio);
4707 	if (err)
4708 		return err;
4709 
4710 	if (ratio <= 100)
4711 		s->remote_node_defrag_ratio = ratio * 10;
4712 
4713 	return length;
4714 }
4715 SLAB_ATTR(remote_node_defrag_ratio);
4716 #endif
4717 
4718 #ifdef CONFIG_SLUB_STATS
4719 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4720 {
4721 	unsigned long sum  = 0;
4722 	int cpu;
4723 	int len;
4724 	int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4725 
4726 	if (!data)
4727 		return -ENOMEM;
4728 
4729 	for_each_online_cpu(cpu) {
4730 		unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4731 
4732 		data[cpu] = x;
4733 		sum += x;
4734 	}
4735 
4736 	len = sprintf(buf, "%lu", sum);
4737 
4738 #ifdef CONFIG_SMP
4739 	for_each_online_cpu(cpu) {
4740 		if (data[cpu] && len < PAGE_SIZE - 20)
4741 			len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4742 	}
4743 #endif
4744 	kfree(data);
4745 	return len + sprintf(buf + len, "\n");
4746 }
4747 
4748 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4749 {
4750 	int cpu;
4751 
4752 	for_each_online_cpu(cpu)
4753 		per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4754 }
4755 
4756 #define STAT_ATTR(si, text) 					\
4757 static ssize_t text##_show(struct kmem_cache *s, char *buf)	\
4758 {								\
4759 	return show_stat(s, buf, si);				\
4760 }								\
4761 static ssize_t text##_store(struct kmem_cache *s,		\
4762 				const char *buf, size_t length)	\
4763 {								\
4764 	if (buf[0] != '0')					\
4765 		return -EINVAL;					\
4766 	clear_stat(s, si);					\
4767 	return length;						\
4768 }								\
4769 SLAB_ATTR(text);						\
4770 
4771 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4772 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4773 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4774 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4775 STAT_ATTR(FREE_FROZEN, free_frozen);
4776 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4777 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4778 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4779 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4780 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4781 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
4782 STAT_ATTR(FREE_SLAB, free_slab);
4783 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4784 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4785 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4786 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4787 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4788 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4789 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
4790 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4791 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4792 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4793 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
4794 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
4795 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
4796 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
4797 #endif
4798 
4799 static struct attribute *slab_attrs[] = {
4800 	&slab_size_attr.attr,
4801 	&object_size_attr.attr,
4802 	&objs_per_slab_attr.attr,
4803 	&order_attr.attr,
4804 	&min_partial_attr.attr,
4805 	&cpu_partial_attr.attr,
4806 	&objects_attr.attr,
4807 	&objects_partial_attr.attr,
4808 	&partial_attr.attr,
4809 	&cpu_slabs_attr.attr,
4810 	&ctor_attr.attr,
4811 	&aliases_attr.attr,
4812 	&align_attr.attr,
4813 	&hwcache_align_attr.attr,
4814 	&reclaim_account_attr.attr,
4815 	&destroy_by_rcu_attr.attr,
4816 	&shrink_attr.attr,
4817 	&reserved_attr.attr,
4818 	&slabs_cpu_partial_attr.attr,
4819 #ifdef CONFIG_SLUB_DEBUG
4820 	&total_objects_attr.attr,
4821 	&slabs_attr.attr,
4822 	&sanity_checks_attr.attr,
4823 	&trace_attr.attr,
4824 	&red_zone_attr.attr,
4825 	&poison_attr.attr,
4826 	&store_user_attr.attr,
4827 	&validate_attr.attr,
4828 	&alloc_calls_attr.attr,
4829 	&free_calls_attr.attr,
4830 #endif
4831 #ifdef CONFIG_ZONE_DMA
4832 	&cache_dma_attr.attr,
4833 #endif
4834 #ifdef CONFIG_NUMA
4835 	&remote_node_defrag_ratio_attr.attr,
4836 #endif
4837 #ifdef CONFIG_SLUB_STATS
4838 	&alloc_fastpath_attr.attr,
4839 	&alloc_slowpath_attr.attr,
4840 	&free_fastpath_attr.attr,
4841 	&free_slowpath_attr.attr,
4842 	&free_frozen_attr.attr,
4843 	&free_add_partial_attr.attr,
4844 	&free_remove_partial_attr.attr,
4845 	&alloc_from_partial_attr.attr,
4846 	&alloc_slab_attr.attr,
4847 	&alloc_refill_attr.attr,
4848 	&alloc_node_mismatch_attr.attr,
4849 	&free_slab_attr.attr,
4850 	&cpuslab_flush_attr.attr,
4851 	&deactivate_full_attr.attr,
4852 	&deactivate_empty_attr.attr,
4853 	&deactivate_to_head_attr.attr,
4854 	&deactivate_to_tail_attr.attr,
4855 	&deactivate_remote_frees_attr.attr,
4856 	&deactivate_bypass_attr.attr,
4857 	&order_fallback_attr.attr,
4858 	&cmpxchg_double_fail_attr.attr,
4859 	&cmpxchg_double_cpu_fail_attr.attr,
4860 	&cpu_partial_alloc_attr.attr,
4861 	&cpu_partial_free_attr.attr,
4862 	&cpu_partial_node_attr.attr,
4863 	&cpu_partial_drain_attr.attr,
4864 #endif
4865 #ifdef CONFIG_FAILSLAB
4866 	&failslab_attr.attr,
4867 #endif
4868 
4869 	NULL
4870 };
4871 
4872 static struct attribute_group slab_attr_group = {
4873 	.attrs = slab_attrs,
4874 };
4875 
4876 static ssize_t slab_attr_show(struct kobject *kobj,
4877 				struct attribute *attr,
4878 				char *buf)
4879 {
4880 	struct slab_attribute *attribute;
4881 	struct kmem_cache *s;
4882 	int err;
4883 
4884 	attribute = to_slab_attr(attr);
4885 	s = to_slab(kobj);
4886 
4887 	if (!attribute->show)
4888 		return -EIO;
4889 
4890 	err = attribute->show(s, buf);
4891 
4892 	return err;
4893 }
4894 
4895 static ssize_t slab_attr_store(struct kobject *kobj,
4896 				struct attribute *attr,
4897 				const char *buf, size_t len)
4898 {
4899 	struct slab_attribute *attribute;
4900 	struct kmem_cache *s;
4901 	int err;
4902 
4903 	attribute = to_slab_attr(attr);
4904 	s = to_slab(kobj);
4905 
4906 	if (!attribute->store)
4907 		return -EIO;
4908 
4909 	err = attribute->store(s, buf, len);
4910 #ifdef CONFIG_MEMCG_KMEM
4911 	if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
4912 		int i;
4913 
4914 		mutex_lock(&slab_mutex);
4915 		if (s->max_attr_size < len)
4916 			s->max_attr_size = len;
4917 
4918 		/*
4919 		 * This is a best effort propagation, so this function's return
4920 		 * value will be determined by the parent cache only. This is
4921 		 * basically because not all attributes will have a well
4922 		 * defined semantics for rollbacks - most of the actions will
4923 		 * have permanent effects.
4924 		 *
4925 		 * Returning the error value of any of the children that fail
4926 		 * is not 100 % defined, in the sense that users seeing the
4927 		 * error code won't be able to know anything about the state of
4928 		 * the cache.
4929 		 *
4930 		 * Only returning the error code for the parent cache at least
4931 		 * has well defined semantics. The cache being written to
4932 		 * directly either failed or succeeded, in which case we loop
4933 		 * through the descendants with best-effort propagation.
4934 		 */
4935 		for_each_memcg_cache_index(i) {
4936 			struct kmem_cache *c = cache_from_memcg_idx(s, i);
4937 			if (c)
4938 				attribute->store(c, buf, len);
4939 		}
4940 		mutex_unlock(&slab_mutex);
4941 	}
4942 #endif
4943 	return err;
4944 }
4945 
4946 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
4947 {
4948 #ifdef CONFIG_MEMCG_KMEM
4949 	int i;
4950 	char *buffer = NULL;
4951 	struct kmem_cache *root_cache;
4952 
4953 	if (is_root_cache(s))
4954 		return;
4955 
4956 	root_cache = s->memcg_params->root_cache;
4957 
4958 	/*
4959 	 * This mean this cache had no attribute written. Therefore, no point
4960 	 * in copying default values around
4961 	 */
4962 	if (!root_cache->max_attr_size)
4963 		return;
4964 
4965 	for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
4966 		char mbuf[64];
4967 		char *buf;
4968 		struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
4969 
4970 		if (!attr || !attr->store || !attr->show)
4971 			continue;
4972 
4973 		/*
4974 		 * It is really bad that we have to allocate here, so we will
4975 		 * do it only as a fallback. If we actually allocate, though,
4976 		 * we can just use the allocated buffer until the end.
4977 		 *
4978 		 * Most of the slub attributes will tend to be very small in
4979 		 * size, but sysfs allows buffers up to a page, so they can
4980 		 * theoretically happen.
4981 		 */
4982 		if (buffer)
4983 			buf = buffer;
4984 		else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
4985 			buf = mbuf;
4986 		else {
4987 			buffer = (char *) get_zeroed_page(GFP_KERNEL);
4988 			if (WARN_ON(!buffer))
4989 				continue;
4990 			buf = buffer;
4991 		}
4992 
4993 		attr->show(root_cache, buf);
4994 		attr->store(s, buf, strlen(buf));
4995 	}
4996 
4997 	if (buffer)
4998 		free_page((unsigned long)buffer);
4999 #endif
5000 }
5001 
5002 static void kmem_cache_release(struct kobject *k)
5003 {
5004 	slab_kmem_cache_release(to_slab(k));
5005 }
5006 
5007 static const struct sysfs_ops slab_sysfs_ops = {
5008 	.show = slab_attr_show,
5009 	.store = slab_attr_store,
5010 };
5011 
5012 static struct kobj_type slab_ktype = {
5013 	.sysfs_ops = &slab_sysfs_ops,
5014 	.release = kmem_cache_release,
5015 };
5016 
5017 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5018 {
5019 	struct kobj_type *ktype = get_ktype(kobj);
5020 
5021 	if (ktype == &slab_ktype)
5022 		return 1;
5023 	return 0;
5024 }
5025 
5026 static const struct kset_uevent_ops slab_uevent_ops = {
5027 	.filter = uevent_filter,
5028 };
5029 
5030 static struct kset *slab_kset;
5031 
5032 static inline struct kset *cache_kset(struct kmem_cache *s)
5033 {
5034 #ifdef CONFIG_MEMCG_KMEM
5035 	if (!is_root_cache(s))
5036 		return s->memcg_params->root_cache->memcg_kset;
5037 #endif
5038 	return slab_kset;
5039 }
5040 
5041 #define ID_STR_LENGTH 64
5042 
5043 /* Create a unique string id for a slab cache:
5044  *
5045  * Format	:[flags-]size
5046  */
5047 static char *create_unique_id(struct kmem_cache *s)
5048 {
5049 	char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5050 	char *p = name;
5051 
5052 	BUG_ON(!name);
5053 
5054 	*p++ = ':';
5055 	/*
5056 	 * First flags affecting slabcache operations. We will only
5057 	 * get here for aliasable slabs so we do not need to support
5058 	 * too many flags. The flags here must cover all flags that
5059 	 * are matched during merging to guarantee that the id is
5060 	 * unique.
5061 	 */
5062 	if (s->flags & SLAB_CACHE_DMA)
5063 		*p++ = 'd';
5064 	if (s->flags & SLAB_RECLAIM_ACCOUNT)
5065 		*p++ = 'a';
5066 	if (s->flags & SLAB_DEBUG_FREE)
5067 		*p++ = 'F';
5068 	if (!(s->flags & SLAB_NOTRACK))
5069 		*p++ = 't';
5070 	if (p != name + 1)
5071 		*p++ = '-';
5072 	p += sprintf(p, "%07d", s->size);
5073 
5074 	BUG_ON(p > name + ID_STR_LENGTH - 1);
5075 	return name;
5076 }
5077 
5078 static int sysfs_slab_add(struct kmem_cache *s)
5079 {
5080 	int err;
5081 	const char *name;
5082 	int unmergeable = slab_unmergeable(s);
5083 
5084 	if (unmergeable) {
5085 		/*
5086 		 * Slabcache can never be merged so we can use the name proper.
5087 		 * This is typically the case for debug situations. In that
5088 		 * case we can catch duplicate names easily.
5089 		 */
5090 		sysfs_remove_link(&slab_kset->kobj, s->name);
5091 		name = s->name;
5092 	} else {
5093 		/*
5094 		 * Create a unique name for the slab as a target
5095 		 * for the symlinks.
5096 		 */
5097 		name = create_unique_id(s);
5098 	}
5099 
5100 	s->kobj.kset = cache_kset(s);
5101 	err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5102 	if (err)
5103 		goto out_put_kobj;
5104 
5105 	err = sysfs_create_group(&s->kobj, &slab_attr_group);
5106 	if (err)
5107 		goto out_del_kobj;
5108 
5109 #ifdef CONFIG_MEMCG_KMEM
5110 	if (is_root_cache(s)) {
5111 		s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5112 		if (!s->memcg_kset) {
5113 			err = -ENOMEM;
5114 			goto out_del_kobj;
5115 		}
5116 	}
5117 #endif
5118 
5119 	kobject_uevent(&s->kobj, KOBJ_ADD);
5120 	if (!unmergeable) {
5121 		/* Setup first alias */
5122 		sysfs_slab_alias(s, s->name);
5123 	}
5124 out:
5125 	if (!unmergeable)
5126 		kfree(name);
5127 	return err;
5128 out_del_kobj:
5129 	kobject_del(&s->kobj);
5130 out_put_kobj:
5131 	kobject_put(&s->kobj);
5132 	goto out;
5133 }
5134 
5135 void sysfs_slab_remove(struct kmem_cache *s)
5136 {
5137 	if (slab_state < FULL)
5138 		/*
5139 		 * Sysfs has not been setup yet so no need to remove the
5140 		 * cache from sysfs.
5141 		 */
5142 		return;
5143 
5144 #ifdef CONFIG_MEMCG_KMEM
5145 	kset_unregister(s->memcg_kset);
5146 #endif
5147 	kobject_uevent(&s->kobj, KOBJ_REMOVE);
5148 	kobject_del(&s->kobj);
5149 	kobject_put(&s->kobj);
5150 }
5151 
5152 /*
5153  * Need to buffer aliases during bootup until sysfs becomes
5154  * available lest we lose that information.
5155  */
5156 struct saved_alias {
5157 	struct kmem_cache *s;
5158 	const char *name;
5159 	struct saved_alias *next;
5160 };
5161 
5162 static struct saved_alias *alias_list;
5163 
5164 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5165 {
5166 	struct saved_alias *al;
5167 
5168 	if (slab_state == FULL) {
5169 		/*
5170 		 * If we have a leftover link then remove it.
5171 		 */
5172 		sysfs_remove_link(&slab_kset->kobj, name);
5173 		return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5174 	}
5175 
5176 	al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5177 	if (!al)
5178 		return -ENOMEM;
5179 
5180 	al->s = s;
5181 	al->name = name;
5182 	al->next = alias_list;
5183 	alias_list = al;
5184 	return 0;
5185 }
5186 
5187 static int __init slab_sysfs_init(void)
5188 {
5189 	struct kmem_cache *s;
5190 	int err;
5191 
5192 	mutex_lock(&slab_mutex);
5193 
5194 	slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5195 	if (!slab_kset) {
5196 		mutex_unlock(&slab_mutex);
5197 		pr_err("Cannot register slab subsystem.\n");
5198 		return -ENOSYS;
5199 	}
5200 
5201 	slab_state = FULL;
5202 
5203 	list_for_each_entry(s, &slab_caches, list) {
5204 		err = sysfs_slab_add(s);
5205 		if (err)
5206 			pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5207 			       s->name);
5208 	}
5209 
5210 	while (alias_list) {
5211 		struct saved_alias *al = alias_list;
5212 
5213 		alias_list = alias_list->next;
5214 		err = sysfs_slab_alias(al->s, al->name);
5215 		if (err)
5216 			pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5217 			       al->name);
5218 		kfree(al);
5219 	}
5220 
5221 	mutex_unlock(&slab_mutex);
5222 	resiliency_test();
5223 	return 0;
5224 }
5225 
5226 __initcall(slab_sysfs_init);
5227 #endif /* CONFIG_SYSFS */
5228 
5229 /*
5230  * The /proc/slabinfo ABI
5231  */
5232 #ifdef CONFIG_SLABINFO
5233 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5234 {
5235 	unsigned long nr_slabs = 0;
5236 	unsigned long nr_objs = 0;
5237 	unsigned long nr_free = 0;
5238 	int node;
5239 	struct kmem_cache_node *n;
5240 
5241 	for_each_kmem_cache_node(s, node, n) {
5242 		nr_slabs += node_nr_slabs(n);
5243 		nr_objs += node_nr_objs(n);
5244 		nr_free += count_partial(n, count_free);
5245 	}
5246 
5247 	sinfo->active_objs = nr_objs - nr_free;
5248 	sinfo->num_objs = nr_objs;
5249 	sinfo->active_slabs = nr_slabs;
5250 	sinfo->num_slabs = nr_slabs;
5251 	sinfo->objects_per_slab = oo_objects(s->oo);
5252 	sinfo->cache_order = oo_order(s->oo);
5253 }
5254 
5255 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5256 {
5257 }
5258 
5259 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5260 		       size_t count, loff_t *ppos)
5261 {
5262 	return -EIO;
5263 }
5264 #endif /* CONFIG_SLABINFO */
5265