xref: /openbmc/linux/mm/slub.c (revision b9ddb2500e7e544410f38476ab928fc2fe01e381)
1  // SPDX-License-Identifier: GPL-2.0
2  /*
3   * SLUB: A slab allocator that limits cache line use instead of queuing
4   * objects in per cpu and per node lists.
5   *
6   * The allocator synchronizes using per slab locks or atomic operatios
7   * and only uses a centralized lock to manage a pool of partial slabs.
8   *
9   * (C) 2007 SGI, Christoph Lameter
10   * (C) 2011 Linux Foundation, Christoph Lameter
11   */
12  
13  #include <linux/mm.h>
14  #include <linux/swap.h> /* struct reclaim_state */
15  #include <linux/module.h>
16  #include <linux/bit_spinlock.h>
17  #include <linux/interrupt.h>
18  #include <linux/bitops.h>
19  #include <linux/slab.h>
20  #include "slab.h"
21  #include <linux/proc_fs.h>
22  #include <linux/seq_file.h>
23  #include <linux/kasan.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  #include <linux/random.h>
37  
38  #include <trace/events/kmem.h>
39  
40  #include "internal.h"
41  
42  /*
43   * Lock order:
44   *   1. slab_mutex (Global Mutex)
45   *   2. node->list_lock
46   *   3. slab_lock(page) (Only on some arches and for debugging)
47   *
48   *   slab_mutex
49   *
50   *   The role of the slab_mutex is to protect the list of all the slabs
51   *   and to synchronize major metadata changes to slab cache structures.
52   *
53   *   The slab_lock is only used for debugging and on arches that do not
54   *   have the ability to do a cmpxchg_double. It only protects:
55   *	A. page->freelist	-> List of object free in a page
56   *	B. page->inuse		-> Number of objects in use
57   *	C. page->objects	-> Number of objects in page
58   *	D. page->frozen		-> frozen state
59   *
60   *   If a slab is frozen then it is exempt from list management. It is not
61   *   on any list except per cpu partial list. The processor that froze the
62   *   slab is the one who can perform list operations on the page. Other
63   *   processors may put objects onto the freelist but the processor that
64   *   froze the slab is the only one that can retrieve the objects from the
65   *   page's freelist.
66   *
67   *   The list_lock protects the partial and full list on each node and
68   *   the partial slab counter. If taken then no new slabs may be added or
69   *   removed from the lists nor make the number of partial slabs be modified.
70   *   (Note that the total number of slabs is an atomic value that may be
71   *   modified without taking the list lock).
72   *
73   *   The list_lock is a centralized lock and thus we avoid taking it as
74   *   much as possible. As long as SLUB does not have to handle partial
75   *   slabs, operations can continue without any centralized lock. F.e.
76   *   allocating a long series of objects that fill up slabs does not require
77   *   the list lock.
78   *   Interrupts are disabled during allocation and deallocation in order to
79   *   make the slab allocator safe to use in the context of an irq. In addition
80   *   interrupts are disabled to ensure that the processor does not change
81   *   while handling per_cpu slabs, due to kernel preemption.
82   *
83   * SLUB assigns one slab for allocation to each processor.
84   * Allocations only occur from these slabs called cpu slabs.
85   *
86   * Slabs with free elements are kept on a partial list and during regular
87   * operations no list for full slabs is used. If an object in a full slab is
88   * freed then the slab will show up again on the partial lists.
89   * We track full slabs for debugging purposes though because otherwise we
90   * cannot scan all objects.
91   *
92   * Slabs are freed when they become empty. Teardown and setup is
93   * minimal so we rely on the page allocators per cpu caches for
94   * fast frees and allocs.
95   *
96   * page->frozen		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   * SLAB_DEBUG_FLAGS	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  #ifdef CONFIG_SLUB_DEBUG
118  #ifdef CONFIG_SLUB_DEBUG_ON
119  DEFINE_STATIC_KEY_TRUE(slub_debug_enabled);
120  #else
121  DEFINE_STATIC_KEY_FALSE(slub_debug_enabled);
122  #endif
123  #endif
124  
125  static inline bool kmem_cache_debug(struct kmem_cache *s)
126  {
127  	return kmem_cache_debug_flags(s, SLAB_DEBUG_FLAGS);
128  }
129  
130  void *fixup_red_left(struct kmem_cache *s, void *p)
131  {
132  	if (kmem_cache_debug_flags(s, SLAB_RED_ZONE))
133  		p += s->red_left_pad;
134  
135  	return p;
136  }
137  
138  static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
139  {
140  #ifdef CONFIG_SLUB_CPU_PARTIAL
141  	return !kmem_cache_debug(s);
142  #else
143  	return false;
144  #endif
145  }
146  
147  /*
148   * Issues still to be resolved:
149   *
150   * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
151   *
152   * - Variable sizing of the per node arrays
153   */
154  
155  /* Enable to test recovery from slab corruption on boot */
156  #undef SLUB_RESILIENCY_TEST
157  
158  /* Enable to log cmpxchg failures */
159  #undef SLUB_DEBUG_CMPXCHG
160  
161  /*
162   * Mininum number of partial slabs. These will be left on the partial
163   * lists even if they are empty. kmem_cache_shrink may reclaim them.
164   */
165  #define MIN_PARTIAL 5
166  
167  /*
168   * Maximum number of desirable partial slabs.
169   * The existence of more partial slabs makes kmem_cache_shrink
170   * sort the partial list by the number of objects in use.
171   */
172  #define MAX_PARTIAL 10
173  
174  #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
175  				SLAB_POISON | SLAB_STORE_USER)
176  
177  /*
178   * These debug flags cannot use CMPXCHG because there might be consistency
179   * issues when checking or reading debug information
180   */
181  #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
182  				SLAB_TRACE)
183  
184  
185  /*
186   * Debugging flags that require metadata to be stored in the slab.  These get
187   * disabled when slub_debug=O is used and a cache's min order increases with
188   * metadata.
189   */
190  #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
191  
192  #define OO_SHIFT	16
193  #define OO_MASK		((1 << OO_SHIFT) - 1)
194  #define MAX_OBJS_PER_PAGE	32767 /* since page.objects is u15 */
195  
196  /* Internal SLUB flags */
197  /* Poison object */
198  #define __OBJECT_POISON		((slab_flags_t __force)0x80000000U)
199  /* Use cmpxchg_double */
200  #define __CMPXCHG_DOUBLE	((slab_flags_t __force)0x40000000U)
201  
202  /*
203   * Tracking user of a slab.
204   */
205  #define TRACK_ADDRS_COUNT 16
206  struct track {
207  	unsigned long addr;	/* Called from address */
208  #ifdef CONFIG_STACKTRACE
209  	unsigned long addrs[TRACK_ADDRS_COUNT];	/* Called from address */
210  #endif
211  	int cpu;		/* Was running on cpu */
212  	int pid;		/* Pid context */
213  	unsigned long when;	/* When did the operation occur */
214  };
215  
216  enum track_item { TRACK_ALLOC, TRACK_FREE };
217  
218  #ifdef CONFIG_SYSFS
219  static int sysfs_slab_add(struct kmem_cache *);
220  static int sysfs_slab_alias(struct kmem_cache *, const char *);
221  #else
222  static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
223  static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
224  							{ return 0; }
225  #endif
226  
227  static inline void stat(const struct kmem_cache *s, enum stat_item si)
228  {
229  #ifdef CONFIG_SLUB_STATS
230  	/*
231  	 * The rmw is racy on a preemptible kernel but this is acceptable, so
232  	 * avoid this_cpu_add()'s irq-disable overhead.
233  	 */
234  	raw_cpu_inc(s->cpu_slab->stat[si]);
235  #endif
236  }
237  
238  /********************************************************************
239   * 			Core slab cache functions
240   *******************************************************************/
241  
242  /*
243   * Returns freelist pointer (ptr). With hardening, this is obfuscated
244   * with an XOR of the address where the pointer is held and a per-cache
245   * random number.
246   */
247  static inline void *freelist_ptr(const struct kmem_cache *s, void *ptr,
248  				 unsigned long ptr_addr)
249  {
250  #ifdef CONFIG_SLAB_FREELIST_HARDENED
251  	/*
252  	 * When CONFIG_KASAN_SW/HW_TAGS is enabled, ptr_addr might be tagged.
253  	 * Normally, this doesn't cause any issues, as both set_freepointer()
254  	 * and get_freepointer() are called with a pointer with the same tag.
255  	 * However, there are some issues with CONFIG_SLUB_DEBUG code. For
256  	 * example, when __free_slub() iterates over objects in a cache, it
257  	 * passes untagged pointers to check_object(). check_object() in turns
258  	 * calls get_freepointer() with an untagged pointer, which causes the
259  	 * freepointer to be restored incorrectly.
260  	 */
261  	return (void *)((unsigned long)ptr ^ s->random ^
262  			swab((unsigned long)kasan_reset_tag((void *)ptr_addr)));
263  #else
264  	return ptr;
265  #endif
266  }
267  
268  /* Returns the freelist pointer recorded at location ptr_addr. */
269  static inline void *freelist_dereference(const struct kmem_cache *s,
270  					 void *ptr_addr)
271  {
272  	return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr),
273  			    (unsigned long)ptr_addr);
274  }
275  
276  static inline void *get_freepointer(struct kmem_cache *s, void *object)
277  {
278  	object = kasan_reset_tag(object);
279  	return freelist_dereference(s, object + s->offset);
280  }
281  
282  static void prefetch_freepointer(const struct kmem_cache *s, void *object)
283  {
284  	prefetch(object + s->offset);
285  }
286  
287  static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
288  {
289  	unsigned long freepointer_addr;
290  	void *p;
291  
292  	if (!debug_pagealloc_enabled_static())
293  		return get_freepointer(s, object);
294  
295  	freepointer_addr = (unsigned long)object + s->offset;
296  	copy_from_kernel_nofault(&p, (void **)freepointer_addr, sizeof(p));
297  	return freelist_ptr(s, p, freepointer_addr);
298  }
299  
300  static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
301  {
302  	unsigned long freeptr_addr = (unsigned long)object + s->offset;
303  
304  #ifdef CONFIG_SLAB_FREELIST_HARDENED
305  	BUG_ON(object == fp); /* naive detection of double free or corruption */
306  #endif
307  
308  	freeptr_addr = (unsigned long)kasan_reset_tag((void *)freeptr_addr);
309  	*(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr);
310  }
311  
312  /* Loop over all objects in a slab */
313  #define for_each_object(__p, __s, __addr, __objects) \
314  	for (__p = fixup_red_left(__s, __addr); \
315  		__p < (__addr) + (__objects) * (__s)->size; \
316  		__p += (__s)->size)
317  
318  static inline unsigned int order_objects(unsigned int order, unsigned int size)
319  {
320  	return ((unsigned int)PAGE_SIZE << order) / size;
321  }
322  
323  static inline struct kmem_cache_order_objects oo_make(unsigned int order,
324  		unsigned int size)
325  {
326  	struct kmem_cache_order_objects x = {
327  		(order << OO_SHIFT) + order_objects(order, size)
328  	};
329  
330  	return x;
331  }
332  
333  static inline unsigned int oo_order(struct kmem_cache_order_objects x)
334  {
335  	return x.x >> OO_SHIFT;
336  }
337  
338  static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
339  {
340  	return x.x & OO_MASK;
341  }
342  
343  /*
344   * Per slab locking using the pagelock
345   */
346  static __always_inline void slab_lock(struct page *page)
347  {
348  	VM_BUG_ON_PAGE(PageTail(page), page);
349  	bit_spin_lock(PG_locked, &page->flags);
350  }
351  
352  static __always_inline void slab_unlock(struct page *page)
353  {
354  	VM_BUG_ON_PAGE(PageTail(page), page);
355  	__bit_spin_unlock(PG_locked, &page->flags);
356  }
357  
358  /* Interrupts must be disabled (for the fallback code to work right) */
359  static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
360  		void *freelist_old, unsigned long counters_old,
361  		void *freelist_new, unsigned long counters_new,
362  		const char *n)
363  {
364  	VM_BUG_ON(!irqs_disabled());
365  #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
366      defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
367  	if (s->flags & __CMPXCHG_DOUBLE) {
368  		if (cmpxchg_double(&page->freelist, &page->counters,
369  				   freelist_old, counters_old,
370  				   freelist_new, counters_new))
371  			return true;
372  	} else
373  #endif
374  	{
375  		slab_lock(page);
376  		if (page->freelist == freelist_old &&
377  					page->counters == counters_old) {
378  			page->freelist = freelist_new;
379  			page->counters = counters_new;
380  			slab_unlock(page);
381  			return true;
382  		}
383  		slab_unlock(page);
384  	}
385  
386  	cpu_relax();
387  	stat(s, CMPXCHG_DOUBLE_FAIL);
388  
389  #ifdef SLUB_DEBUG_CMPXCHG
390  	pr_info("%s %s: cmpxchg double redo ", n, s->name);
391  #endif
392  
393  	return false;
394  }
395  
396  static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
397  		void *freelist_old, unsigned long counters_old,
398  		void *freelist_new, unsigned long counters_new,
399  		const char *n)
400  {
401  #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
402      defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
403  	if (s->flags & __CMPXCHG_DOUBLE) {
404  		if (cmpxchg_double(&page->freelist, &page->counters,
405  				   freelist_old, counters_old,
406  				   freelist_new, counters_new))
407  			return true;
408  	} else
409  #endif
410  	{
411  		unsigned long flags;
412  
413  		local_irq_save(flags);
414  		slab_lock(page);
415  		if (page->freelist == freelist_old &&
416  					page->counters == counters_old) {
417  			page->freelist = freelist_new;
418  			page->counters = counters_new;
419  			slab_unlock(page);
420  			local_irq_restore(flags);
421  			return true;
422  		}
423  		slab_unlock(page);
424  		local_irq_restore(flags);
425  	}
426  
427  	cpu_relax();
428  	stat(s, CMPXCHG_DOUBLE_FAIL);
429  
430  #ifdef SLUB_DEBUG_CMPXCHG
431  	pr_info("%s %s: cmpxchg double redo ", n, s->name);
432  #endif
433  
434  	return false;
435  }
436  
437  #ifdef CONFIG_SLUB_DEBUG
438  static unsigned long object_map[BITS_TO_LONGS(MAX_OBJS_PER_PAGE)];
439  static DEFINE_SPINLOCK(object_map_lock);
440  
441  /*
442   * Determine a map of object in use on a page.
443   *
444   * Node listlock must be held to guarantee that the page does
445   * not vanish from under us.
446   */
447  static unsigned long *get_map(struct kmem_cache *s, struct page *page)
448  	__acquires(&object_map_lock)
449  {
450  	void *p;
451  	void *addr = page_address(page);
452  
453  	VM_BUG_ON(!irqs_disabled());
454  
455  	spin_lock(&object_map_lock);
456  
457  	bitmap_zero(object_map, page->objects);
458  
459  	for (p = page->freelist; p; p = get_freepointer(s, p))
460  		set_bit(__obj_to_index(s, addr, p), object_map);
461  
462  	return object_map;
463  }
464  
465  static void put_map(unsigned long *map) __releases(&object_map_lock)
466  {
467  	VM_BUG_ON(map != object_map);
468  	spin_unlock(&object_map_lock);
469  }
470  
471  static inline unsigned int size_from_object(struct kmem_cache *s)
472  {
473  	if (s->flags & SLAB_RED_ZONE)
474  		return s->size - s->red_left_pad;
475  
476  	return s->size;
477  }
478  
479  static inline void *restore_red_left(struct kmem_cache *s, void *p)
480  {
481  	if (s->flags & SLAB_RED_ZONE)
482  		p -= s->red_left_pad;
483  
484  	return p;
485  }
486  
487  /*
488   * Debug settings:
489   */
490  #if defined(CONFIG_SLUB_DEBUG_ON)
491  static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
492  #else
493  static slab_flags_t slub_debug;
494  #endif
495  
496  static char *slub_debug_string;
497  static int disable_higher_order_debug;
498  
499  /*
500   * slub is about to manipulate internal object metadata.  This memory lies
501   * outside the range of the allocated object, so accessing it would normally
502   * be reported by kasan as a bounds error.  metadata_access_enable() is used
503   * to tell kasan that these accesses are OK.
504   */
505  static inline void metadata_access_enable(void)
506  {
507  	kasan_disable_current();
508  }
509  
510  static inline void metadata_access_disable(void)
511  {
512  	kasan_enable_current();
513  }
514  
515  /*
516   * Object debugging
517   */
518  
519  /* Verify that a pointer has an address that is valid within a slab page */
520  static inline int check_valid_pointer(struct kmem_cache *s,
521  				struct page *page, void *object)
522  {
523  	void *base;
524  
525  	if (!object)
526  		return 1;
527  
528  	base = page_address(page);
529  	object = kasan_reset_tag(object);
530  	object = restore_red_left(s, object);
531  	if (object < base || object >= base + page->objects * s->size ||
532  		(object - base) % s->size) {
533  		return 0;
534  	}
535  
536  	return 1;
537  }
538  
539  static void print_section(char *level, char *text, u8 *addr,
540  			  unsigned int length)
541  {
542  	metadata_access_enable();
543  	print_hex_dump(level, kasan_reset_tag(text), DUMP_PREFIX_ADDRESS,
544  			16, 1, addr, length, 1);
545  	metadata_access_disable();
546  }
547  
548  /*
549   * See comment in calculate_sizes().
550   */
551  static inline bool freeptr_outside_object(struct kmem_cache *s)
552  {
553  	return s->offset >= s->inuse;
554  }
555  
556  /*
557   * Return offset of the end of info block which is inuse + free pointer if
558   * not overlapping with object.
559   */
560  static inline unsigned int get_info_end(struct kmem_cache *s)
561  {
562  	if (freeptr_outside_object(s))
563  		return s->inuse + sizeof(void *);
564  	else
565  		return s->inuse;
566  }
567  
568  static struct track *get_track(struct kmem_cache *s, void *object,
569  	enum track_item alloc)
570  {
571  	struct track *p;
572  
573  	p = object + get_info_end(s);
574  
575  	return kasan_reset_tag(p + alloc);
576  }
577  
578  static void set_track(struct kmem_cache *s, void *object,
579  			enum track_item alloc, unsigned long addr)
580  {
581  	struct track *p = get_track(s, object, alloc);
582  
583  	if (addr) {
584  #ifdef CONFIG_STACKTRACE
585  		unsigned int nr_entries;
586  
587  		metadata_access_enable();
588  		nr_entries = stack_trace_save(kasan_reset_tag(p->addrs),
589  					      TRACK_ADDRS_COUNT, 3);
590  		metadata_access_disable();
591  
592  		if (nr_entries < TRACK_ADDRS_COUNT)
593  			p->addrs[nr_entries] = 0;
594  #endif
595  		p->addr = addr;
596  		p->cpu = smp_processor_id();
597  		p->pid = current->pid;
598  		p->when = jiffies;
599  	} else {
600  		memset(p, 0, sizeof(struct track));
601  	}
602  }
603  
604  static void init_tracking(struct kmem_cache *s, void *object)
605  {
606  	if (!(s->flags & SLAB_STORE_USER))
607  		return;
608  
609  	set_track(s, object, TRACK_FREE, 0UL);
610  	set_track(s, object, TRACK_ALLOC, 0UL);
611  }
612  
613  static void print_track(const char *s, struct track *t, unsigned long pr_time)
614  {
615  	if (!t->addr)
616  		return;
617  
618  	pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
619  	       s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
620  #ifdef CONFIG_STACKTRACE
621  	{
622  		int i;
623  		for (i = 0; i < TRACK_ADDRS_COUNT; i++)
624  			if (t->addrs[i])
625  				pr_err("\t%pS\n", (void *)t->addrs[i]);
626  			else
627  				break;
628  	}
629  #endif
630  }
631  
632  void print_tracking(struct kmem_cache *s, void *object)
633  {
634  	unsigned long pr_time = jiffies;
635  	if (!(s->flags & SLAB_STORE_USER))
636  		return;
637  
638  	print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
639  	print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
640  }
641  
642  static void print_page_info(struct page *page)
643  {
644  	pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
645  	       page, page->objects, page->inuse, page->freelist, page->flags);
646  
647  }
648  
649  static void slab_bug(struct kmem_cache *s, char *fmt, ...)
650  {
651  	struct va_format vaf;
652  	va_list args;
653  
654  	va_start(args, fmt);
655  	vaf.fmt = fmt;
656  	vaf.va = &args;
657  	pr_err("=============================================================================\n");
658  	pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
659  	pr_err("-----------------------------------------------------------------------------\n\n");
660  
661  	add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
662  	va_end(args);
663  }
664  
665  static void slab_fix(struct kmem_cache *s, char *fmt, ...)
666  {
667  	struct va_format vaf;
668  	va_list args;
669  
670  	va_start(args, fmt);
671  	vaf.fmt = fmt;
672  	vaf.va = &args;
673  	pr_err("FIX %s: %pV\n", s->name, &vaf);
674  	va_end(args);
675  }
676  
677  static bool freelist_corrupted(struct kmem_cache *s, struct page *page,
678  			       void **freelist, void *nextfree)
679  {
680  	if ((s->flags & SLAB_CONSISTENCY_CHECKS) &&
681  	    !check_valid_pointer(s, page, nextfree) && freelist) {
682  		object_err(s, page, *freelist, "Freechain corrupt");
683  		*freelist = NULL;
684  		slab_fix(s, "Isolate corrupted freechain");
685  		return true;
686  	}
687  
688  	return false;
689  }
690  
691  static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
692  {
693  	unsigned int off;	/* Offset of last byte */
694  	u8 *addr = page_address(page);
695  
696  	print_tracking(s, p);
697  
698  	print_page_info(page);
699  
700  	pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
701  	       p, p - addr, get_freepointer(s, p));
702  
703  	if (s->flags & SLAB_RED_ZONE)
704  		print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
705  			      s->red_left_pad);
706  	else if (p > addr + 16)
707  		print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
708  
709  	print_section(KERN_ERR, "Object ", p,
710  		      min_t(unsigned int, s->object_size, PAGE_SIZE));
711  	if (s->flags & SLAB_RED_ZONE)
712  		print_section(KERN_ERR, "Redzone ", p + s->object_size,
713  			s->inuse - s->object_size);
714  
715  	off = get_info_end(s);
716  
717  	if (s->flags & SLAB_STORE_USER)
718  		off += 2 * sizeof(struct track);
719  
720  	off += kasan_metadata_size(s);
721  
722  	if (off != size_from_object(s))
723  		/* Beginning of the filler is the free pointer */
724  		print_section(KERN_ERR, "Padding ", p + off,
725  			      size_from_object(s) - off);
726  
727  	dump_stack();
728  }
729  
730  void object_err(struct kmem_cache *s, struct page *page,
731  			u8 *object, char *reason)
732  {
733  	slab_bug(s, "%s", reason);
734  	print_trailer(s, page, object);
735  }
736  
737  static __printf(3, 4) void slab_err(struct kmem_cache *s, struct page *page,
738  			const char *fmt, ...)
739  {
740  	va_list args;
741  	char buf[100];
742  
743  	va_start(args, fmt);
744  	vsnprintf(buf, sizeof(buf), fmt, args);
745  	va_end(args);
746  	slab_bug(s, "%s", buf);
747  	print_page_info(page);
748  	dump_stack();
749  }
750  
751  static void init_object(struct kmem_cache *s, void *object, u8 val)
752  {
753  	u8 *p = kasan_reset_tag(object);
754  
755  	if (s->flags & SLAB_RED_ZONE)
756  		memset(p - s->red_left_pad, val, s->red_left_pad);
757  
758  	if (s->flags & __OBJECT_POISON) {
759  		memset(p, POISON_FREE, s->object_size - 1);
760  		p[s->object_size - 1] = POISON_END;
761  	}
762  
763  	if (s->flags & SLAB_RED_ZONE)
764  		memset(p + s->object_size, val, s->inuse - s->object_size);
765  }
766  
767  static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
768  						void *from, void *to)
769  {
770  	slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
771  	memset(from, data, to - from);
772  }
773  
774  static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
775  			u8 *object, char *what,
776  			u8 *start, unsigned int value, unsigned int bytes)
777  {
778  	u8 *fault;
779  	u8 *end;
780  	u8 *addr = page_address(page);
781  
782  	metadata_access_enable();
783  	fault = memchr_inv(kasan_reset_tag(start), value, bytes);
784  	metadata_access_disable();
785  	if (!fault)
786  		return 1;
787  
788  	end = start + bytes;
789  	while (end > fault && end[-1] == value)
790  		end--;
791  
792  	slab_bug(s, "%s overwritten", what);
793  	pr_err("INFO: 0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
794  					fault, end - 1, fault - addr,
795  					fault[0], value);
796  	print_trailer(s, page, object);
797  
798  	restore_bytes(s, what, value, fault, end);
799  	return 0;
800  }
801  
802  /*
803   * Object layout:
804   *
805   * object address
806   * 	Bytes of the object to be managed.
807   * 	If the freepointer may overlay the object then the free
808   *	pointer is at the middle of the object.
809   *
810   * 	Poisoning uses 0x6b (POISON_FREE) and the last byte is
811   * 	0xa5 (POISON_END)
812   *
813   * object + s->object_size
814   * 	Padding to reach word boundary. This is also used for Redzoning.
815   * 	Padding is extended by another word if Redzoning is enabled and
816   * 	object_size == inuse.
817   *
818   * 	We fill with 0xbb (RED_INACTIVE) for inactive objects and with
819   * 	0xcc (RED_ACTIVE) for objects in use.
820   *
821   * object + s->inuse
822   * 	Meta data starts here.
823   *
824   * 	A. Free pointer (if we cannot overwrite object on free)
825   * 	B. Tracking data for SLAB_STORE_USER
826   * 	C. Padding to reach required alignment boundary or at mininum
827   * 		one word if debugging is on to be able to detect writes
828   * 		before the word boundary.
829   *
830   *	Padding is done using 0x5a (POISON_INUSE)
831   *
832   * object + s->size
833   * 	Nothing is used beyond s->size.
834   *
835   * If slabcaches are merged then the object_size and inuse boundaries are mostly
836   * ignored. And therefore no slab options that rely on these boundaries
837   * may be used with merged slabcaches.
838   */
839  
840  static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
841  {
842  	unsigned long off = get_info_end(s);	/* The end of info */
843  
844  	if (s->flags & SLAB_STORE_USER)
845  		/* We also have user information there */
846  		off += 2 * sizeof(struct track);
847  
848  	off += kasan_metadata_size(s);
849  
850  	if (size_from_object(s) == off)
851  		return 1;
852  
853  	return check_bytes_and_report(s, page, p, "Object padding",
854  			p + off, POISON_INUSE, size_from_object(s) - off);
855  }
856  
857  /* Check the pad bytes at the end of a slab page */
858  static int slab_pad_check(struct kmem_cache *s, struct page *page)
859  {
860  	u8 *start;
861  	u8 *fault;
862  	u8 *end;
863  	u8 *pad;
864  	int length;
865  	int remainder;
866  
867  	if (!(s->flags & SLAB_POISON))
868  		return 1;
869  
870  	start = page_address(page);
871  	length = page_size(page);
872  	end = start + length;
873  	remainder = length % s->size;
874  	if (!remainder)
875  		return 1;
876  
877  	pad = end - remainder;
878  	metadata_access_enable();
879  	fault = memchr_inv(kasan_reset_tag(pad), POISON_INUSE, remainder);
880  	metadata_access_disable();
881  	if (!fault)
882  		return 1;
883  	while (end > fault && end[-1] == POISON_INUSE)
884  		end--;
885  
886  	slab_err(s, page, "Padding overwritten. 0x%p-0x%p @offset=%tu",
887  			fault, end - 1, fault - start);
888  	print_section(KERN_ERR, "Padding ", pad, remainder);
889  
890  	restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
891  	return 0;
892  }
893  
894  static int check_object(struct kmem_cache *s, struct page *page,
895  					void *object, u8 val)
896  {
897  	u8 *p = object;
898  	u8 *endobject = object + s->object_size;
899  
900  	if (s->flags & SLAB_RED_ZONE) {
901  		if (!check_bytes_and_report(s, page, object, "Redzone",
902  			object - s->red_left_pad, val, s->red_left_pad))
903  			return 0;
904  
905  		if (!check_bytes_and_report(s, page, object, "Redzone",
906  			endobject, val, s->inuse - s->object_size))
907  			return 0;
908  	} else {
909  		if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
910  			check_bytes_and_report(s, page, p, "Alignment padding",
911  				endobject, POISON_INUSE,
912  				s->inuse - s->object_size);
913  		}
914  	}
915  
916  	if (s->flags & SLAB_POISON) {
917  		if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
918  			(!check_bytes_and_report(s, page, p, "Poison", p,
919  					POISON_FREE, s->object_size - 1) ||
920  			 !check_bytes_and_report(s, page, p, "Poison",
921  				p + s->object_size - 1, POISON_END, 1)))
922  			return 0;
923  		/*
924  		 * check_pad_bytes cleans up on its own.
925  		 */
926  		check_pad_bytes(s, page, p);
927  	}
928  
929  	if (!freeptr_outside_object(s) && val == SLUB_RED_ACTIVE)
930  		/*
931  		 * Object and freepointer overlap. Cannot check
932  		 * freepointer while object is allocated.
933  		 */
934  		return 1;
935  
936  	/* Check free pointer validity */
937  	if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
938  		object_err(s, page, p, "Freepointer corrupt");
939  		/*
940  		 * No choice but to zap it and thus lose the remainder
941  		 * of the free objects in this slab. May cause
942  		 * another error because the object count is now wrong.
943  		 */
944  		set_freepointer(s, p, NULL);
945  		return 0;
946  	}
947  	return 1;
948  }
949  
950  static int check_slab(struct kmem_cache *s, struct page *page)
951  {
952  	int maxobj;
953  
954  	VM_BUG_ON(!irqs_disabled());
955  
956  	if (!PageSlab(page)) {
957  		slab_err(s, page, "Not a valid slab page");
958  		return 0;
959  	}
960  
961  	maxobj = order_objects(compound_order(page), s->size);
962  	if (page->objects > maxobj) {
963  		slab_err(s, page, "objects %u > max %u",
964  			page->objects, maxobj);
965  		return 0;
966  	}
967  	if (page->inuse > page->objects) {
968  		slab_err(s, page, "inuse %u > max %u",
969  			page->inuse, page->objects);
970  		return 0;
971  	}
972  	/* Slab_pad_check fixes things up after itself */
973  	slab_pad_check(s, page);
974  	return 1;
975  }
976  
977  /*
978   * Determine if a certain object on a page is on the freelist. Must hold the
979   * slab lock to guarantee that the chains are in a consistent state.
980   */
981  static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
982  {
983  	int nr = 0;
984  	void *fp;
985  	void *object = NULL;
986  	int max_objects;
987  
988  	fp = page->freelist;
989  	while (fp && nr <= page->objects) {
990  		if (fp == search)
991  			return 1;
992  		if (!check_valid_pointer(s, page, fp)) {
993  			if (object) {
994  				object_err(s, page, object,
995  					"Freechain corrupt");
996  				set_freepointer(s, object, NULL);
997  			} else {
998  				slab_err(s, page, "Freepointer corrupt");
999  				page->freelist = NULL;
1000  				page->inuse = page->objects;
1001  				slab_fix(s, "Freelist cleared");
1002  				return 0;
1003  			}
1004  			break;
1005  		}
1006  		object = fp;
1007  		fp = get_freepointer(s, object);
1008  		nr++;
1009  	}
1010  
1011  	max_objects = order_objects(compound_order(page), s->size);
1012  	if (max_objects > MAX_OBJS_PER_PAGE)
1013  		max_objects = MAX_OBJS_PER_PAGE;
1014  
1015  	if (page->objects != max_objects) {
1016  		slab_err(s, page, "Wrong number of objects. Found %d but should be %d",
1017  			 page->objects, max_objects);
1018  		page->objects = max_objects;
1019  		slab_fix(s, "Number of objects adjusted.");
1020  	}
1021  	if (page->inuse != page->objects - nr) {
1022  		slab_err(s, page, "Wrong object count. Counter is %d but counted were %d",
1023  			 page->inuse, page->objects - nr);
1024  		page->inuse = page->objects - nr;
1025  		slab_fix(s, "Object count adjusted.");
1026  	}
1027  	return search == NULL;
1028  }
1029  
1030  static void trace(struct kmem_cache *s, struct page *page, void *object,
1031  								int alloc)
1032  {
1033  	if (s->flags & SLAB_TRACE) {
1034  		pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1035  			s->name,
1036  			alloc ? "alloc" : "free",
1037  			object, page->inuse,
1038  			page->freelist);
1039  
1040  		if (!alloc)
1041  			print_section(KERN_INFO, "Object ", (void *)object,
1042  					s->object_size);
1043  
1044  		dump_stack();
1045  	}
1046  }
1047  
1048  /*
1049   * Tracking of fully allocated slabs for debugging purposes.
1050   */
1051  static void add_full(struct kmem_cache *s,
1052  	struct kmem_cache_node *n, struct page *page)
1053  {
1054  	if (!(s->flags & SLAB_STORE_USER))
1055  		return;
1056  
1057  	lockdep_assert_held(&n->list_lock);
1058  	list_add(&page->slab_list, &n->full);
1059  }
1060  
1061  static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
1062  {
1063  	if (!(s->flags & SLAB_STORE_USER))
1064  		return;
1065  
1066  	lockdep_assert_held(&n->list_lock);
1067  	list_del(&page->slab_list);
1068  }
1069  
1070  /* Tracking of the number of slabs for debugging purposes */
1071  static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1072  {
1073  	struct kmem_cache_node *n = get_node(s, node);
1074  
1075  	return atomic_long_read(&n->nr_slabs);
1076  }
1077  
1078  static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1079  {
1080  	return atomic_long_read(&n->nr_slabs);
1081  }
1082  
1083  static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1084  {
1085  	struct kmem_cache_node *n = get_node(s, node);
1086  
1087  	/*
1088  	 * May be called early in order to allocate a slab for the
1089  	 * kmem_cache_node structure. Solve the chicken-egg
1090  	 * dilemma by deferring the increment of the count during
1091  	 * bootstrap (see early_kmem_cache_node_alloc).
1092  	 */
1093  	if (likely(n)) {
1094  		atomic_long_inc(&n->nr_slabs);
1095  		atomic_long_add(objects, &n->total_objects);
1096  	}
1097  }
1098  static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1099  {
1100  	struct kmem_cache_node *n = get_node(s, node);
1101  
1102  	atomic_long_dec(&n->nr_slabs);
1103  	atomic_long_sub(objects, &n->total_objects);
1104  }
1105  
1106  /* Object debug checks for alloc/free paths */
1107  static void setup_object_debug(struct kmem_cache *s, struct page *page,
1108  								void *object)
1109  {
1110  	if (!kmem_cache_debug_flags(s, SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))
1111  		return;
1112  
1113  	init_object(s, object, SLUB_RED_INACTIVE);
1114  	init_tracking(s, object);
1115  }
1116  
1117  static
1118  void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr)
1119  {
1120  	if (!kmem_cache_debug_flags(s, SLAB_POISON))
1121  		return;
1122  
1123  	metadata_access_enable();
1124  	memset(kasan_reset_tag(addr), POISON_INUSE, page_size(page));
1125  	metadata_access_disable();
1126  }
1127  
1128  static inline int alloc_consistency_checks(struct kmem_cache *s,
1129  					struct page *page, void *object)
1130  {
1131  	if (!check_slab(s, page))
1132  		return 0;
1133  
1134  	if (!check_valid_pointer(s, page, object)) {
1135  		object_err(s, page, object, "Freelist Pointer check fails");
1136  		return 0;
1137  	}
1138  
1139  	if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1140  		return 0;
1141  
1142  	return 1;
1143  }
1144  
1145  static noinline int alloc_debug_processing(struct kmem_cache *s,
1146  					struct page *page,
1147  					void *object, unsigned long addr)
1148  {
1149  	if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1150  		if (!alloc_consistency_checks(s, page, object))
1151  			goto bad;
1152  	}
1153  
1154  	/* Success perform special debug activities for allocs */
1155  	if (s->flags & SLAB_STORE_USER)
1156  		set_track(s, object, TRACK_ALLOC, addr);
1157  	trace(s, page, object, 1);
1158  	init_object(s, object, SLUB_RED_ACTIVE);
1159  	return 1;
1160  
1161  bad:
1162  	if (PageSlab(page)) {
1163  		/*
1164  		 * If this is a slab page then lets do the best we can
1165  		 * to avoid issues in the future. Marking all objects
1166  		 * as used avoids touching the remaining objects.
1167  		 */
1168  		slab_fix(s, "Marking all objects used");
1169  		page->inuse = page->objects;
1170  		page->freelist = NULL;
1171  	}
1172  	return 0;
1173  }
1174  
1175  static inline int free_consistency_checks(struct kmem_cache *s,
1176  		struct page *page, void *object, unsigned long addr)
1177  {
1178  	if (!check_valid_pointer(s, page, object)) {
1179  		slab_err(s, page, "Invalid object pointer 0x%p", object);
1180  		return 0;
1181  	}
1182  
1183  	if (on_freelist(s, page, object)) {
1184  		object_err(s, page, object, "Object already free");
1185  		return 0;
1186  	}
1187  
1188  	if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1189  		return 0;
1190  
1191  	if (unlikely(s != page->slab_cache)) {
1192  		if (!PageSlab(page)) {
1193  			slab_err(s, page, "Attempt to free object(0x%p) outside of slab",
1194  				 object);
1195  		} else if (!page->slab_cache) {
1196  			pr_err("SLUB <none>: no slab for object 0x%p.\n",
1197  			       object);
1198  			dump_stack();
1199  		} else
1200  			object_err(s, page, object,
1201  					"page slab pointer corrupt.");
1202  		return 0;
1203  	}
1204  	return 1;
1205  }
1206  
1207  /* Supports checking bulk free of a constructed freelist */
1208  static noinline int free_debug_processing(
1209  	struct kmem_cache *s, struct page *page,
1210  	void *head, void *tail, int bulk_cnt,
1211  	unsigned long addr)
1212  {
1213  	struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1214  	void *object = head;
1215  	int cnt = 0;
1216  	unsigned long flags;
1217  	int ret = 0;
1218  
1219  	spin_lock_irqsave(&n->list_lock, flags);
1220  	slab_lock(page);
1221  
1222  	if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1223  		if (!check_slab(s, page))
1224  			goto out;
1225  	}
1226  
1227  next_object:
1228  	cnt++;
1229  
1230  	if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1231  		if (!free_consistency_checks(s, page, object, addr))
1232  			goto out;
1233  	}
1234  
1235  	if (s->flags & SLAB_STORE_USER)
1236  		set_track(s, object, TRACK_FREE, addr);
1237  	trace(s, page, object, 0);
1238  	/* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1239  	init_object(s, object, SLUB_RED_INACTIVE);
1240  
1241  	/* Reached end of constructed freelist yet? */
1242  	if (object != tail) {
1243  		object = get_freepointer(s, object);
1244  		goto next_object;
1245  	}
1246  	ret = 1;
1247  
1248  out:
1249  	if (cnt != bulk_cnt)
1250  		slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1251  			 bulk_cnt, cnt);
1252  
1253  	slab_unlock(page);
1254  	spin_unlock_irqrestore(&n->list_lock, flags);
1255  	if (!ret)
1256  		slab_fix(s, "Object at 0x%p not freed", object);
1257  	return ret;
1258  }
1259  
1260  /*
1261   * Parse a block of slub_debug options. Blocks are delimited by ';'
1262   *
1263   * @str:    start of block
1264   * @flags:  returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified
1265   * @slabs:  return start of list of slabs, or NULL when there's no list
1266   * @init:   assume this is initial parsing and not per-kmem-create parsing
1267   *
1268   * returns the start of next block if there's any, or NULL
1269   */
1270  static char *
1271  parse_slub_debug_flags(char *str, slab_flags_t *flags, char **slabs, bool init)
1272  {
1273  	bool higher_order_disable = false;
1274  
1275  	/* Skip any completely empty blocks */
1276  	while (*str && *str == ';')
1277  		str++;
1278  
1279  	if (*str == ',') {
1280  		/*
1281  		 * No options but restriction on slabs. This means full
1282  		 * debugging for slabs matching a pattern.
1283  		 */
1284  		*flags = DEBUG_DEFAULT_FLAGS;
1285  		goto check_slabs;
1286  	}
1287  	*flags = 0;
1288  
1289  	/* Determine which debug features should be switched on */
1290  	for (; *str && *str != ',' && *str != ';'; str++) {
1291  		switch (tolower(*str)) {
1292  		case '-':
1293  			*flags = 0;
1294  			break;
1295  		case 'f':
1296  			*flags |= SLAB_CONSISTENCY_CHECKS;
1297  			break;
1298  		case 'z':
1299  			*flags |= SLAB_RED_ZONE;
1300  			break;
1301  		case 'p':
1302  			*flags |= SLAB_POISON;
1303  			break;
1304  		case 'u':
1305  			*flags |= SLAB_STORE_USER;
1306  			break;
1307  		case 't':
1308  			*flags |= SLAB_TRACE;
1309  			break;
1310  		case 'a':
1311  			*flags |= SLAB_FAILSLAB;
1312  			break;
1313  		case 'o':
1314  			/*
1315  			 * Avoid enabling debugging on caches if its minimum
1316  			 * order would increase as a result.
1317  			 */
1318  			higher_order_disable = true;
1319  			break;
1320  		default:
1321  			if (init)
1322  				pr_err("slub_debug option '%c' unknown. skipped\n", *str);
1323  		}
1324  	}
1325  check_slabs:
1326  	if (*str == ',')
1327  		*slabs = ++str;
1328  	else
1329  		*slabs = NULL;
1330  
1331  	/* Skip over the slab list */
1332  	while (*str && *str != ';')
1333  		str++;
1334  
1335  	/* Skip any completely empty blocks */
1336  	while (*str && *str == ';')
1337  		str++;
1338  
1339  	if (init && higher_order_disable)
1340  		disable_higher_order_debug = 1;
1341  
1342  	if (*str)
1343  		return str;
1344  	else
1345  		return NULL;
1346  }
1347  
1348  static int __init setup_slub_debug(char *str)
1349  {
1350  	slab_flags_t flags;
1351  	char *saved_str;
1352  	char *slab_list;
1353  	bool global_slub_debug_changed = false;
1354  	bool slab_list_specified = false;
1355  
1356  	slub_debug = DEBUG_DEFAULT_FLAGS;
1357  	if (*str++ != '=' || !*str)
1358  		/*
1359  		 * No options specified. Switch on full debugging.
1360  		 */
1361  		goto out;
1362  
1363  	saved_str = str;
1364  	while (str) {
1365  		str = parse_slub_debug_flags(str, &flags, &slab_list, true);
1366  
1367  		if (!slab_list) {
1368  			slub_debug = flags;
1369  			global_slub_debug_changed = true;
1370  		} else {
1371  			slab_list_specified = true;
1372  		}
1373  	}
1374  
1375  	/*
1376  	 * For backwards compatibility, a single list of flags with list of
1377  	 * slabs means debugging is only enabled for those slabs, so the global
1378  	 * slub_debug should be 0. We can extended that to multiple lists as
1379  	 * long as there is no option specifying flags without a slab list.
1380  	 */
1381  	if (slab_list_specified) {
1382  		if (!global_slub_debug_changed)
1383  			slub_debug = 0;
1384  		slub_debug_string = saved_str;
1385  	}
1386  out:
1387  	if (slub_debug != 0 || slub_debug_string)
1388  		static_branch_enable(&slub_debug_enabled);
1389  	if ((static_branch_unlikely(&init_on_alloc) ||
1390  	     static_branch_unlikely(&init_on_free)) &&
1391  	    (slub_debug & SLAB_POISON))
1392  		pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1393  	return 1;
1394  }
1395  
1396  __setup("slub_debug", setup_slub_debug);
1397  
1398  /*
1399   * kmem_cache_flags - apply debugging options to the cache
1400   * @object_size:	the size of an object without meta data
1401   * @flags:		flags to set
1402   * @name:		name of the cache
1403   * @ctor:		constructor function
1404   *
1405   * Debug option(s) are applied to @flags. In addition to the debug
1406   * option(s), if a slab name (or multiple) is specified i.e.
1407   * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1408   * then only the select slabs will receive the debug option(s).
1409   */
1410  slab_flags_t kmem_cache_flags(unsigned int object_size,
1411  	slab_flags_t flags, const char *name,
1412  	void (*ctor)(void *))
1413  {
1414  	char *iter;
1415  	size_t len;
1416  	char *next_block;
1417  	slab_flags_t block_flags;
1418  
1419  	len = strlen(name);
1420  	next_block = slub_debug_string;
1421  	/* Go through all blocks of debug options, see if any matches our slab's name */
1422  	while (next_block) {
1423  		next_block = parse_slub_debug_flags(next_block, &block_flags, &iter, false);
1424  		if (!iter)
1425  			continue;
1426  		/* Found a block that has a slab list, search it */
1427  		while (*iter) {
1428  			char *end, *glob;
1429  			size_t cmplen;
1430  
1431  			end = strchrnul(iter, ',');
1432  			if (next_block && next_block < end)
1433  				end = next_block - 1;
1434  
1435  			glob = strnchr(iter, end - iter, '*');
1436  			if (glob)
1437  				cmplen = glob - iter;
1438  			else
1439  				cmplen = max_t(size_t, len, (end - iter));
1440  
1441  			if (!strncmp(name, iter, cmplen)) {
1442  				flags |= block_flags;
1443  				return flags;
1444  			}
1445  
1446  			if (!*end || *end == ';')
1447  				break;
1448  			iter = end + 1;
1449  		}
1450  	}
1451  
1452  	return flags | slub_debug;
1453  }
1454  #else /* !CONFIG_SLUB_DEBUG */
1455  static inline void setup_object_debug(struct kmem_cache *s,
1456  			struct page *page, void *object) {}
1457  static inline
1458  void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr) {}
1459  
1460  static inline int alloc_debug_processing(struct kmem_cache *s,
1461  	struct page *page, void *object, unsigned long addr) { return 0; }
1462  
1463  static inline int free_debug_processing(
1464  	struct kmem_cache *s, struct page *page,
1465  	void *head, void *tail, int bulk_cnt,
1466  	unsigned long addr) { return 0; }
1467  
1468  static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1469  			{ return 1; }
1470  static inline int check_object(struct kmem_cache *s, struct page *page,
1471  			void *object, u8 val) { return 1; }
1472  static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1473  					struct page *page) {}
1474  static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1475  					struct page *page) {}
1476  slab_flags_t kmem_cache_flags(unsigned int object_size,
1477  	slab_flags_t flags, const char *name,
1478  	void (*ctor)(void *))
1479  {
1480  	return flags;
1481  }
1482  #define slub_debug 0
1483  
1484  #define disable_higher_order_debug 0
1485  
1486  static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1487  							{ return 0; }
1488  static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1489  							{ return 0; }
1490  static inline void inc_slabs_node(struct kmem_cache *s, int node,
1491  							int objects) {}
1492  static inline void dec_slabs_node(struct kmem_cache *s, int node,
1493  							int objects) {}
1494  
1495  static bool freelist_corrupted(struct kmem_cache *s, struct page *page,
1496  			       void **freelist, void *nextfree)
1497  {
1498  	return false;
1499  }
1500  #endif /* CONFIG_SLUB_DEBUG */
1501  
1502  /*
1503   * Hooks for other subsystems that check memory allocations. In a typical
1504   * production configuration these hooks all should produce no code at all.
1505   */
1506  static inline void *kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1507  {
1508  	ptr = kasan_kmalloc_large(ptr, size, flags);
1509  	/* As ptr might get tagged, call kmemleak hook after KASAN. */
1510  	kmemleak_alloc(ptr, size, 1, flags);
1511  	return ptr;
1512  }
1513  
1514  static __always_inline void kfree_hook(void *x)
1515  {
1516  	kmemleak_free(x);
1517  	kasan_kfree_large(x, _RET_IP_);
1518  }
1519  
1520  static __always_inline bool slab_free_hook(struct kmem_cache *s, void *x)
1521  {
1522  	kmemleak_free_recursive(x, s->flags);
1523  
1524  	/*
1525  	 * Trouble is that we may no longer disable interrupts in the fast path
1526  	 * So in order to make the debug calls that expect irqs to be
1527  	 * disabled we need to disable interrupts temporarily.
1528  	 */
1529  #ifdef CONFIG_LOCKDEP
1530  	{
1531  		unsigned long flags;
1532  
1533  		local_irq_save(flags);
1534  		debug_check_no_locks_freed(x, s->object_size);
1535  		local_irq_restore(flags);
1536  	}
1537  #endif
1538  	if (!(s->flags & SLAB_DEBUG_OBJECTS))
1539  		debug_check_no_obj_freed(x, s->object_size);
1540  
1541  	/* Use KCSAN to help debug racy use-after-free. */
1542  	if (!(s->flags & SLAB_TYPESAFE_BY_RCU))
1543  		__kcsan_check_access(x, s->object_size,
1544  				     KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT);
1545  
1546  	/* KASAN might put x into memory quarantine, delaying its reuse */
1547  	return kasan_slab_free(s, x, _RET_IP_);
1548  }
1549  
1550  static inline bool slab_free_freelist_hook(struct kmem_cache *s,
1551  					   void **head, void **tail)
1552  {
1553  
1554  	void *object;
1555  	void *next = *head;
1556  	void *old_tail = *tail ? *tail : *head;
1557  	int rsize;
1558  
1559  	/* Head and tail of the reconstructed freelist */
1560  	*head = NULL;
1561  	*tail = NULL;
1562  
1563  	do {
1564  		object = next;
1565  		next = get_freepointer(s, object);
1566  
1567  		if (slab_want_init_on_free(s)) {
1568  			/*
1569  			 * Clear the object and the metadata, but don't touch
1570  			 * the redzone.
1571  			 */
1572  			memset(kasan_reset_tag(object), 0, s->object_size);
1573  			rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad
1574  							   : 0;
1575  			memset((char *)kasan_reset_tag(object) + s->inuse, 0,
1576  			       s->size - s->inuse - rsize);
1577  
1578  		}
1579  		/* If object's reuse doesn't have to be delayed */
1580  		if (!slab_free_hook(s, object)) {
1581  			/* Move object to the new freelist */
1582  			set_freepointer(s, object, *head);
1583  			*head = object;
1584  			if (!*tail)
1585  				*tail = object;
1586  		}
1587  	} while (object != old_tail);
1588  
1589  	if (*head == *tail)
1590  		*tail = NULL;
1591  
1592  	return *head != NULL;
1593  }
1594  
1595  static void *setup_object(struct kmem_cache *s, struct page *page,
1596  				void *object)
1597  {
1598  	setup_object_debug(s, page, object);
1599  	object = kasan_init_slab_obj(s, object);
1600  	if (unlikely(s->ctor)) {
1601  		kasan_unpoison_object_data(s, object);
1602  		s->ctor(object);
1603  		kasan_poison_object_data(s, object);
1604  	}
1605  	return object;
1606  }
1607  
1608  /*
1609   * Slab allocation and freeing
1610   */
1611  static inline struct page *alloc_slab_page(struct kmem_cache *s,
1612  		gfp_t flags, int node, struct kmem_cache_order_objects oo)
1613  {
1614  	struct page *page;
1615  	unsigned int order = oo_order(oo);
1616  
1617  	if (node == NUMA_NO_NODE)
1618  		page = alloc_pages(flags, order);
1619  	else
1620  		page = __alloc_pages_node(node, flags, order);
1621  
1622  	if (page)
1623  		account_slab_page(page, order, s);
1624  
1625  	return page;
1626  }
1627  
1628  #ifdef CONFIG_SLAB_FREELIST_RANDOM
1629  /* Pre-initialize the random sequence cache */
1630  static int init_cache_random_seq(struct kmem_cache *s)
1631  {
1632  	unsigned int count = oo_objects(s->oo);
1633  	int err;
1634  
1635  	/* Bailout if already initialised */
1636  	if (s->random_seq)
1637  		return 0;
1638  
1639  	err = cache_random_seq_create(s, count, GFP_KERNEL);
1640  	if (err) {
1641  		pr_err("SLUB: Unable to initialize free list for %s\n",
1642  			s->name);
1643  		return err;
1644  	}
1645  
1646  	/* Transform to an offset on the set of pages */
1647  	if (s->random_seq) {
1648  		unsigned int i;
1649  
1650  		for (i = 0; i < count; i++)
1651  			s->random_seq[i] *= s->size;
1652  	}
1653  	return 0;
1654  }
1655  
1656  /* Initialize each random sequence freelist per cache */
1657  static void __init init_freelist_randomization(void)
1658  {
1659  	struct kmem_cache *s;
1660  
1661  	mutex_lock(&slab_mutex);
1662  
1663  	list_for_each_entry(s, &slab_caches, list)
1664  		init_cache_random_seq(s);
1665  
1666  	mutex_unlock(&slab_mutex);
1667  }
1668  
1669  /* Get the next entry on the pre-computed freelist randomized */
1670  static void *next_freelist_entry(struct kmem_cache *s, struct page *page,
1671  				unsigned long *pos, void *start,
1672  				unsigned long page_limit,
1673  				unsigned long freelist_count)
1674  {
1675  	unsigned int idx;
1676  
1677  	/*
1678  	 * If the target page allocation failed, the number of objects on the
1679  	 * page might be smaller than the usual size defined by the cache.
1680  	 */
1681  	do {
1682  		idx = s->random_seq[*pos];
1683  		*pos += 1;
1684  		if (*pos >= freelist_count)
1685  			*pos = 0;
1686  	} while (unlikely(idx >= page_limit));
1687  
1688  	return (char *)start + idx;
1689  }
1690  
1691  /* Shuffle the single linked freelist based on a random pre-computed sequence */
1692  static bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1693  {
1694  	void *start;
1695  	void *cur;
1696  	void *next;
1697  	unsigned long idx, pos, page_limit, freelist_count;
1698  
1699  	if (page->objects < 2 || !s->random_seq)
1700  		return false;
1701  
1702  	freelist_count = oo_objects(s->oo);
1703  	pos = get_random_int() % freelist_count;
1704  
1705  	page_limit = page->objects * s->size;
1706  	start = fixup_red_left(s, page_address(page));
1707  
1708  	/* First entry is used as the base of the freelist */
1709  	cur = next_freelist_entry(s, page, &pos, start, page_limit,
1710  				freelist_count);
1711  	cur = setup_object(s, page, cur);
1712  	page->freelist = cur;
1713  
1714  	for (idx = 1; idx < page->objects; idx++) {
1715  		next = next_freelist_entry(s, page, &pos, start, page_limit,
1716  			freelist_count);
1717  		next = setup_object(s, page, next);
1718  		set_freepointer(s, cur, next);
1719  		cur = next;
1720  	}
1721  	set_freepointer(s, cur, NULL);
1722  
1723  	return true;
1724  }
1725  #else
1726  static inline int init_cache_random_seq(struct kmem_cache *s)
1727  {
1728  	return 0;
1729  }
1730  static inline void init_freelist_randomization(void) { }
1731  static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1732  {
1733  	return false;
1734  }
1735  #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1736  
1737  static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1738  {
1739  	struct page *page;
1740  	struct kmem_cache_order_objects oo = s->oo;
1741  	gfp_t alloc_gfp;
1742  	void *start, *p, *next;
1743  	int idx;
1744  	bool shuffle;
1745  
1746  	flags &= gfp_allowed_mask;
1747  
1748  	if (gfpflags_allow_blocking(flags))
1749  		local_irq_enable();
1750  
1751  	flags |= s->allocflags;
1752  
1753  	/*
1754  	 * Let the initial higher-order allocation fail under memory pressure
1755  	 * so we fall-back to the minimum order allocation.
1756  	 */
1757  	alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1758  	if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1759  		alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
1760  
1761  	page = alloc_slab_page(s, alloc_gfp, node, oo);
1762  	if (unlikely(!page)) {
1763  		oo = s->min;
1764  		alloc_gfp = flags;
1765  		/*
1766  		 * Allocation may have failed due to fragmentation.
1767  		 * Try a lower order alloc if possible
1768  		 */
1769  		page = alloc_slab_page(s, alloc_gfp, node, oo);
1770  		if (unlikely(!page))
1771  			goto out;
1772  		stat(s, ORDER_FALLBACK);
1773  	}
1774  
1775  	page->objects = oo_objects(oo);
1776  
1777  	page->slab_cache = s;
1778  	__SetPageSlab(page);
1779  	if (page_is_pfmemalloc(page))
1780  		SetPageSlabPfmemalloc(page);
1781  
1782  	kasan_poison_slab(page);
1783  
1784  	start = page_address(page);
1785  
1786  	setup_page_debug(s, page, start);
1787  
1788  	shuffle = shuffle_freelist(s, page);
1789  
1790  	if (!shuffle) {
1791  		start = fixup_red_left(s, start);
1792  		start = setup_object(s, page, start);
1793  		page->freelist = start;
1794  		for (idx = 0, p = start; idx < page->objects - 1; idx++) {
1795  			next = p + s->size;
1796  			next = setup_object(s, page, next);
1797  			set_freepointer(s, p, next);
1798  			p = next;
1799  		}
1800  		set_freepointer(s, p, NULL);
1801  	}
1802  
1803  	page->inuse = page->objects;
1804  	page->frozen = 1;
1805  
1806  out:
1807  	if (gfpflags_allow_blocking(flags))
1808  		local_irq_disable();
1809  	if (!page)
1810  		return NULL;
1811  
1812  	inc_slabs_node(s, page_to_nid(page), page->objects);
1813  
1814  	return page;
1815  }
1816  
1817  static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1818  {
1819  	if (unlikely(flags & GFP_SLAB_BUG_MASK))
1820  		flags = kmalloc_fix_flags(flags);
1821  
1822  	return allocate_slab(s,
1823  		flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1824  }
1825  
1826  static void __free_slab(struct kmem_cache *s, struct page *page)
1827  {
1828  	int order = compound_order(page);
1829  	int pages = 1 << order;
1830  
1831  	if (kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS)) {
1832  		void *p;
1833  
1834  		slab_pad_check(s, page);
1835  		for_each_object(p, s, page_address(page),
1836  						page->objects)
1837  			check_object(s, page, p, SLUB_RED_INACTIVE);
1838  	}
1839  
1840  	__ClearPageSlabPfmemalloc(page);
1841  	__ClearPageSlab(page);
1842  	/* In union with page->mapping where page allocator expects NULL */
1843  	page->slab_cache = NULL;
1844  	if (current->reclaim_state)
1845  		current->reclaim_state->reclaimed_slab += pages;
1846  	unaccount_slab_page(page, order, s);
1847  	__free_pages(page, order);
1848  }
1849  
1850  static void rcu_free_slab(struct rcu_head *h)
1851  {
1852  	struct page *page = container_of(h, struct page, rcu_head);
1853  
1854  	__free_slab(page->slab_cache, page);
1855  }
1856  
1857  static void free_slab(struct kmem_cache *s, struct page *page)
1858  {
1859  	if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) {
1860  		call_rcu(&page->rcu_head, rcu_free_slab);
1861  	} else
1862  		__free_slab(s, page);
1863  }
1864  
1865  static void discard_slab(struct kmem_cache *s, struct page *page)
1866  {
1867  	dec_slabs_node(s, page_to_nid(page), page->objects);
1868  	free_slab(s, page);
1869  }
1870  
1871  /*
1872   * Management of partially allocated slabs.
1873   */
1874  static inline void
1875  __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1876  {
1877  	n->nr_partial++;
1878  	if (tail == DEACTIVATE_TO_TAIL)
1879  		list_add_tail(&page->slab_list, &n->partial);
1880  	else
1881  		list_add(&page->slab_list, &n->partial);
1882  }
1883  
1884  static inline void add_partial(struct kmem_cache_node *n,
1885  				struct page *page, int tail)
1886  {
1887  	lockdep_assert_held(&n->list_lock);
1888  	__add_partial(n, page, tail);
1889  }
1890  
1891  static inline void remove_partial(struct kmem_cache_node *n,
1892  					struct page *page)
1893  {
1894  	lockdep_assert_held(&n->list_lock);
1895  	list_del(&page->slab_list);
1896  	n->nr_partial--;
1897  }
1898  
1899  /*
1900   * Remove slab from the partial list, freeze it and
1901   * return the pointer to the freelist.
1902   *
1903   * Returns a list of objects or NULL if it fails.
1904   */
1905  static inline void *acquire_slab(struct kmem_cache *s,
1906  		struct kmem_cache_node *n, struct page *page,
1907  		int mode, int *objects)
1908  {
1909  	void *freelist;
1910  	unsigned long counters;
1911  	struct page new;
1912  
1913  	lockdep_assert_held(&n->list_lock);
1914  
1915  	/*
1916  	 * Zap the freelist and set the frozen bit.
1917  	 * The old freelist is the list of objects for the
1918  	 * per cpu allocation list.
1919  	 */
1920  	freelist = page->freelist;
1921  	counters = page->counters;
1922  	new.counters = counters;
1923  	*objects = new.objects - new.inuse;
1924  	if (mode) {
1925  		new.inuse = page->objects;
1926  		new.freelist = NULL;
1927  	} else {
1928  		new.freelist = freelist;
1929  	}
1930  
1931  	VM_BUG_ON(new.frozen);
1932  	new.frozen = 1;
1933  
1934  	if (!__cmpxchg_double_slab(s, page,
1935  			freelist, counters,
1936  			new.freelist, new.counters,
1937  			"acquire_slab"))
1938  		return NULL;
1939  
1940  	remove_partial(n, page);
1941  	WARN_ON(!freelist);
1942  	return freelist;
1943  }
1944  
1945  static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1946  static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1947  
1948  /*
1949   * Try to allocate a partial slab from a specific node.
1950   */
1951  static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1952  				struct kmem_cache_cpu *c, gfp_t flags)
1953  {
1954  	struct page *page, *page2;
1955  	void *object = NULL;
1956  	unsigned int available = 0;
1957  	int objects;
1958  
1959  	/*
1960  	 * Racy check. If we mistakenly see no partial slabs then we
1961  	 * just allocate an empty slab. If we mistakenly try to get a
1962  	 * partial slab and there is none available then get_partial()
1963  	 * will return NULL.
1964  	 */
1965  	if (!n || !n->nr_partial)
1966  		return NULL;
1967  
1968  	spin_lock(&n->list_lock);
1969  	list_for_each_entry_safe(page, page2, &n->partial, slab_list) {
1970  		void *t;
1971  
1972  		if (!pfmemalloc_match(page, flags))
1973  			continue;
1974  
1975  		t = acquire_slab(s, n, page, object == NULL, &objects);
1976  		if (!t)
1977  			break;
1978  
1979  		available += objects;
1980  		if (!object) {
1981  			c->page = page;
1982  			stat(s, ALLOC_FROM_PARTIAL);
1983  			object = t;
1984  		} else {
1985  			put_cpu_partial(s, page, 0);
1986  			stat(s, CPU_PARTIAL_NODE);
1987  		}
1988  		if (!kmem_cache_has_cpu_partial(s)
1989  			|| available > slub_cpu_partial(s) / 2)
1990  			break;
1991  
1992  	}
1993  	spin_unlock(&n->list_lock);
1994  	return object;
1995  }
1996  
1997  /*
1998   * Get a page from somewhere. Search in increasing NUMA distances.
1999   */
2000  static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
2001  		struct kmem_cache_cpu *c)
2002  {
2003  #ifdef CONFIG_NUMA
2004  	struct zonelist *zonelist;
2005  	struct zoneref *z;
2006  	struct zone *zone;
2007  	enum zone_type highest_zoneidx = gfp_zone(flags);
2008  	void *object;
2009  	unsigned int cpuset_mems_cookie;
2010  
2011  	/*
2012  	 * The defrag ratio allows a configuration of the tradeoffs between
2013  	 * inter node defragmentation and node local allocations. A lower
2014  	 * defrag_ratio increases the tendency to do local allocations
2015  	 * instead of attempting to obtain partial slabs from other nodes.
2016  	 *
2017  	 * If the defrag_ratio is set to 0 then kmalloc() always
2018  	 * returns node local objects. If the ratio is higher then kmalloc()
2019  	 * may return off node objects because partial slabs are obtained
2020  	 * from other nodes and filled up.
2021  	 *
2022  	 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
2023  	 * (which makes defrag_ratio = 1000) then every (well almost)
2024  	 * allocation will first attempt to defrag slab caches on other nodes.
2025  	 * This means scanning over all nodes to look for partial slabs which
2026  	 * may be expensive if we do it every time we are trying to find a slab
2027  	 * with available objects.
2028  	 */
2029  	if (!s->remote_node_defrag_ratio ||
2030  			get_cycles() % 1024 > s->remote_node_defrag_ratio)
2031  		return NULL;
2032  
2033  	do {
2034  		cpuset_mems_cookie = read_mems_allowed_begin();
2035  		zonelist = node_zonelist(mempolicy_slab_node(), flags);
2036  		for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) {
2037  			struct kmem_cache_node *n;
2038  
2039  			n = get_node(s, zone_to_nid(zone));
2040  
2041  			if (n && cpuset_zone_allowed(zone, flags) &&
2042  					n->nr_partial > s->min_partial) {
2043  				object = get_partial_node(s, n, c, flags);
2044  				if (object) {
2045  					/*
2046  					 * Don't check read_mems_allowed_retry()
2047  					 * here - if mems_allowed was updated in
2048  					 * parallel, that was a harmless race
2049  					 * between allocation and the cpuset
2050  					 * update
2051  					 */
2052  					return object;
2053  				}
2054  			}
2055  		}
2056  	} while (read_mems_allowed_retry(cpuset_mems_cookie));
2057  #endif	/* CONFIG_NUMA */
2058  	return NULL;
2059  }
2060  
2061  /*
2062   * Get a partial page, lock it and return it.
2063   */
2064  static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
2065  		struct kmem_cache_cpu *c)
2066  {
2067  	void *object;
2068  	int searchnode = node;
2069  
2070  	if (node == NUMA_NO_NODE)
2071  		searchnode = numa_mem_id();
2072  
2073  	object = get_partial_node(s, get_node(s, searchnode), c, flags);
2074  	if (object || node != NUMA_NO_NODE)
2075  		return object;
2076  
2077  	return get_any_partial(s, flags, c);
2078  }
2079  
2080  #ifdef CONFIG_PREEMPTION
2081  /*
2082   * Calculate the next globally unique transaction for disambiguation
2083   * during cmpxchg. The transactions start with the cpu number and are then
2084   * incremented by CONFIG_NR_CPUS.
2085   */
2086  #define TID_STEP  roundup_pow_of_two(CONFIG_NR_CPUS)
2087  #else
2088  /*
2089   * No preemption supported therefore also no need to check for
2090   * different cpus.
2091   */
2092  #define TID_STEP 1
2093  #endif
2094  
2095  static inline unsigned long next_tid(unsigned long tid)
2096  {
2097  	return tid + TID_STEP;
2098  }
2099  
2100  #ifdef SLUB_DEBUG_CMPXCHG
2101  static inline unsigned int tid_to_cpu(unsigned long tid)
2102  {
2103  	return tid % TID_STEP;
2104  }
2105  
2106  static inline unsigned long tid_to_event(unsigned long tid)
2107  {
2108  	return tid / TID_STEP;
2109  }
2110  #endif
2111  
2112  static inline unsigned int init_tid(int cpu)
2113  {
2114  	return cpu;
2115  }
2116  
2117  static inline void note_cmpxchg_failure(const char *n,
2118  		const struct kmem_cache *s, unsigned long tid)
2119  {
2120  #ifdef SLUB_DEBUG_CMPXCHG
2121  	unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
2122  
2123  	pr_info("%s %s: cmpxchg redo ", n, s->name);
2124  
2125  #ifdef CONFIG_PREEMPTION
2126  	if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
2127  		pr_warn("due to cpu change %d -> %d\n",
2128  			tid_to_cpu(tid), tid_to_cpu(actual_tid));
2129  	else
2130  #endif
2131  	if (tid_to_event(tid) != tid_to_event(actual_tid))
2132  		pr_warn("due to cpu running other code. Event %ld->%ld\n",
2133  			tid_to_event(tid), tid_to_event(actual_tid));
2134  	else
2135  		pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2136  			actual_tid, tid, next_tid(tid));
2137  #endif
2138  	stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2139  }
2140  
2141  static void init_kmem_cache_cpus(struct kmem_cache *s)
2142  {
2143  	int cpu;
2144  
2145  	for_each_possible_cpu(cpu)
2146  		per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
2147  }
2148  
2149  /*
2150   * Remove the cpu slab
2151   */
2152  static void deactivate_slab(struct kmem_cache *s, struct page *page,
2153  				void *freelist, struct kmem_cache_cpu *c)
2154  {
2155  	enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
2156  	struct kmem_cache_node *n = get_node(s, page_to_nid(page));
2157  	int lock = 0;
2158  	enum slab_modes l = M_NONE, m = M_NONE;
2159  	void *nextfree;
2160  	int tail = DEACTIVATE_TO_HEAD;
2161  	struct page new;
2162  	struct page old;
2163  
2164  	if (page->freelist) {
2165  		stat(s, DEACTIVATE_REMOTE_FREES);
2166  		tail = DEACTIVATE_TO_TAIL;
2167  	}
2168  
2169  	/*
2170  	 * Stage one: Free all available per cpu objects back
2171  	 * to the page freelist while it is still frozen. Leave the
2172  	 * last one.
2173  	 *
2174  	 * There is no need to take the list->lock because the page
2175  	 * is still frozen.
2176  	 */
2177  	while (freelist && (nextfree = get_freepointer(s, freelist))) {
2178  		void *prior;
2179  		unsigned long counters;
2180  
2181  		/*
2182  		 * If 'nextfree' is invalid, it is possible that the object at
2183  		 * 'freelist' is already corrupted.  So isolate all objects
2184  		 * starting at 'freelist'.
2185  		 */
2186  		if (freelist_corrupted(s, page, &freelist, nextfree))
2187  			break;
2188  
2189  		do {
2190  			prior = page->freelist;
2191  			counters = page->counters;
2192  			set_freepointer(s, freelist, prior);
2193  			new.counters = counters;
2194  			new.inuse--;
2195  			VM_BUG_ON(!new.frozen);
2196  
2197  		} while (!__cmpxchg_double_slab(s, page,
2198  			prior, counters,
2199  			freelist, new.counters,
2200  			"drain percpu freelist"));
2201  
2202  		freelist = nextfree;
2203  	}
2204  
2205  	/*
2206  	 * Stage two: Ensure that the page is unfrozen while the
2207  	 * list presence reflects the actual number of objects
2208  	 * during unfreeze.
2209  	 *
2210  	 * We setup the list membership and then perform a cmpxchg
2211  	 * with the count. If there is a mismatch then the page
2212  	 * is not unfrozen but the page is on the wrong list.
2213  	 *
2214  	 * Then we restart the process which may have to remove
2215  	 * the page from the list that we just put it on again
2216  	 * because the number of objects in the slab may have
2217  	 * changed.
2218  	 */
2219  redo:
2220  
2221  	old.freelist = page->freelist;
2222  	old.counters = page->counters;
2223  	VM_BUG_ON(!old.frozen);
2224  
2225  	/* Determine target state of the slab */
2226  	new.counters = old.counters;
2227  	if (freelist) {
2228  		new.inuse--;
2229  		set_freepointer(s, freelist, old.freelist);
2230  		new.freelist = freelist;
2231  	} else
2232  		new.freelist = old.freelist;
2233  
2234  	new.frozen = 0;
2235  
2236  	if (!new.inuse && n->nr_partial >= s->min_partial)
2237  		m = M_FREE;
2238  	else if (new.freelist) {
2239  		m = M_PARTIAL;
2240  		if (!lock) {
2241  			lock = 1;
2242  			/*
2243  			 * Taking the spinlock removes the possibility
2244  			 * that acquire_slab() will see a slab page that
2245  			 * is frozen
2246  			 */
2247  			spin_lock(&n->list_lock);
2248  		}
2249  	} else {
2250  		m = M_FULL;
2251  		if (kmem_cache_debug_flags(s, SLAB_STORE_USER) && !lock) {
2252  			lock = 1;
2253  			/*
2254  			 * This also ensures that the scanning of full
2255  			 * slabs from diagnostic functions will not see
2256  			 * any frozen slabs.
2257  			 */
2258  			spin_lock(&n->list_lock);
2259  		}
2260  	}
2261  
2262  	if (l != m) {
2263  		if (l == M_PARTIAL)
2264  			remove_partial(n, page);
2265  		else if (l == M_FULL)
2266  			remove_full(s, n, page);
2267  
2268  		if (m == M_PARTIAL)
2269  			add_partial(n, page, tail);
2270  		else if (m == M_FULL)
2271  			add_full(s, n, page);
2272  	}
2273  
2274  	l = m;
2275  	if (!__cmpxchg_double_slab(s, page,
2276  				old.freelist, old.counters,
2277  				new.freelist, new.counters,
2278  				"unfreezing slab"))
2279  		goto redo;
2280  
2281  	if (lock)
2282  		spin_unlock(&n->list_lock);
2283  
2284  	if (m == M_PARTIAL)
2285  		stat(s, tail);
2286  	else if (m == M_FULL)
2287  		stat(s, DEACTIVATE_FULL);
2288  	else if (m == M_FREE) {
2289  		stat(s, DEACTIVATE_EMPTY);
2290  		discard_slab(s, page);
2291  		stat(s, FREE_SLAB);
2292  	}
2293  
2294  	c->page = NULL;
2295  	c->freelist = NULL;
2296  }
2297  
2298  /*
2299   * Unfreeze all the cpu partial slabs.
2300   *
2301   * This function must be called with interrupts disabled
2302   * for the cpu using c (or some other guarantee must be there
2303   * to guarantee no concurrent accesses).
2304   */
2305  static void unfreeze_partials(struct kmem_cache *s,
2306  		struct kmem_cache_cpu *c)
2307  {
2308  #ifdef CONFIG_SLUB_CPU_PARTIAL
2309  	struct kmem_cache_node *n = NULL, *n2 = NULL;
2310  	struct page *page, *discard_page = NULL;
2311  
2312  	while ((page = slub_percpu_partial(c))) {
2313  		struct page new;
2314  		struct page old;
2315  
2316  		slub_set_percpu_partial(c, page);
2317  
2318  		n2 = get_node(s, page_to_nid(page));
2319  		if (n != n2) {
2320  			if (n)
2321  				spin_unlock(&n->list_lock);
2322  
2323  			n = n2;
2324  			spin_lock(&n->list_lock);
2325  		}
2326  
2327  		do {
2328  
2329  			old.freelist = page->freelist;
2330  			old.counters = page->counters;
2331  			VM_BUG_ON(!old.frozen);
2332  
2333  			new.counters = old.counters;
2334  			new.freelist = old.freelist;
2335  
2336  			new.frozen = 0;
2337  
2338  		} while (!__cmpxchg_double_slab(s, page,
2339  				old.freelist, old.counters,
2340  				new.freelist, new.counters,
2341  				"unfreezing slab"));
2342  
2343  		if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2344  			page->next = discard_page;
2345  			discard_page = page;
2346  		} else {
2347  			add_partial(n, page, DEACTIVATE_TO_TAIL);
2348  			stat(s, FREE_ADD_PARTIAL);
2349  		}
2350  	}
2351  
2352  	if (n)
2353  		spin_unlock(&n->list_lock);
2354  
2355  	while (discard_page) {
2356  		page = discard_page;
2357  		discard_page = discard_page->next;
2358  
2359  		stat(s, DEACTIVATE_EMPTY);
2360  		discard_slab(s, page);
2361  		stat(s, FREE_SLAB);
2362  	}
2363  #endif	/* CONFIG_SLUB_CPU_PARTIAL */
2364  }
2365  
2366  /*
2367   * Put a page that was just frozen (in __slab_free|get_partial_node) into a
2368   * partial page slot if available.
2369   *
2370   * If we did not find a slot then simply move all the partials to the
2371   * per node partial list.
2372   */
2373  static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2374  {
2375  #ifdef CONFIG_SLUB_CPU_PARTIAL
2376  	struct page *oldpage;
2377  	int pages;
2378  	int pobjects;
2379  
2380  	preempt_disable();
2381  	do {
2382  		pages = 0;
2383  		pobjects = 0;
2384  		oldpage = this_cpu_read(s->cpu_slab->partial);
2385  
2386  		if (oldpage) {
2387  			pobjects = oldpage->pobjects;
2388  			pages = oldpage->pages;
2389  			if (drain && pobjects > slub_cpu_partial(s)) {
2390  				unsigned long flags;
2391  				/*
2392  				 * partial array is full. Move the existing
2393  				 * set to the per node partial list.
2394  				 */
2395  				local_irq_save(flags);
2396  				unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2397  				local_irq_restore(flags);
2398  				oldpage = NULL;
2399  				pobjects = 0;
2400  				pages = 0;
2401  				stat(s, CPU_PARTIAL_DRAIN);
2402  			}
2403  		}
2404  
2405  		pages++;
2406  		pobjects += page->objects - page->inuse;
2407  
2408  		page->pages = pages;
2409  		page->pobjects = pobjects;
2410  		page->next = oldpage;
2411  
2412  	} while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2413  								!= oldpage);
2414  	if (unlikely(!slub_cpu_partial(s))) {
2415  		unsigned long flags;
2416  
2417  		local_irq_save(flags);
2418  		unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2419  		local_irq_restore(flags);
2420  	}
2421  	preempt_enable();
2422  #endif	/* CONFIG_SLUB_CPU_PARTIAL */
2423  }
2424  
2425  static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2426  {
2427  	stat(s, CPUSLAB_FLUSH);
2428  	deactivate_slab(s, c->page, c->freelist, c);
2429  
2430  	c->tid = next_tid(c->tid);
2431  }
2432  
2433  /*
2434   * Flush cpu slab.
2435   *
2436   * Called from IPI handler with interrupts disabled.
2437   */
2438  static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2439  {
2440  	struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2441  
2442  	if (c->page)
2443  		flush_slab(s, c);
2444  
2445  	unfreeze_partials(s, c);
2446  }
2447  
2448  static void flush_cpu_slab(void *d)
2449  {
2450  	struct kmem_cache *s = d;
2451  
2452  	__flush_cpu_slab(s, smp_processor_id());
2453  }
2454  
2455  static bool has_cpu_slab(int cpu, void *info)
2456  {
2457  	struct kmem_cache *s = info;
2458  	struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2459  
2460  	return c->page || slub_percpu_partial(c);
2461  }
2462  
2463  static void flush_all(struct kmem_cache *s)
2464  {
2465  	on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1);
2466  }
2467  
2468  /*
2469   * Use the cpu notifier to insure that the cpu slabs are flushed when
2470   * necessary.
2471   */
2472  static int slub_cpu_dead(unsigned int cpu)
2473  {
2474  	struct kmem_cache *s;
2475  	unsigned long flags;
2476  
2477  	mutex_lock(&slab_mutex);
2478  	list_for_each_entry(s, &slab_caches, list) {
2479  		local_irq_save(flags);
2480  		__flush_cpu_slab(s, cpu);
2481  		local_irq_restore(flags);
2482  	}
2483  	mutex_unlock(&slab_mutex);
2484  	return 0;
2485  }
2486  
2487  /*
2488   * Check if the objects in a per cpu structure fit numa
2489   * locality expectations.
2490   */
2491  static inline int node_match(struct page *page, int node)
2492  {
2493  #ifdef CONFIG_NUMA
2494  	if (node != NUMA_NO_NODE && page_to_nid(page) != node)
2495  		return 0;
2496  #endif
2497  	return 1;
2498  }
2499  
2500  #ifdef CONFIG_SLUB_DEBUG
2501  static int count_free(struct page *page)
2502  {
2503  	return page->objects - page->inuse;
2504  }
2505  
2506  static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2507  {
2508  	return atomic_long_read(&n->total_objects);
2509  }
2510  #endif /* CONFIG_SLUB_DEBUG */
2511  
2512  #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2513  static unsigned long count_partial(struct kmem_cache_node *n,
2514  					int (*get_count)(struct page *))
2515  {
2516  	unsigned long flags;
2517  	unsigned long x = 0;
2518  	struct page *page;
2519  
2520  	spin_lock_irqsave(&n->list_lock, flags);
2521  	list_for_each_entry(page, &n->partial, slab_list)
2522  		x += get_count(page);
2523  	spin_unlock_irqrestore(&n->list_lock, flags);
2524  	return x;
2525  }
2526  #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2527  
2528  static noinline void
2529  slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2530  {
2531  #ifdef CONFIG_SLUB_DEBUG
2532  	static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2533  				      DEFAULT_RATELIMIT_BURST);
2534  	int node;
2535  	struct kmem_cache_node *n;
2536  
2537  	if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2538  		return;
2539  
2540  	pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2541  		nid, gfpflags, &gfpflags);
2542  	pr_warn("  cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2543  		s->name, s->object_size, s->size, oo_order(s->oo),
2544  		oo_order(s->min));
2545  
2546  	if (oo_order(s->min) > get_order(s->object_size))
2547  		pr_warn("  %s debugging increased min order, use slub_debug=O to disable.\n",
2548  			s->name);
2549  
2550  	for_each_kmem_cache_node(s, node, n) {
2551  		unsigned long nr_slabs;
2552  		unsigned long nr_objs;
2553  		unsigned long nr_free;
2554  
2555  		nr_free  = count_partial(n, count_free);
2556  		nr_slabs = node_nr_slabs(n);
2557  		nr_objs  = node_nr_objs(n);
2558  
2559  		pr_warn("  node %d: slabs: %ld, objs: %ld, free: %ld\n",
2560  			node, nr_slabs, nr_objs, nr_free);
2561  	}
2562  #endif
2563  }
2564  
2565  static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2566  			int node, struct kmem_cache_cpu **pc)
2567  {
2568  	void *freelist;
2569  	struct kmem_cache_cpu *c = *pc;
2570  	struct page *page;
2571  
2572  	WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
2573  
2574  	freelist = get_partial(s, flags, node, c);
2575  
2576  	if (freelist)
2577  		return freelist;
2578  
2579  	page = new_slab(s, flags, node);
2580  	if (page) {
2581  		c = raw_cpu_ptr(s->cpu_slab);
2582  		if (c->page)
2583  			flush_slab(s, c);
2584  
2585  		/*
2586  		 * No other reference to the page yet so we can
2587  		 * muck around with it freely without cmpxchg
2588  		 */
2589  		freelist = page->freelist;
2590  		page->freelist = NULL;
2591  
2592  		stat(s, ALLOC_SLAB);
2593  		c->page = page;
2594  		*pc = c;
2595  	}
2596  
2597  	return freelist;
2598  }
2599  
2600  static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2601  {
2602  	if (unlikely(PageSlabPfmemalloc(page)))
2603  		return gfp_pfmemalloc_allowed(gfpflags);
2604  
2605  	return true;
2606  }
2607  
2608  /*
2609   * Check the page->freelist of a page and either transfer the freelist to the
2610   * per cpu freelist or deactivate the page.
2611   *
2612   * The page is still frozen if the return value is not NULL.
2613   *
2614   * If this function returns NULL then the page has been unfrozen.
2615   *
2616   * This function must be called with interrupt disabled.
2617   */
2618  static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2619  {
2620  	struct page new;
2621  	unsigned long counters;
2622  	void *freelist;
2623  
2624  	do {
2625  		freelist = page->freelist;
2626  		counters = page->counters;
2627  
2628  		new.counters = counters;
2629  		VM_BUG_ON(!new.frozen);
2630  
2631  		new.inuse = page->objects;
2632  		new.frozen = freelist != NULL;
2633  
2634  	} while (!__cmpxchg_double_slab(s, page,
2635  		freelist, counters,
2636  		NULL, new.counters,
2637  		"get_freelist"));
2638  
2639  	return freelist;
2640  }
2641  
2642  /*
2643   * Slow path. The lockless freelist is empty or we need to perform
2644   * debugging duties.
2645   *
2646   * Processing is still very fast if new objects have been freed to the
2647   * regular freelist. In that case we simply take over the regular freelist
2648   * as the lockless freelist and zap the regular freelist.
2649   *
2650   * If that is not working then we fall back to the partial lists. We take the
2651   * first element of the freelist as the object to allocate now and move the
2652   * rest of the freelist to the lockless freelist.
2653   *
2654   * And if we were unable to get a new slab from the partial slab lists then
2655   * we need to allocate a new slab. This is the slowest path since it involves
2656   * a call to the page allocator and the setup of a new slab.
2657   *
2658   * Version of __slab_alloc to use when we know that interrupts are
2659   * already disabled (which is the case for bulk allocation).
2660   */
2661  static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2662  			  unsigned long addr, struct kmem_cache_cpu *c)
2663  {
2664  	void *freelist;
2665  	struct page *page;
2666  
2667  	stat(s, ALLOC_SLOWPATH);
2668  
2669  	page = c->page;
2670  	if (!page) {
2671  		/*
2672  		 * if the node is not online or has no normal memory, just
2673  		 * ignore the node constraint
2674  		 */
2675  		if (unlikely(node != NUMA_NO_NODE &&
2676  			     !node_state(node, N_NORMAL_MEMORY)))
2677  			node = NUMA_NO_NODE;
2678  		goto new_slab;
2679  	}
2680  redo:
2681  
2682  	if (unlikely(!node_match(page, node))) {
2683  		/*
2684  		 * same as above but node_match() being false already
2685  		 * implies node != NUMA_NO_NODE
2686  		 */
2687  		if (!node_state(node, N_NORMAL_MEMORY)) {
2688  			node = NUMA_NO_NODE;
2689  			goto redo;
2690  		} else {
2691  			stat(s, ALLOC_NODE_MISMATCH);
2692  			deactivate_slab(s, page, c->freelist, c);
2693  			goto new_slab;
2694  		}
2695  	}
2696  
2697  	/*
2698  	 * By rights, we should be searching for a slab page that was
2699  	 * PFMEMALLOC but right now, we are losing the pfmemalloc
2700  	 * information when the page leaves the per-cpu allocator
2701  	 */
2702  	if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2703  		deactivate_slab(s, page, c->freelist, c);
2704  		goto new_slab;
2705  	}
2706  
2707  	/* must check again c->freelist in case of cpu migration or IRQ */
2708  	freelist = c->freelist;
2709  	if (freelist)
2710  		goto load_freelist;
2711  
2712  	freelist = get_freelist(s, page);
2713  
2714  	if (!freelist) {
2715  		c->page = NULL;
2716  		stat(s, DEACTIVATE_BYPASS);
2717  		goto new_slab;
2718  	}
2719  
2720  	stat(s, ALLOC_REFILL);
2721  
2722  load_freelist:
2723  	/*
2724  	 * freelist is pointing to the list of objects to be used.
2725  	 * page is pointing to the page from which the objects are obtained.
2726  	 * That page must be frozen for per cpu allocations to work.
2727  	 */
2728  	VM_BUG_ON(!c->page->frozen);
2729  	c->freelist = get_freepointer(s, freelist);
2730  	c->tid = next_tid(c->tid);
2731  	return freelist;
2732  
2733  new_slab:
2734  
2735  	if (slub_percpu_partial(c)) {
2736  		page = c->page = slub_percpu_partial(c);
2737  		slub_set_percpu_partial(c, page);
2738  		stat(s, CPU_PARTIAL_ALLOC);
2739  		goto redo;
2740  	}
2741  
2742  	freelist = new_slab_objects(s, gfpflags, node, &c);
2743  
2744  	if (unlikely(!freelist)) {
2745  		slab_out_of_memory(s, gfpflags, node);
2746  		return NULL;
2747  	}
2748  
2749  	page = c->page;
2750  	if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2751  		goto load_freelist;
2752  
2753  	/* Only entered in the debug case */
2754  	if (kmem_cache_debug(s) &&
2755  			!alloc_debug_processing(s, page, freelist, addr))
2756  		goto new_slab;	/* Slab failed checks. Next slab needed */
2757  
2758  	deactivate_slab(s, page, get_freepointer(s, freelist), c);
2759  	return freelist;
2760  }
2761  
2762  /*
2763   * Another one that disabled interrupt and compensates for possible
2764   * cpu changes by refetching the per cpu area pointer.
2765   */
2766  static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2767  			  unsigned long addr, struct kmem_cache_cpu *c)
2768  {
2769  	void *p;
2770  	unsigned long flags;
2771  
2772  	local_irq_save(flags);
2773  #ifdef CONFIG_PREEMPTION
2774  	/*
2775  	 * We may have been preempted and rescheduled on a different
2776  	 * cpu before disabling interrupts. Need to reload cpu area
2777  	 * pointer.
2778  	 */
2779  	c = this_cpu_ptr(s->cpu_slab);
2780  #endif
2781  
2782  	p = ___slab_alloc(s, gfpflags, node, addr, c);
2783  	local_irq_restore(flags);
2784  	return p;
2785  }
2786  
2787  /*
2788   * If the object has been wiped upon free, make sure it's fully initialized by
2789   * zeroing out freelist pointer.
2790   */
2791  static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s,
2792  						   void *obj)
2793  {
2794  	if (unlikely(slab_want_init_on_free(s)) && obj)
2795  		memset((void *)((char *)obj + s->offset), 0, sizeof(void *));
2796  }
2797  
2798  /*
2799   * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2800   * have the fastpath folded into their functions. So no function call
2801   * overhead for requests that can be satisfied on the fastpath.
2802   *
2803   * The fastpath works by first checking if the lockless freelist can be used.
2804   * If not then __slab_alloc is called for slow processing.
2805   *
2806   * Otherwise we can simply pick the next object from the lockless free list.
2807   */
2808  static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2809  		gfp_t gfpflags, int node, unsigned long addr)
2810  {
2811  	void *object;
2812  	struct kmem_cache_cpu *c;
2813  	struct page *page;
2814  	unsigned long tid;
2815  	struct obj_cgroup *objcg = NULL;
2816  
2817  	s = slab_pre_alloc_hook(s, &objcg, 1, gfpflags);
2818  	if (!s)
2819  		return NULL;
2820  redo:
2821  	/*
2822  	 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2823  	 * enabled. We may switch back and forth between cpus while
2824  	 * reading from one cpu area. That does not matter as long
2825  	 * as we end up on the original cpu again when doing the cmpxchg.
2826  	 *
2827  	 * We should guarantee that tid and kmem_cache are retrieved on
2828  	 * the same cpu. It could be different if CONFIG_PREEMPTION so we need
2829  	 * to check if it is matched or not.
2830  	 */
2831  	do {
2832  		tid = this_cpu_read(s->cpu_slab->tid);
2833  		c = raw_cpu_ptr(s->cpu_slab);
2834  	} while (IS_ENABLED(CONFIG_PREEMPTION) &&
2835  		 unlikely(tid != READ_ONCE(c->tid)));
2836  
2837  	/*
2838  	 * Irqless object alloc/free algorithm used here depends on sequence
2839  	 * of fetching cpu_slab's data. tid should be fetched before anything
2840  	 * on c to guarantee that object and page associated with previous tid
2841  	 * won't be used with current tid. If we fetch tid first, object and
2842  	 * page could be one associated with next tid and our alloc/free
2843  	 * request will be failed. In this case, we will retry. So, no problem.
2844  	 */
2845  	barrier();
2846  
2847  	/*
2848  	 * The transaction ids are globally unique per cpu and per operation on
2849  	 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2850  	 * occurs on the right processor and that there was no operation on the
2851  	 * linked list in between.
2852  	 */
2853  
2854  	object = c->freelist;
2855  	page = c->page;
2856  	if (unlikely(!object || !page || !node_match(page, node))) {
2857  		object = __slab_alloc(s, gfpflags, node, addr, c);
2858  	} else {
2859  		void *next_object = get_freepointer_safe(s, object);
2860  
2861  		/*
2862  		 * The cmpxchg will only match if there was no additional
2863  		 * operation and if we are on the right processor.
2864  		 *
2865  		 * The cmpxchg does the following atomically (without lock
2866  		 * semantics!)
2867  		 * 1. Relocate first pointer to the current per cpu area.
2868  		 * 2. Verify that tid and freelist have not been changed
2869  		 * 3. If they were not changed replace tid and freelist
2870  		 *
2871  		 * Since this is without lock semantics the protection is only
2872  		 * against code executing on this cpu *not* from access by
2873  		 * other cpus.
2874  		 */
2875  		if (unlikely(!this_cpu_cmpxchg_double(
2876  				s->cpu_slab->freelist, s->cpu_slab->tid,
2877  				object, tid,
2878  				next_object, next_tid(tid)))) {
2879  
2880  			note_cmpxchg_failure("slab_alloc", s, tid);
2881  			goto redo;
2882  		}
2883  		prefetch_freepointer(s, next_object);
2884  		stat(s, ALLOC_FASTPATH);
2885  	}
2886  
2887  	maybe_wipe_obj_freeptr(s, kasan_reset_tag(object));
2888  
2889  	if (unlikely(slab_want_init_on_alloc(gfpflags, s)) && object)
2890  		memset(kasan_reset_tag(object), 0, s->object_size);
2891  
2892  	slab_post_alloc_hook(s, objcg, gfpflags, 1, &object);
2893  
2894  	return object;
2895  }
2896  
2897  static __always_inline void *slab_alloc(struct kmem_cache *s,
2898  		gfp_t gfpflags, unsigned long addr)
2899  {
2900  	return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2901  }
2902  
2903  void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2904  {
2905  	void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2906  
2907  	trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2908  				s->size, gfpflags);
2909  
2910  	return ret;
2911  }
2912  EXPORT_SYMBOL(kmem_cache_alloc);
2913  
2914  #ifdef CONFIG_TRACING
2915  void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2916  {
2917  	void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2918  	trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2919  	ret = kasan_kmalloc(s, ret, size, gfpflags);
2920  	return ret;
2921  }
2922  EXPORT_SYMBOL(kmem_cache_alloc_trace);
2923  #endif
2924  
2925  #ifdef CONFIG_NUMA
2926  void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2927  {
2928  	void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2929  
2930  	trace_kmem_cache_alloc_node(_RET_IP_, ret,
2931  				    s->object_size, s->size, gfpflags, node);
2932  
2933  	return ret;
2934  }
2935  EXPORT_SYMBOL(kmem_cache_alloc_node);
2936  
2937  #ifdef CONFIG_TRACING
2938  void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2939  				    gfp_t gfpflags,
2940  				    int node, size_t size)
2941  {
2942  	void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2943  
2944  	trace_kmalloc_node(_RET_IP_, ret,
2945  			   size, s->size, gfpflags, node);
2946  
2947  	ret = kasan_kmalloc(s, ret, size, gfpflags);
2948  	return ret;
2949  }
2950  EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2951  #endif
2952  #endif	/* CONFIG_NUMA */
2953  
2954  /*
2955   * Slow path handling. This may still be called frequently since objects
2956   * have a longer lifetime than the cpu slabs in most processing loads.
2957   *
2958   * So we still attempt to reduce cache line usage. Just take the slab
2959   * lock and free the item. If there is no additional partial page
2960   * handling required then we can return immediately.
2961   */
2962  static void __slab_free(struct kmem_cache *s, struct page *page,
2963  			void *head, void *tail, int cnt,
2964  			unsigned long addr)
2965  
2966  {
2967  	void *prior;
2968  	int was_frozen;
2969  	struct page new;
2970  	unsigned long counters;
2971  	struct kmem_cache_node *n = NULL;
2972  	unsigned long flags;
2973  
2974  	stat(s, FREE_SLOWPATH);
2975  
2976  	if (kmem_cache_debug(s) &&
2977  	    !free_debug_processing(s, page, head, tail, cnt, addr))
2978  		return;
2979  
2980  	do {
2981  		if (unlikely(n)) {
2982  			spin_unlock_irqrestore(&n->list_lock, flags);
2983  			n = NULL;
2984  		}
2985  		prior = page->freelist;
2986  		counters = page->counters;
2987  		set_freepointer(s, tail, prior);
2988  		new.counters = counters;
2989  		was_frozen = new.frozen;
2990  		new.inuse -= cnt;
2991  		if ((!new.inuse || !prior) && !was_frozen) {
2992  
2993  			if (kmem_cache_has_cpu_partial(s) && !prior) {
2994  
2995  				/*
2996  				 * Slab was on no list before and will be
2997  				 * partially empty
2998  				 * We can defer the list move and instead
2999  				 * freeze it.
3000  				 */
3001  				new.frozen = 1;
3002  
3003  			} else { /* Needs to be taken off a list */
3004  
3005  				n = get_node(s, page_to_nid(page));
3006  				/*
3007  				 * Speculatively acquire the list_lock.
3008  				 * If the cmpxchg does not succeed then we may
3009  				 * drop the list_lock without any processing.
3010  				 *
3011  				 * Otherwise the list_lock will synchronize with
3012  				 * other processors updating the list of slabs.
3013  				 */
3014  				spin_lock_irqsave(&n->list_lock, flags);
3015  
3016  			}
3017  		}
3018  
3019  	} while (!cmpxchg_double_slab(s, page,
3020  		prior, counters,
3021  		head, new.counters,
3022  		"__slab_free"));
3023  
3024  	if (likely(!n)) {
3025  
3026  		if (likely(was_frozen)) {
3027  			/*
3028  			 * The list lock was not taken therefore no list
3029  			 * activity can be necessary.
3030  			 */
3031  			stat(s, FREE_FROZEN);
3032  		} else if (new.frozen) {
3033  			/*
3034  			 * If we just froze the page then put it onto the
3035  			 * per cpu partial list.
3036  			 */
3037  			put_cpu_partial(s, page, 1);
3038  			stat(s, CPU_PARTIAL_FREE);
3039  		}
3040  
3041  		return;
3042  	}
3043  
3044  	if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
3045  		goto slab_empty;
3046  
3047  	/*
3048  	 * Objects left in the slab. If it was not on the partial list before
3049  	 * then add it.
3050  	 */
3051  	if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
3052  		remove_full(s, n, page);
3053  		add_partial(n, page, DEACTIVATE_TO_TAIL);
3054  		stat(s, FREE_ADD_PARTIAL);
3055  	}
3056  	spin_unlock_irqrestore(&n->list_lock, flags);
3057  	return;
3058  
3059  slab_empty:
3060  	if (prior) {
3061  		/*
3062  		 * Slab on the partial list.
3063  		 */
3064  		remove_partial(n, page);
3065  		stat(s, FREE_REMOVE_PARTIAL);
3066  	} else {
3067  		/* Slab must be on the full list */
3068  		remove_full(s, n, page);
3069  	}
3070  
3071  	spin_unlock_irqrestore(&n->list_lock, flags);
3072  	stat(s, FREE_SLAB);
3073  	discard_slab(s, page);
3074  }
3075  
3076  /*
3077   * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
3078   * can perform fastpath freeing without additional function calls.
3079   *
3080   * The fastpath is only possible if we are freeing to the current cpu slab
3081   * of this processor. This typically the case if we have just allocated
3082   * the item before.
3083   *
3084   * If fastpath is not possible then fall back to __slab_free where we deal
3085   * with all sorts of special processing.
3086   *
3087   * Bulk free of a freelist with several objects (all pointing to the
3088   * same page) possible by specifying head and tail ptr, plus objects
3089   * count (cnt). Bulk free indicated by tail pointer being set.
3090   */
3091  static __always_inline void do_slab_free(struct kmem_cache *s,
3092  				struct page *page, void *head, void *tail,
3093  				int cnt, unsigned long addr)
3094  {
3095  	void *tail_obj = tail ? : head;
3096  	struct kmem_cache_cpu *c;
3097  	unsigned long tid;
3098  
3099  	memcg_slab_free_hook(s, &head, 1);
3100  redo:
3101  	/*
3102  	 * Determine the currently cpus per cpu slab.
3103  	 * The cpu may change afterward. However that does not matter since
3104  	 * data is retrieved via this pointer. If we are on the same cpu
3105  	 * during the cmpxchg then the free will succeed.
3106  	 */
3107  	do {
3108  		tid = this_cpu_read(s->cpu_slab->tid);
3109  		c = raw_cpu_ptr(s->cpu_slab);
3110  	} while (IS_ENABLED(CONFIG_PREEMPTION) &&
3111  		 unlikely(tid != READ_ONCE(c->tid)));
3112  
3113  	/* Same with comment on barrier() in slab_alloc_node() */
3114  	barrier();
3115  
3116  	if (likely(page == c->page)) {
3117  		void **freelist = READ_ONCE(c->freelist);
3118  
3119  		set_freepointer(s, tail_obj, freelist);
3120  
3121  		if (unlikely(!this_cpu_cmpxchg_double(
3122  				s->cpu_slab->freelist, s->cpu_slab->tid,
3123  				freelist, tid,
3124  				head, next_tid(tid)))) {
3125  
3126  			note_cmpxchg_failure("slab_free", s, tid);
3127  			goto redo;
3128  		}
3129  		stat(s, FREE_FASTPATH);
3130  	} else
3131  		__slab_free(s, page, head, tail_obj, cnt, addr);
3132  
3133  }
3134  
3135  static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
3136  				      void *head, void *tail, int cnt,
3137  				      unsigned long addr)
3138  {
3139  	/*
3140  	 * With KASAN enabled slab_free_freelist_hook modifies the freelist
3141  	 * to remove objects, whose reuse must be delayed.
3142  	 */
3143  	if (slab_free_freelist_hook(s, &head, &tail))
3144  		do_slab_free(s, page, head, tail, cnt, addr);
3145  }
3146  
3147  #ifdef CONFIG_KASAN_GENERIC
3148  void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
3149  {
3150  	do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr);
3151  }
3152  #endif
3153  
3154  void kmem_cache_free(struct kmem_cache *s, void *x)
3155  {
3156  	s = cache_from_obj(s, x);
3157  	if (!s)
3158  		return;
3159  	slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
3160  	trace_kmem_cache_free(_RET_IP_, x);
3161  }
3162  EXPORT_SYMBOL(kmem_cache_free);
3163  
3164  struct detached_freelist {
3165  	struct page *page;
3166  	void *tail;
3167  	void *freelist;
3168  	int cnt;
3169  	struct kmem_cache *s;
3170  };
3171  
3172  /*
3173   * This function progressively scans the array with free objects (with
3174   * a limited look ahead) and extract objects belonging to the same
3175   * page.  It builds a detached freelist directly within the given
3176   * page/objects.  This can happen without any need for
3177   * synchronization, because the objects are owned by running process.
3178   * The freelist is build up as a single linked list in the objects.
3179   * The idea is, that this detached freelist can then be bulk
3180   * transferred to the real freelist(s), but only requiring a single
3181   * synchronization primitive.  Look ahead in the array is limited due
3182   * to performance reasons.
3183   */
3184  static inline
3185  int build_detached_freelist(struct kmem_cache *s, size_t size,
3186  			    void **p, struct detached_freelist *df)
3187  {
3188  	size_t first_skipped_index = 0;
3189  	int lookahead = 3;
3190  	void *object;
3191  	struct page *page;
3192  
3193  	/* Always re-init detached_freelist */
3194  	df->page = NULL;
3195  
3196  	do {
3197  		object = p[--size];
3198  		/* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3199  	} while (!object && size);
3200  
3201  	if (!object)
3202  		return 0;
3203  
3204  	page = virt_to_head_page(object);
3205  	if (!s) {
3206  		/* Handle kalloc'ed objects */
3207  		if (unlikely(!PageSlab(page))) {
3208  			BUG_ON(!PageCompound(page));
3209  			kfree_hook(object);
3210  			__free_pages(page, compound_order(page));
3211  			p[size] = NULL; /* mark object processed */
3212  			return size;
3213  		}
3214  		/* Derive kmem_cache from object */
3215  		df->s = page->slab_cache;
3216  	} else {
3217  		df->s = cache_from_obj(s, object); /* Support for memcg */
3218  	}
3219  
3220  	/* Start new detached freelist */
3221  	df->page = page;
3222  	set_freepointer(df->s, object, NULL);
3223  	df->tail = object;
3224  	df->freelist = object;
3225  	p[size] = NULL; /* mark object processed */
3226  	df->cnt = 1;
3227  
3228  	while (size) {
3229  		object = p[--size];
3230  		if (!object)
3231  			continue; /* Skip processed objects */
3232  
3233  		/* df->page is always set at this point */
3234  		if (df->page == virt_to_head_page(object)) {
3235  			/* Opportunity build freelist */
3236  			set_freepointer(df->s, object, df->freelist);
3237  			df->freelist = object;
3238  			df->cnt++;
3239  			p[size] = NULL; /* mark object processed */
3240  
3241  			continue;
3242  		}
3243  
3244  		/* Limit look ahead search */
3245  		if (!--lookahead)
3246  			break;
3247  
3248  		if (!first_skipped_index)
3249  			first_skipped_index = size + 1;
3250  	}
3251  
3252  	return first_skipped_index;
3253  }
3254  
3255  /* Note that interrupts must be enabled when calling this function. */
3256  void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3257  {
3258  	if (WARN_ON(!size))
3259  		return;
3260  
3261  	memcg_slab_free_hook(s, p, size);
3262  	do {
3263  		struct detached_freelist df;
3264  
3265  		size = build_detached_freelist(s, size, p, &df);
3266  		if (!df.page)
3267  			continue;
3268  
3269  		slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
3270  	} while (likely(size));
3271  }
3272  EXPORT_SYMBOL(kmem_cache_free_bulk);
3273  
3274  /* Note that interrupts must be enabled when calling this function. */
3275  int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3276  			  void **p)
3277  {
3278  	struct kmem_cache_cpu *c;
3279  	int i;
3280  	struct obj_cgroup *objcg = NULL;
3281  
3282  	/* memcg and kmem_cache debug support */
3283  	s = slab_pre_alloc_hook(s, &objcg, size, flags);
3284  	if (unlikely(!s))
3285  		return false;
3286  	/*
3287  	 * Drain objects in the per cpu slab, while disabling local
3288  	 * IRQs, which protects against PREEMPT and interrupts
3289  	 * handlers invoking normal fastpath.
3290  	 */
3291  	local_irq_disable();
3292  	c = this_cpu_ptr(s->cpu_slab);
3293  
3294  	for (i = 0; i < size; i++) {
3295  		void *object = c->freelist;
3296  
3297  		if (unlikely(!object)) {
3298  			/*
3299  			 * We may have removed an object from c->freelist using
3300  			 * the fastpath in the previous iteration; in that case,
3301  			 * c->tid has not been bumped yet.
3302  			 * Since ___slab_alloc() may reenable interrupts while
3303  			 * allocating memory, we should bump c->tid now.
3304  			 */
3305  			c->tid = next_tid(c->tid);
3306  
3307  			/*
3308  			 * Invoking slow path likely have side-effect
3309  			 * of re-populating per CPU c->freelist
3310  			 */
3311  			p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3312  					    _RET_IP_, c);
3313  			if (unlikely(!p[i]))
3314  				goto error;
3315  
3316  			c = this_cpu_ptr(s->cpu_slab);
3317  			maybe_wipe_obj_freeptr(s, p[i]);
3318  
3319  			continue; /* goto for-loop */
3320  		}
3321  		c->freelist = get_freepointer(s, object);
3322  		p[i] = object;
3323  		maybe_wipe_obj_freeptr(s, p[i]);
3324  	}
3325  	c->tid = next_tid(c->tid);
3326  	local_irq_enable();
3327  
3328  	/* Clear memory outside IRQ disabled fastpath loop */
3329  	if (unlikely(slab_want_init_on_alloc(flags, s))) {
3330  		int j;
3331  
3332  		for (j = 0; j < i; j++)
3333  			memset(p[j], 0, s->object_size);
3334  	}
3335  
3336  	/* memcg and kmem_cache debug support */
3337  	slab_post_alloc_hook(s, objcg, flags, size, p);
3338  	return i;
3339  error:
3340  	local_irq_enable();
3341  	slab_post_alloc_hook(s, objcg, flags, i, p);
3342  	__kmem_cache_free_bulk(s, i, p);
3343  	return 0;
3344  }
3345  EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3346  
3347  
3348  /*
3349   * Object placement in a slab is made very easy because we always start at
3350   * offset 0. If we tune the size of the object to the alignment then we can
3351   * get the required alignment by putting one properly sized object after
3352   * another.
3353   *
3354   * Notice that the allocation order determines the sizes of the per cpu
3355   * caches. Each processor has always one slab available for allocations.
3356   * Increasing the allocation order reduces the number of times that slabs
3357   * must be moved on and off the partial lists and is therefore a factor in
3358   * locking overhead.
3359   */
3360  
3361  /*
3362   * Mininum / Maximum order of slab pages. This influences locking overhead
3363   * and slab fragmentation. A higher order reduces the number of partial slabs
3364   * and increases the number of allocations possible without having to
3365   * take the list_lock.
3366   */
3367  static unsigned int slub_min_order;
3368  static unsigned int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3369  static unsigned int slub_min_objects;
3370  
3371  /*
3372   * Calculate the order of allocation given an slab object size.
3373   *
3374   * The order of allocation has significant impact on performance and other
3375   * system components. Generally order 0 allocations should be preferred since
3376   * order 0 does not cause fragmentation in the page allocator. Larger objects
3377   * be problematic to put into order 0 slabs because there may be too much
3378   * unused space left. We go to a higher order if more than 1/16th of the slab
3379   * would be wasted.
3380   *
3381   * In order to reach satisfactory performance we must ensure that a minimum
3382   * number of objects is in one slab. Otherwise we may generate too much
3383   * activity on the partial lists which requires taking the list_lock. This is
3384   * less a concern for large slabs though which are rarely used.
3385   *
3386   * slub_max_order specifies the order where we begin to stop considering the
3387   * number of objects in a slab as critical. If we reach slub_max_order then
3388   * we try to keep the page order as low as possible. So we accept more waste
3389   * of space in favor of a small page order.
3390   *
3391   * Higher order allocations also allow the placement of more objects in a
3392   * slab and thereby reduce object handling overhead. If the user has
3393   * requested a higher mininum order then we start with that one instead of
3394   * the smallest order which will fit the object.
3395   */
3396  static inline unsigned int slab_order(unsigned int size,
3397  		unsigned int min_objects, unsigned int max_order,
3398  		unsigned int fract_leftover)
3399  {
3400  	unsigned int min_order = slub_min_order;
3401  	unsigned int order;
3402  
3403  	if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
3404  		return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3405  
3406  	for (order = max(min_order, (unsigned int)get_order(min_objects * size));
3407  			order <= max_order; order++) {
3408  
3409  		unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
3410  		unsigned int rem;
3411  
3412  		rem = slab_size % size;
3413  
3414  		if (rem <= slab_size / fract_leftover)
3415  			break;
3416  	}
3417  
3418  	return order;
3419  }
3420  
3421  static inline int calculate_order(unsigned int size)
3422  {
3423  	unsigned int order;
3424  	unsigned int min_objects;
3425  	unsigned int max_objects;
3426  
3427  	/*
3428  	 * Attempt to find best configuration for a slab. This
3429  	 * works by first attempting to generate a layout with
3430  	 * the best configuration and backing off gradually.
3431  	 *
3432  	 * First we increase the acceptable waste in a slab. Then
3433  	 * we reduce the minimum objects required in a slab.
3434  	 */
3435  	min_objects = slub_min_objects;
3436  	if (!min_objects)
3437  		min_objects = 4 * (fls(num_online_cpus()) + 1);
3438  	max_objects = order_objects(slub_max_order, size);
3439  	min_objects = min(min_objects, max_objects);
3440  
3441  	while (min_objects > 1) {
3442  		unsigned int fraction;
3443  
3444  		fraction = 16;
3445  		while (fraction >= 4) {
3446  			order = slab_order(size, min_objects,
3447  					slub_max_order, fraction);
3448  			if (order <= slub_max_order)
3449  				return order;
3450  			fraction /= 2;
3451  		}
3452  		min_objects--;
3453  	}
3454  
3455  	/*
3456  	 * We were unable to place multiple objects in a slab. Now
3457  	 * lets see if we can place a single object there.
3458  	 */
3459  	order = slab_order(size, 1, slub_max_order, 1);
3460  	if (order <= slub_max_order)
3461  		return order;
3462  
3463  	/*
3464  	 * Doh this slab cannot be placed using slub_max_order.
3465  	 */
3466  	order = slab_order(size, 1, MAX_ORDER, 1);
3467  	if (order < MAX_ORDER)
3468  		return order;
3469  	return -ENOSYS;
3470  }
3471  
3472  static void
3473  init_kmem_cache_node(struct kmem_cache_node *n)
3474  {
3475  	n->nr_partial = 0;
3476  	spin_lock_init(&n->list_lock);
3477  	INIT_LIST_HEAD(&n->partial);
3478  #ifdef CONFIG_SLUB_DEBUG
3479  	atomic_long_set(&n->nr_slabs, 0);
3480  	atomic_long_set(&n->total_objects, 0);
3481  	INIT_LIST_HEAD(&n->full);
3482  #endif
3483  }
3484  
3485  static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3486  {
3487  	BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3488  			KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3489  
3490  	/*
3491  	 * Must align to double word boundary for the double cmpxchg
3492  	 * instructions to work; see __pcpu_double_call_return_bool().
3493  	 */
3494  	s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3495  				     2 * sizeof(void *));
3496  
3497  	if (!s->cpu_slab)
3498  		return 0;
3499  
3500  	init_kmem_cache_cpus(s);
3501  
3502  	return 1;
3503  }
3504  
3505  static struct kmem_cache *kmem_cache_node;
3506  
3507  /*
3508   * No kmalloc_node yet so do it by hand. We know that this is the first
3509   * slab on the node for this slabcache. There are no concurrent accesses
3510   * possible.
3511   *
3512   * Note that this function only works on the kmem_cache_node
3513   * when allocating for the kmem_cache_node. This is used for bootstrapping
3514   * memory on a fresh node that has no slab structures yet.
3515   */
3516  static void early_kmem_cache_node_alloc(int node)
3517  {
3518  	struct page *page;
3519  	struct kmem_cache_node *n;
3520  
3521  	BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3522  
3523  	page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3524  
3525  	BUG_ON(!page);
3526  	if (page_to_nid(page) != node) {
3527  		pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3528  		pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3529  	}
3530  
3531  	n = page->freelist;
3532  	BUG_ON(!n);
3533  #ifdef CONFIG_SLUB_DEBUG
3534  	init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3535  	init_tracking(kmem_cache_node, n);
3536  #endif
3537  	n = kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node),
3538  		      GFP_KERNEL);
3539  	page->freelist = get_freepointer(kmem_cache_node, n);
3540  	page->inuse = 1;
3541  	page->frozen = 0;
3542  	kmem_cache_node->node[node] = n;
3543  	init_kmem_cache_node(n);
3544  	inc_slabs_node(kmem_cache_node, node, page->objects);
3545  
3546  	/*
3547  	 * No locks need to be taken here as it has just been
3548  	 * initialized and there is no concurrent access.
3549  	 */
3550  	__add_partial(n, page, DEACTIVATE_TO_HEAD);
3551  }
3552  
3553  static void free_kmem_cache_nodes(struct kmem_cache *s)
3554  {
3555  	int node;
3556  	struct kmem_cache_node *n;
3557  
3558  	for_each_kmem_cache_node(s, node, n) {
3559  		s->node[node] = NULL;
3560  		kmem_cache_free(kmem_cache_node, n);
3561  	}
3562  }
3563  
3564  void __kmem_cache_release(struct kmem_cache *s)
3565  {
3566  	cache_random_seq_destroy(s);
3567  	free_percpu(s->cpu_slab);
3568  	free_kmem_cache_nodes(s);
3569  }
3570  
3571  static int init_kmem_cache_nodes(struct kmem_cache *s)
3572  {
3573  	int node;
3574  
3575  	for_each_node_state(node, N_NORMAL_MEMORY) {
3576  		struct kmem_cache_node *n;
3577  
3578  		if (slab_state == DOWN) {
3579  			early_kmem_cache_node_alloc(node);
3580  			continue;
3581  		}
3582  		n = kmem_cache_alloc_node(kmem_cache_node,
3583  						GFP_KERNEL, node);
3584  
3585  		if (!n) {
3586  			free_kmem_cache_nodes(s);
3587  			return 0;
3588  		}
3589  
3590  		init_kmem_cache_node(n);
3591  		s->node[node] = n;
3592  	}
3593  	return 1;
3594  }
3595  
3596  static void set_min_partial(struct kmem_cache *s, unsigned long min)
3597  {
3598  	if (min < MIN_PARTIAL)
3599  		min = MIN_PARTIAL;
3600  	else if (min > MAX_PARTIAL)
3601  		min = MAX_PARTIAL;
3602  	s->min_partial = min;
3603  }
3604  
3605  static void set_cpu_partial(struct kmem_cache *s)
3606  {
3607  #ifdef CONFIG_SLUB_CPU_PARTIAL
3608  	/*
3609  	 * cpu_partial determined the maximum number of objects kept in the
3610  	 * per cpu partial lists of a processor.
3611  	 *
3612  	 * Per cpu partial lists mainly contain slabs that just have one
3613  	 * object freed. If they are used for allocation then they can be
3614  	 * filled up again with minimal effort. The slab will never hit the
3615  	 * per node partial lists and therefore no locking will be required.
3616  	 *
3617  	 * This setting also determines
3618  	 *
3619  	 * A) The number of objects from per cpu partial slabs dumped to the
3620  	 *    per node list when we reach the limit.
3621  	 * B) The number of objects in cpu partial slabs to extract from the
3622  	 *    per node list when we run out of per cpu objects. We only fetch
3623  	 *    50% to keep some capacity around for frees.
3624  	 */
3625  	if (!kmem_cache_has_cpu_partial(s))
3626  		slub_set_cpu_partial(s, 0);
3627  	else if (s->size >= PAGE_SIZE)
3628  		slub_set_cpu_partial(s, 2);
3629  	else if (s->size >= 1024)
3630  		slub_set_cpu_partial(s, 6);
3631  	else if (s->size >= 256)
3632  		slub_set_cpu_partial(s, 13);
3633  	else
3634  		slub_set_cpu_partial(s, 30);
3635  #endif
3636  }
3637  
3638  /*
3639   * calculate_sizes() determines the order and the distribution of data within
3640   * a slab object.
3641   */
3642  static int calculate_sizes(struct kmem_cache *s, int forced_order)
3643  {
3644  	slab_flags_t flags = s->flags;
3645  	unsigned int size = s->object_size;
3646  	unsigned int freepointer_area;
3647  	unsigned int order;
3648  
3649  	/*
3650  	 * Round up object size to the next word boundary. We can only
3651  	 * place the free pointer at word boundaries and this determines
3652  	 * the possible location of the free pointer.
3653  	 */
3654  	size = ALIGN(size, sizeof(void *));
3655  	/*
3656  	 * This is the area of the object where a freepointer can be
3657  	 * safely written. If redzoning adds more to the inuse size, we
3658  	 * can't use that portion for writing the freepointer, so
3659  	 * s->offset must be limited within this for the general case.
3660  	 */
3661  	freepointer_area = size;
3662  
3663  #ifdef CONFIG_SLUB_DEBUG
3664  	/*
3665  	 * Determine if we can poison the object itself. If the user of
3666  	 * the slab may touch the object after free or before allocation
3667  	 * then we should never poison the object itself.
3668  	 */
3669  	if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
3670  			!s->ctor)
3671  		s->flags |= __OBJECT_POISON;
3672  	else
3673  		s->flags &= ~__OBJECT_POISON;
3674  
3675  
3676  	/*
3677  	 * If we are Redzoning then check if there is some space between the
3678  	 * end of the object and the free pointer. If not then add an
3679  	 * additional word to have some bytes to store Redzone information.
3680  	 */
3681  	if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3682  		size += sizeof(void *);
3683  #endif
3684  
3685  	/*
3686  	 * With that we have determined the number of bytes in actual use
3687  	 * by the object. This is the potential offset to the free pointer.
3688  	 */
3689  	s->inuse = size;
3690  
3691  	if (((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
3692  		s->ctor)) {
3693  		/*
3694  		 * Relocate free pointer after the object if it is not
3695  		 * permitted to overwrite the first word of the object on
3696  		 * kmem_cache_free.
3697  		 *
3698  		 * This is the case if we do RCU, have a constructor or
3699  		 * destructor or are poisoning the objects.
3700  		 *
3701  		 * The assumption that s->offset >= s->inuse means free
3702  		 * pointer is outside of the object is used in the
3703  		 * freeptr_outside_object() function. If that is no
3704  		 * longer true, the function needs to be modified.
3705  		 */
3706  		s->offset = size;
3707  		size += sizeof(void *);
3708  	} else if (freepointer_area > sizeof(void *)) {
3709  		/*
3710  		 * Store freelist pointer near middle of object to keep
3711  		 * it away from the edges of the object to avoid small
3712  		 * sized over/underflows from neighboring allocations.
3713  		 */
3714  		s->offset = ALIGN(freepointer_area / 2, sizeof(void *));
3715  	}
3716  
3717  #ifdef CONFIG_SLUB_DEBUG
3718  	if (flags & SLAB_STORE_USER)
3719  		/*
3720  		 * Need to store information about allocs and frees after
3721  		 * the object.
3722  		 */
3723  		size += 2 * sizeof(struct track);
3724  #endif
3725  
3726  	kasan_cache_create(s, &size, &s->flags);
3727  #ifdef CONFIG_SLUB_DEBUG
3728  	if (flags & SLAB_RED_ZONE) {
3729  		/*
3730  		 * Add some empty padding so that we can catch
3731  		 * overwrites from earlier objects rather than let
3732  		 * tracking information or the free pointer be
3733  		 * corrupted if a user writes before the start
3734  		 * of the object.
3735  		 */
3736  		size += sizeof(void *);
3737  
3738  		s->red_left_pad = sizeof(void *);
3739  		s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3740  		size += s->red_left_pad;
3741  	}
3742  #endif
3743  
3744  	/*
3745  	 * SLUB stores one object immediately after another beginning from
3746  	 * offset 0. In order to align the objects we have to simply size
3747  	 * each object to conform to the alignment.
3748  	 */
3749  	size = ALIGN(size, s->align);
3750  	s->size = size;
3751  	s->reciprocal_size = reciprocal_value(size);
3752  	if (forced_order >= 0)
3753  		order = forced_order;
3754  	else
3755  		order = calculate_order(size);
3756  
3757  	if ((int)order < 0)
3758  		return 0;
3759  
3760  	s->allocflags = 0;
3761  	if (order)
3762  		s->allocflags |= __GFP_COMP;
3763  
3764  	if (s->flags & SLAB_CACHE_DMA)
3765  		s->allocflags |= GFP_DMA;
3766  
3767  	if (s->flags & SLAB_CACHE_DMA32)
3768  		s->allocflags |= GFP_DMA32;
3769  
3770  	if (s->flags & SLAB_RECLAIM_ACCOUNT)
3771  		s->allocflags |= __GFP_RECLAIMABLE;
3772  
3773  	/*
3774  	 * Determine the number of objects per slab
3775  	 */
3776  	s->oo = oo_make(order, size);
3777  	s->min = oo_make(get_order(size), size);
3778  	if (oo_objects(s->oo) > oo_objects(s->max))
3779  		s->max = s->oo;
3780  
3781  	return !!oo_objects(s->oo);
3782  }
3783  
3784  static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
3785  {
3786  	s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3787  #ifdef CONFIG_SLAB_FREELIST_HARDENED
3788  	s->random = get_random_long();
3789  #endif
3790  
3791  	if (!calculate_sizes(s, -1))
3792  		goto error;
3793  	if (disable_higher_order_debug) {
3794  		/*
3795  		 * Disable debugging flags that store metadata if the min slab
3796  		 * order increased.
3797  		 */
3798  		if (get_order(s->size) > get_order(s->object_size)) {
3799  			s->flags &= ~DEBUG_METADATA_FLAGS;
3800  			s->offset = 0;
3801  			if (!calculate_sizes(s, -1))
3802  				goto error;
3803  		}
3804  	}
3805  
3806  #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3807      defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3808  	if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
3809  		/* Enable fast mode */
3810  		s->flags |= __CMPXCHG_DOUBLE;
3811  #endif
3812  
3813  	/*
3814  	 * The larger the object size is, the more pages we want on the partial
3815  	 * list to avoid pounding the page allocator excessively.
3816  	 */
3817  	set_min_partial(s, ilog2(s->size) / 2);
3818  
3819  	set_cpu_partial(s);
3820  
3821  #ifdef CONFIG_NUMA
3822  	s->remote_node_defrag_ratio = 1000;
3823  #endif
3824  
3825  	/* Initialize the pre-computed randomized freelist if slab is up */
3826  	if (slab_state >= UP) {
3827  		if (init_cache_random_seq(s))
3828  			goto error;
3829  	}
3830  
3831  	if (!init_kmem_cache_nodes(s))
3832  		goto error;
3833  
3834  	if (alloc_kmem_cache_cpus(s))
3835  		return 0;
3836  
3837  	free_kmem_cache_nodes(s);
3838  error:
3839  	return -EINVAL;
3840  }
3841  
3842  static void list_slab_objects(struct kmem_cache *s, struct page *page,
3843  			      const char *text)
3844  {
3845  #ifdef CONFIG_SLUB_DEBUG
3846  	void *addr = page_address(page);
3847  	unsigned long *map;
3848  	void *p;
3849  
3850  	slab_err(s, page, text, s->name);
3851  	slab_lock(page);
3852  
3853  	map = get_map(s, page);
3854  	for_each_object(p, s, addr, page->objects) {
3855  
3856  		if (!test_bit(__obj_to_index(s, addr, p), map)) {
3857  			pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3858  			print_tracking(s, p);
3859  		}
3860  	}
3861  	put_map(map);
3862  	slab_unlock(page);
3863  #endif
3864  }
3865  
3866  /*
3867   * Attempt to free all partial slabs on a node.
3868   * This is called from __kmem_cache_shutdown(). We must take list_lock
3869   * because sysfs file might still access partial list after the shutdowning.
3870   */
3871  static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3872  {
3873  	LIST_HEAD(discard);
3874  	struct page *page, *h;
3875  
3876  	BUG_ON(irqs_disabled());
3877  	spin_lock_irq(&n->list_lock);
3878  	list_for_each_entry_safe(page, h, &n->partial, slab_list) {
3879  		if (!page->inuse) {
3880  			remove_partial(n, page);
3881  			list_add(&page->slab_list, &discard);
3882  		} else {
3883  			list_slab_objects(s, page,
3884  			  "Objects remaining in %s on __kmem_cache_shutdown()");
3885  		}
3886  	}
3887  	spin_unlock_irq(&n->list_lock);
3888  
3889  	list_for_each_entry_safe(page, h, &discard, slab_list)
3890  		discard_slab(s, page);
3891  }
3892  
3893  bool __kmem_cache_empty(struct kmem_cache *s)
3894  {
3895  	int node;
3896  	struct kmem_cache_node *n;
3897  
3898  	for_each_kmem_cache_node(s, node, n)
3899  		if (n->nr_partial || slabs_node(s, node))
3900  			return false;
3901  	return true;
3902  }
3903  
3904  /*
3905   * Release all resources used by a slab cache.
3906   */
3907  int __kmem_cache_shutdown(struct kmem_cache *s)
3908  {
3909  	int node;
3910  	struct kmem_cache_node *n;
3911  
3912  	flush_all(s);
3913  	/* Attempt to free all objects */
3914  	for_each_kmem_cache_node(s, node, n) {
3915  		free_partial(s, n);
3916  		if (n->nr_partial || slabs_node(s, node))
3917  			return 1;
3918  	}
3919  	return 0;
3920  }
3921  
3922  /********************************************************************
3923   *		Kmalloc subsystem
3924   *******************************************************************/
3925  
3926  static int __init setup_slub_min_order(char *str)
3927  {
3928  	get_option(&str, (int *)&slub_min_order);
3929  
3930  	return 1;
3931  }
3932  
3933  __setup("slub_min_order=", setup_slub_min_order);
3934  
3935  static int __init setup_slub_max_order(char *str)
3936  {
3937  	get_option(&str, (int *)&slub_max_order);
3938  	slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1);
3939  
3940  	return 1;
3941  }
3942  
3943  __setup("slub_max_order=", setup_slub_max_order);
3944  
3945  static int __init setup_slub_min_objects(char *str)
3946  {
3947  	get_option(&str, (int *)&slub_min_objects);
3948  
3949  	return 1;
3950  }
3951  
3952  __setup("slub_min_objects=", setup_slub_min_objects);
3953  
3954  void *__kmalloc(size_t size, gfp_t flags)
3955  {
3956  	struct kmem_cache *s;
3957  	void *ret;
3958  
3959  	if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3960  		return kmalloc_large(size, flags);
3961  
3962  	s = kmalloc_slab(size, flags);
3963  
3964  	if (unlikely(ZERO_OR_NULL_PTR(s)))
3965  		return s;
3966  
3967  	ret = slab_alloc(s, flags, _RET_IP_);
3968  
3969  	trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3970  
3971  	ret = kasan_kmalloc(s, ret, size, flags);
3972  
3973  	return ret;
3974  }
3975  EXPORT_SYMBOL(__kmalloc);
3976  
3977  #ifdef CONFIG_NUMA
3978  static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3979  {
3980  	struct page *page;
3981  	void *ptr = NULL;
3982  	unsigned int order = get_order(size);
3983  
3984  	flags |= __GFP_COMP;
3985  	page = alloc_pages_node(node, flags, order);
3986  	if (page) {
3987  		ptr = page_address(page);
3988  		mod_node_page_state(page_pgdat(page), NR_SLAB_UNRECLAIMABLE_B,
3989  				    PAGE_SIZE << order);
3990  	}
3991  
3992  	return kmalloc_large_node_hook(ptr, size, flags);
3993  }
3994  
3995  void *__kmalloc_node(size_t size, gfp_t flags, int node)
3996  {
3997  	struct kmem_cache *s;
3998  	void *ret;
3999  
4000  	if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4001  		ret = kmalloc_large_node(size, flags, node);
4002  
4003  		trace_kmalloc_node(_RET_IP_, ret,
4004  				   size, PAGE_SIZE << get_order(size),
4005  				   flags, node);
4006  
4007  		return ret;
4008  	}
4009  
4010  	s = kmalloc_slab(size, flags);
4011  
4012  	if (unlikely(ZERO_OR_NULL_PTR(s)))
4013  		return s;
4014  
4015  	ret = slab_alloc_node(s, flags, node, _RET_IP_);
4016  
4017  	trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
4018  
4019  	ret = kasan_kmalloc(s, ret, size, flags);
4020  
4021  	return ret;
4022  }
4023  EXPORT_SYMBOL(__kmalloc_node);
4024  #endif	/* CONFIG_NUMA */
4025  
4026  #ifdef CONFIG_HARDENED_USERCOPY
4027  /*
4028   * Rejects incorrectly sized objects and objects that are to be copied
4029   * to/from userspace but do not fall entirely within the containing slab
4030   * cache's usercopy region.
4031   *
4032   * Returns NULL if check passes, otherwise const char * to name of cache
4033   * to indicate an error.
4034   */
4035  void __check_heap_object(const void *ptr, unsigned long n, struct page *page,
4036  			 bool to_user)
4037  {
4038  	struct kmem_cache *s;
4039  	unsigned int offset;
4040  	size_t object_size;
4041  
4042  	ptr = kasan_reset_tag(ptr);
4043  
4044  	/* Find object and usable object size. */
4045  	s = page->slab_cache;
4046  
4047  	/* Reject impossible pointers. */
4048  	if (ptr < page_address(page))
4049  		usercopy_abort("SLUB object not in SLUB page?!", NULL,
4050  			       to_user, 0, n);
4051  
4052  	/* Find offset within object. */
4053  	offset = (ptr - page_address(page)) % s->size;
4054  
4055  	/* Adjust for redzone and reject if within the redzone. */
4056  	if (kmem_cache_debug_flags(s, SLAB_RED_ZONE)) {
4057  		if (offset < s->red_left_pad)
4058  			usercopy_abort("SLUB object in left red zone",
4059  				       s->name, to_user, offset, n);
4060  		offset -= s->red_left_pad;
4061  	}
4062  
4063  	/* Allow address range falling entirely within usercopy region. */
4064  	if (offset >= s->useroffset &&
4065  	    offset - s->useroffset <= s->usersize &&
4066  	    n <= s->useroffset - offset + s->usersize)
4067  		return;
4068  
4069  	/*
4070  	 * If the copy is still within the allocated object, produce
4071  	 * a warning instead of rejecting the copy. This is intended
4072  	 * to be a temporary method to find any missing usercopy
4073  	 * whitelists.
4074  	 */
4075  	object_size = slab_ksize(s);
4076  	if (usercopy_fallback &&
4077  	    offset <= object_size && n <= object_size - offset) {
4078  		usercopy_warn("SLUB object", s->name, to_user, offset, n);
4079  		return;
4080  	}
4081  
4082  	usercopy_abort("SLUB object", s->name, to_user, offset, n);
4083  }
4084  #endif /* CONFIG_HARDENED_USERCOPY */
4085  
4086  size_t __ksize(const void *object)
4087  {
4088  	struct page *page;
4089  
4090  	if (unlikely(object == ZERO_SIZE_PTR))
4091  		return 0;
4092  
4093  	page = virt_to_head_page(object);
4094  
4095  	if (unlikely(!PageSlab(page))) {
4096  		WARN_ON(!PageCompound(page));
4097  		return page_size(page);
4098  	}
4099  
4100  	return slab_ksize(page->slab_cache);
4101  }
4102  EXPORT_SYMBOL(__ksize);
4103  
4104  void kfree(const void *x)
4105  {
4106  	struct page *page;
4107  	void *object = (void *)x;
4108  
4109  	trace_kfree(_RET_IP_, x);
4110  
4111  	if (unlikely(ZERO_OR_NULL_PTR(x)))
4112  		return;
4113  
4114  	page = virt_to_head_page(x);
4115  	if (unlikely(!PageSlab(page))) {
4116  		unsigned int order = compound_order(page);
4117  
4118  		BUG_ON(!PageCompound(page));
4119  		kfree_hook(object);
4120  		mod_node_page_state(page_pgdat(page), NR_SLAB_UNRECLAIMABLE_B,
4121  				    -(PAGE_SIZE << order));
4122  		__free_pages(page, order);
4123  		return;
4124  	}
4125  	slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
4126  }
4127  EXPORT_SYMBOL(kfree);
4128  
4129  #define SHRINK_PROMOTE_MAX 32
4130  
4131  /*
4132   * kmem_cache_shrink discards empty slabs and promotes the slabs filled
4133   * up most to the head of the partial lists. New allocations will then
4134   * fill those up and thus they can be removed from the partial lists.
4135   *
4136   * The slabs with the least items are placed last. This results in them
4137   * being allocated from last increasing the chance that the last objects
4138   * are freed in them.
4139   */
4140  int __kmem_cache_shrink(struct kmem_cache *s)
4141  {
4142  	int node;
4143  	int i;
4144  	struct kmem_cache_node *n;
4145  	struct page *page;
4146  	struct page *t;
4147  	struct list_head discard;
4148  	struct list_head promote[SHRINK_PROMOTE_MAX];
4149  	unsigned long flags;
4150  	int ret = 0;
4151  
4152  	flush_all(s);
4153  	for_each_kmem_cache_node(s, node, n) {
4154  		INIT_LIST_HEAD(&discard);
4155  		for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
4156  			INIT_LIST_HEAD(promote + i);
4157  
4158  		spin_lock_irqsave(&n->list_lock, flags);
4159  
4160  		/*
4161  		 * Build lists of slabs to discard or promote.
4162  		 *
4163  		 * Note that concurrent frees may occur while we hold the
4164  		 * list_lock. page->inuse here is the upper limit.
4165  		 */
4166  		list_for_each_entry_safe(page, t, &n->partial, slab_list) {
4167  			int free = page->objects - page->inuse;
4168  
4169  			/* Do not reread page->inuse */
4170  			barrier();
4171  
4172  			/* We do not keep full slabs on the list */
4173  			BUG_ON(free <= 0);
4174  
4175  			if (free == page->objects) {
4176  				list_move(&page->slab_list, &discard);
4177  				n->nr_partial--;
4178  			} else if (free <= SHRINK_PROMOTE_MAX)
4179  				list_move(&page->slab_list, promote + free - 1);
4180  		}
4181  
4182  		/*
4183  		 * Promote the slabs filled up most to the head of the
4184  		 * partial list.
4185  		 */
4186  		for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
4187  			list_splice(promote + i, &n->partial);
4188  
4189  		spin_unlock_irqrestore(&n->list_lock, flags);
4190  
4191  		/* Release empty slabs */
4192  		list_for_each_entry_safe(page, t, &discard, slab_list)
4193  			discard_slab(s, page);
4194  
4195  		if (slabs_node(s, node))
4196  			ret = 1;
4197  	}
4198  
4199  	return ret;
4200  }
4201  
4202  static int slab_mem_going_offline_callback(void *arg)
4203  {
4204  	struct kmem_cache *s;
4205  
4206  	mutex_lock(&slab_mutex);
4207  	list_for_each_entry(s, &slab_caches, list)
4208  		__kmem_cache_shrink(s);
4209  	mutex_unlock(&slab_mutex);
4210  
4211  	return 0;
4212  }
4213  
4214  static void slab_mem_offline_callback(void *arg)
4215  {
4216  	struct kmem_cache_node *n;
4217  	struct kmem_cache *s;
4218  	struct memory_notify *marg = arg;
4219  	int offline_node;
4220  
4221  	offline_node = marg->status_change_nid_normal;
4222  
4223  	/*
4224  	 * If the node still has available memory. we need kmem_cache_node
4225  	 * for it yet.
4226  	 */
4227  	if (offline_node < 0)
4228  		return;
4229  
4230  	mutex_lock(&slab_mutex);
4231  	list_for_each_entry(s, &slab_caches, list) {
4232  		n = get_node(s, offline_node);
4233  		if (n) {
4234  			/*
4235  			 * if n->nr_slabs > 0, slabs still exist on the node
4236  			 * that is going down. We were unable to free them,
4237  			 * and offline_pages() function shouldn't call this
4238  			 * callback. So, we must fail.
4239  			 */
4240  			BUG_ON(slabs_node(s, offline_node));
4241  
4242  			s->node[offline_node] = NULL;
4243  			kmem_cache_free(kmem_cache_node, n);
4244  		}
4245  	}
4246  	mutex_unlock(&slab_mutex);
4247  }
4248  
4249  static int slab_mem_going_online_callback(void *arg)
4250  {
4251  	struct kmem_cache_node *n;
4252  	struct kmem_cache *s;
4253  	struct memory_notify *marg = arg;
4254  	int nid = marg->status_change_nid_normal;
4255  	int ret = 0;
4256  
4257  	/*
4258  	 * If the node's memory is already available, then kmem_cache_node is
4259  	 * already created. Nothing to do.
4260  	 */
4261  	if (nid < 0)
4262  		return 0;
4263  
4264  	/*
4265  	 * We are bringing a node online. No memory is available yet. We must
4266  	 * allocate a kmem_cache_node structure in order to bring the node
4267  	 * online.
4268  	 */
4269  	mutex_lock(&slab_mutex);
4270  	list_for_each_entry(s, &slab_caches, list) {
4271  		/*
4272  		 * XXX: kmem_cache_alloc_node will fallback to other nodes
4273  		 *      since memory is not yet available from the node that
4274  		 *      is brought up.
4275  		 */
4276  		n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4277  		if (!n) {
4278  			ret = -ENOMEM;
4279  			goto out;
4280  		}
4281  		init_kmem_cache_node(n);
4282  		s->node[nid] = n;
4283  	}
4284  out:
4285  	mutex_unlock(&slab_mutex);
4286  	return ret;
4287  }
4288  
4289  static int slab_memory_callback(struct notifier_block *self,
4290  				unsigned long action, void *arg)
4291  {
4292  	int ret = 0;
4293  
4294  	switch (action) {
4295  	case MEM_GOING_ONLINE:
4296  		ret = slab_mem_going_online_callback(arg);
4297  		break;
4298  	case MEM_GOING_OFFLINE:
4299  		ret = slab_mem_going_offline_callback(arg);
4300  		break;
4301  	case MEM_OFFLINE:
4302  	case MEM_CANCEL_ONLINE:
4303  		slab_mem_offline_callback(arg);
4304  		break;
4305  	case MEM_ONLINE:
4306  	case MEM_CANCEL_OFFLINE:
4307  		break;
4308  	}
4309  	if (ret)
4310  		ret = notifier_from_errno(ret);
4311  	else
4312  		ret = NOTIFY_OK;
4313  	return ret;
4314  }
4315  
4316  static struct notifier_block slab_memory_callback_nb = {
4317  	.notifier_call = slab_memory_callback,
4318  	.priority = SLAB_CALLBACK_PRI,
4319  };
4320  
4321  /********************************************************************
4322   *			Basic setup of slabs
4323   *******************************************************************/
4324  
4325  /*
4326   * Used for early kmem_cache structures that were allocated using
4327   * the page allocator. Allocate them properly then fix up the pointers
4328   * that may be pointing to the wrong kmem_cache structure.
4329   */
4330  
4331  static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4332  {
4333  	int node;
4334  	struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4335  	struct kmem_cache_node *n;
4336  
4337  	memcpy(s, static_cache, kmem_cache->object_size);
4338  
4339  	/*
4340  	 * This runs very early, and only the boot processor is supposed to be
4341  	 * up.  Even if it weren't true, IRQs are not up so we couldn't fire
4342  	 * IPIs around.
4343  	 */
4344  	__flush_cpu_slab(s, smp_processor_id());
4345  	for_each_kmem_cache_node(s, node, n) {
4346  		struct page *p;
4347  
4348  		list_for_each_entry(p, &n->partial, slab_list)
4349  			p->slab_cache = s;
4350  
4351  #ifdef CONFIG_SLUB_DEBUG
4352  		list_for_each_entry(p, &n->full, slab_list)
4353  			p->slab_cache = s;
4354  #endif
4355  	}
4356  	list_add(&s->list, &slab_caches);
4357  	return s;
4358  }
4359  
4360  void __init kmem_cache_init(void)
4361  {
4362  	static __initdata struct kmem_cache boot_kmem_cache,
4363  		boot_kmem_cache_node;
4364  
4365  	if (debug_guardpage_minorder())
4366  		slub_max_order = 0;
4367  
4368  	kmem_cache_node = &boot_kmem_cache_node;
4369  	kmem_cache = &boot_kmem_cache;
4370  
4371  	create_boot_cache(kmem_cache_node, "kmem_cache_node",
4372  		sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
4373  
4374  	register_hotmemory_notifier(&slab_memory_callback_nb);
4375  
4376  	/* Able to allocate the per node structures */
4377  	slab_state = PARTIAL;
4378  
4379  	create_boot_cache(kmem_cache, "kmem_cache",
4380  			offsetof(struct kmem_cache, node) +
4381  				nr_node_ids * sizeof(struct kmem_cache_node *),
4382  		       SLAB_HWCACHE_ALIGN, 0, 0);
4383  
4384  	kmem_cache = bootstrap(&boot_kmem_cache);
4385  	kmem_cache_node = bootstrap(&boot_kmem_cache_node);
4386  
4387  	/* Now we can use the kmem_cache to allocate kmalloc slabs */
4388  	setup_kmalloc_cache_index_table();
4389  	create_kmalloc_caches(0);
4390  
4391  	/* Setup random freelists for each cache */
4392  	init_freelist_randomization();
4393  
4394  	cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
4395  				  slub_cpu_dead);
4396  
4397  	pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
4398  		cache_line_size(),
4399  		slub_min_order, slub_max_order, slub_min_objects,
4400  		nr_cpu_ids, nr_node_ids);
4401  }
4402  
4403  void __init kmem_cache_init_late(void)
4404  {
4405  }
4406  
4407  struct kmem_cache *
4408  __kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
4409  		   slab_flags_t flags, void (*ctor)(void *))
4410  {
4411  	struct kmem_cache *s;
4412  
4413  	s = find_mergeable(size, align, flags, name, ctor);
4414  	if (s) {
4415  		s->refcount++;
4416  
4417  		/*
4418  		 * Adjust the object sizes so that we clear
4419  		 * the complete object on kzalloc.
4420  		 */
4421  		s->object_size = max(s->object_size, size);
4422  		s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
4423  
4424  		if (sysfs_slab_alias(s, name)) {
4425  			s->refcount--;
4426  			s = NULL;
4427  		}
4428  	}
4429  
4430  	return s;
4431  }
4432  
4433  int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
4434  {
4435  	int err;
4436  
4437  	err = kmem_cache_open(s, flags);
4438  	if (err)
4439  		return err;
4440  
4441  	/* Mutex is not taken during early boot */
4442  	if (slab_state <= UP)
4443  		return 0;
4444  
4445  	err = sysfs_slab_add(s);
4446  	if (err)
4447  		__kmem_cache_release(s);
4448  
4449  	return err;
4450  }
4451  
4452  void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4453  {
4454  	struct kmem_cache *s;
4455  	void *ret;
4456  
4457  	if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4458  		return kmalloc_large(size, gfpflags);
4459  
4460  	s = kmalloc_slab(size, gfpflags);
4461  
4462  	if (unlikely(ZERO_OR_NULL_PTR(s)))
4463  		return s;
4464  
4465  	ret = slab_alloc(s, gfpflags, caller);
4466  
4467  	/* Honor the call site pointer we received. */
4468  	trace_kmalloc(caller, ret, size, s->size, gfpflags);
4469  
4470  	return ret;
4471  }
4472  EXPORT_SYMBOL(__kmalloc_track_caller);
4473  
4474  #ifdef CONFIG_NUMA
4475  void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4476  					int node, unsigned long caller)
4477  {
4478  	struct kmem_cache *s;
4479  	void *ret;
4480  
4481  	if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4482  		ret = kmalloc_large_node(size, gfpflags, node);
4483  
4484  		trace_kmalloc_node(caller, ret,
4485  				   size, PAGE_SIZE << get_order(size),
4486  				   gfpflags, node);
4487  
4488  		return ret;
4489  	}
4490  
4491  	s = kmalloc_slab(size, gfpflags);
4492  
4493  	if (unlikely(ZERO_OR_NULL_PTR(s)))
4494  		return s;
4495  
4496  	ret = slab_alloc_node(s, gfpflags, node, caller);
4497  
4498  	/* Honor the call site pointer we received. */
4499  	trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4500  
4501  	return ret;
4502  }
4503  EXPORT_SYMBOL(__kmalloc_node_track_caller);
4504  #endif
4505  
4506  #ifdef CONFIG_SYSFS
4507  static int count_inuse(struct page *page)
4508  {
4509  	return page->inuse;
4510  }
4511  
4512  static int count_total(struct page *page)
4513  {
4514  	return page->objects;
4515  }
4516  #endif
4517  
4518  #ifdef CONFIG_SLUB_DEBUG
4519  static void validate_slab(struct kmem_cache *s, struct page *page)
4520  {
4521  	void *p;
4522  	void *addr = page_address(page);
4523  	unsigned long *map;
4524  
4525  	slab_lock(page);
4526  
4527  	if (!check_slab(s, page) || !on_freelist(s, page, NULL))
4528  		goto unlock;
4529  
4530  	/* Now we know that a valid freelist exists */
4531  	map = get_map(s, page);
4532  	for_each_object(p, s, addr, page->objects) {
4533  		u8 val = test_bit(__obj_to_index(s, addr, p), map) ?
4534  			 SLUB_RED_INACTIVE : SLUB_RED_ACTIVE;
4535  
4536  		if (!check_object(s, page, p, val))
4537  			break;
4538  	}
4539  	put_map(map);
4540  unlock:
4541  	slab_unlock(page);
4542  }
4543  
4544  static int validate_slab_node(struct kmem_cache *s,
4545  		struct kmem_cache_node *n)
4546  {
4547  	unsigned long count = 0;
4548  	struct page *page;
4549  	unsigned long flags;
4550  
4551  	spin_lock_irqsave(&n->list_lock, flags);
4552  
4553  	list_for_each_entry(page, &n->partial, slab_list) {
4554  		validate_slab(s, page);
4555  		count++;
4556  	}
4557  	if (count != n->nr_partial)
4558  		pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4559  		       s->name, count, n->nr_partial);
4560  
4561  	if (!(s->flags & SLAB_STORE_USER))
4562  		goto out;
4563  
4564  	list_for_each_entry(page, &n->full, slab_list) {
4565  		validate_slab(s, page);
4566  		count++;
4567  	}
4568  	if (count != atomic_long_read(&n->nr_slabs))
4569  		pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4570  		       s->name, count, atomic_long_read(&n->nr_slabs));
4571  
4572  out:
4573  	spin_unlock_irqrestore(&n->list_lock, flags);
4574  	return count;
4575  }
4576  
4577  static long validate_slab_cache(struct kmem_cache *s)
4578  {
4579  	int node;
4580  	unsigned long count = 0;
4581  	struct kmem_cache_node *n;
4582  
4583  	flush_all(s);
4584  	for_each_kmem_cache_node(s, node, n)
4585  		count += validate_slab_node(s, n);
4586  
4587  	return count;
4588  }
4589  /*
4590   * Generate lists of code addresses where slabcache objects are allocated
4591   * and freed.
4592   */
4593  
4594  struct location {
4595  	unsigned long count;
4596  	unsigned long addr;
4597  	long long sum_time;
4598  	long min_time;
4599  	long max_time;
4600  	long min_pid;
4601  	long max_pid;
4602  	DECLARE_BITMAP(cpus, NR_CPUS);
4603  	nodemask_t nodes;
4604  };
4605  
4606  struct loc_track {
4607  	unsigned long max;
4608  	unsigned long count;
4609  	struct location *loc;
4610  };
4611  
4612  static void free_loc_track(struct loc_track *t)
4613  {
4614  	if (t->max)
4615  		free_pages((unsigned long)t->loc,
4616  			get_order(sizeof(struct location) * t->max));
4617  }
4618  
4619  static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4620  {
4621  	struct location *l;
4622  	int order;
4623  
4624  	order = get_order(sizeof(struct location) * max);
4625  
4626  	l = (void *)__get_free_pages(flags, order);
4627  	if (!l)
4628  		return 0;
4629  
4630  	if (t->count) {
4631  		memcpy(l, t->loc, sizeof(struct location) * t->count);
4632  		free_loc_track(t);
4633  	}
4634  	t->max = max;
4635  	t->loc = l;
4636  	return 1;
4637  }
4638  
4639  static int add_location(struct loc_track *t, struct kmem_cache *s,
4640  				const struct track *track)
4641  {
4642  	long start, end, pos;
4643  	struct location *l;
4644  	unsigned long caddr;
4645  	unsigned long age = jiffies - track->when;
4646  
4647  	start = -1;
4648  	end = t->count;
4649  
4650  	for ( ; ; ) {
4651  		pos = start + (end - start + 1) / 2;
4652  
4653  		/*
4654  		 * There is nothing at "end". If we end up there
4655  		 * we need to add something to before end.
4656  		 */
4657  		if (pos == end)
4658  			break;
4659  
4660  		caddr = t->loc[pos].addr;
4661  		if (track->addr == caddr) {
4662  
4663  			l = &t->loc[pos];
4664  			l->count++;
4665  			if (track->when) {
4666  				l->sum_time += age;
4667  				if (age < l->min_time)
4668  					l->min_time = age;
4669  				if (age > l->max_time)
4670  					l->max_time = age;
4671  
4672  				if (track->pid < l->min_pid)
4673  					l->min_pid = track->pid;
4674  				if (track->pid > l->max_pid)
4675  					l->max_pid = track->pid;
4676  
4677  				cpumask_set_cpu(track->cpu,
4678  						to_cpumask(l->cpus));
4679  			}
4680  			node_set(page_to_nid(virt_to_page(track)), l->nodes);
4681  			return 1;
4682  		}
4683  
4684  		if (track->addr < caddr)
4685  			end = pos;
4686  		else
4687  			start = pos;
4688  	}
4689  
4690  	/*
4691  	 * Not found. Insert new tracking element.
4692  	 */
4693  	if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4694  		return 0;
4695  
4696  	l = t->loc + pos;
4697  	if (pos < t->count)
4698  		memmove(l + 1, l,
4699  			(t->count - pos) * sizeof(struct location));
4700  	t->count++;
4701  	l->count = 1;
4702  	l->addr = track->addr;
4703  	l->sum_time = age;
4704  	l->min_time = age;
4705  	l->max_time = age;
4706  	l->min_pid = track->pid;
4707  	l->max_pid = track->pid;
4708  	cpumask_clear(to_cpumask(l->cpus));
4709  	cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4710  	nodes_clear(l->nodes);
4711  	node_set(page_to_nid(virt_to_page(track)), l->nodes);
4712  	return 1;
4713  }
4714  
4715  static void process_slab(struct loc_track *t, struct kmem_cache *s,
4716  		struct page *page, enum track_item alloc)
4717  {
4718  	void *addr = page_address(page);
4719  	void *p;
4720  	unsigned long *map;
4721  
4722  	map = get_map(s, page);
4723  	for_each_object(p, s, addr, page->objects)
4724  		if (!test_bit(__obj_to_index(s, addr, p), map))
4725  			add_location(t, s, get_track(s, p, alloc));
4726  	put_map(map);
4727  }
4728  
4729  static int list_locations(struct kmem_cache *s, char *buf,
4730  			  enum track_item alloc)
4731  {
4732  	int len = 0;
4733  	unsigned long i;
4734  	struct loc_track t = { 0, 0, NULL };
4735  	int node;
4736  	struct kmem_cache_node *n;
4737  
4738  	if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4739  			     GFP_KERNEL)) {
4740  		return sysfs_emit(buf, "Out of memory\n");
4741  	}
4742  	/* Push back cpu slabs */
4743  	flush_all(s);
4744  
4745  	for_each_kmem_cache_node(s, node, n) {
4746  		unsigned long flags;
4747  		struct page *page;
4748  
4749  		if (!atomic_long_read(&n->nr_slabs))
4750  			continue;
4751  
4752  		spin_lock_irqsave(&n->list_lock, flags);
4753  		list_for_each_entry(page, &n->partial, slab_list)
4754  			process_slab(&t, s, page, alloc);
4755  		list_for_each_entry(page, &n->full, slab_list)
4756  			process_slab(&t, s, page, alloc);
4757  		spin_unlock_irqrestore(&n->list_lock, flags);
4758  	}
4759  
4760  	for (i = 0; i < t.count; i++) {
4761  		struct location *l = &t.loc[i];
4762  
4763  		len += sysfs_emit_at(buf, len, "%7ld ", l->count);
4764  
4765  		if (l->addr)
4766  			len += sysfs_emit_at(buf, len, "%pS", (void *)l->addr);
4767  		else
4768  			len += sysfs_emit_at(buf, len, "<not-available>");
4769  
4770  		if (l->sum_time != l->min_time)
4771  			len += sysfs_emit_at(buf, len, " age=%ld/%ld/%ld",
4772  					     l->min_time,
4773  					     (long)div_u64(l->sum_time,
4774  							   l->count),
4775  					     l->max_time);
4776  		else
4777  			len += sysfs_emit_at(buf, len, " age=%ld", l->min_time);
4778  
4779  		if (l->min_pid != l->max_pid)
4780  			len += sysfs_emit_at(buf, len, " pid=%ld-%ld",
4781  					     l->min_pid, l->max_pid);
4782  		else
4783  			len += sysfs_emit_at(buf, len, " pid=%ld",
4784  					     l->min_pid);
4785  
4786  		if (num_online_cpus() > 1 &&
4787  		    !cpumask_empty(to_cpumask(l->cpus)))
4788  			len += sysfs_emit_at(buf, len, " cpus=%*pbl",
4789  					     cpumask_pr_args(to_cpumask(l->cpus)));
4790  
4791  		if (nr_online_nodes > 1 && !nodes_empty(l->nodes))
4792  			len += sysfs_emit_at(buf, len, " nodes=%*pbl",
4793  					     nodemask_pr_args(&l->nodes));
4794  
4795  		len += sysfs_emit_at(buf, len, "\n");
4796  	}
4797  
4798  	free_loc_track(&t);
4799  	if (!t.count)
4800  		len += sysfs_emit_at(buf, len, "No data\n");
4801  
4802  	return len;
4803  }
4804  #endif	/* CONFIG_SLUB_DEBUG */
4805  
4806  #ifdef SLUB_RESILIENCY_TEST
4807  static void __init resiliency_test(void)
4808  {
4809  	u8 *p;
4810  	int type = KMALLOC_NORMAL;
4811  
4812  	BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4813  
4814  	pr_err("SLUB resiliency testing\n");
4815  	pr_err("-----------------------\n");
4816  	pr_err("A. Corruption after allocation\n");
4817  
4818  	p = kzalloc(16, GFP_KERNEL);
4819  	p[16] = 0x12;
4820  	pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4821  	       p + 16);
4822  
4823  	validate_slab_cache(kmalloc_caches[type][4]);
4824  
4825  	/* Hmmm... The next two are dangerous */
4826  	p = kzalloc(32, GFP_KERNEL);
4827  	p[32 + sizeof(void *)] = 0x34;
4828  	pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4829  	       p);
4830  	pr_err("If allocated object is overwritten then not detectable\n\n");
4831  
4832  	validate_slab_cache(kmalloc_caches[type][5]);
4833  	p = kzalloc(64, GFP_KERNEL);
4834  	p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4835  	*p = 0x56;
4836  	pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4837  	       p);
4838  	pr_err("If allocated object is overwritten then not detectable\n\n");
4839  	validate_slab_cache(kmalloc_caches[type][6]);
4840  
4841  	pr_err("\nB. Corruption after free\n");
4842  	p = kzalloc(128, GFP_KERNEL);
4843  	kfree(p);
4844  	*p = 0x78;
4845  	pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4846  	validate_slab_cache(kmalloc_caches[type][7]);
4847  
4848  	p = kzalloc(256, GFP_KERNEL);
4849  	kfree(p);
4850  	p[50] = 0x9a;
4851  	pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4852  	validate_slab_cache(kmalloc_caches[type][8]);
4853  
4854  	p = kzalloc(512, GFP_KERNEL);
4855  	kfree(p);
4856  	p[512] = 0xab;
4857  	pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4858  	validate_slab_cache(kmalloc_caches[type][9]);
4859  }
4860  #else
4861  #ifdef CONFIG_SYSFS
4862  static void resiliency_test(void) {};
4863  #endif
4864  #endif	/* SLUB_RESILIENCY_TEST */
4865  
4866  #ifdef CONFIG_SYSFS
4867  enum slab_stat_type {
4868  	SL_ALL,			/* All slabs */
4869  	SL_PARTIAL,		/* Only partially allocated slabs */
4870  	SL_CPU,			/* Only slabs used for cpu caches */
4871  	SL_OBJECTS,		/* Determine allocated objects not slabs */
4872  	SL_TOTAL		/* Determine object capacity not slabs */
4873  };
4874  
4875  #define SO_ALL		(1 << SL_ALL)
4876  #define SO_PARTIAL	(1 << SL_PARTIAL)
4877  #define SO_CPU		(1 << SL_CPU)
4878  #define SO_OBJECTS	(1 << SL_OBJECTS)
4879  #define SO_TOTAL	(1 << SL_TOTAL)
4880  
4881  #ifdef CONFIG_MEMCG
4882  static bool memcg_sysfs_enabled = IS_ENABLED(CONFIG_SLUB_MEMCG_SYSFS_ON);
4883  
4884  static int __init setup_slub_memcg_sysfs(char *str)
4885  {
4886  	int v;
4887  
4888  	if (get_option(&str, &v) > 0)
4889  		memcg_sysfs_enabled = v;
4890  
4891  	return 1;
4892  }
4893  
4894  __setup("slub_memcg_sysfs=", setup_slub_memcg_sysfs);
4895  #endif
4896  
4897  static ssize_t show_slab_objects(struct kmem_cache *s,
4898  				 char *buf, unsigned long flags)
4899  {
4900  	unsigned long total = 0;
4901  	int node;
4902  	int x;
4903  	unsigned long *nodes;
4904  	int len = 0;
4905  
4906  	nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
4907  	if (!nodes)
4908  		return -ENOMEM;
4909  
4910  	if (flags & SO_CPU) {
4911  		int cpu;
4912  
4913  		for_each_possible_cpu(cpu) {
4914  			struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4915  							       cpu);
4916  			int node;
4917  			struct page *page;
4918  
4919  			page = READ_ONCE(c->page);
4920  			if (!page)
4921  				continue;
4922  
4923  			node = page_to_nid(page);
4924  			if (flags & SO_TOTAL)
4925  				x = page->objects;
4926  			else if (flags & SO_OBJECTS)
4927  				x = page->inuse;
4928  			else
4929  				x = 1;
4930  
4931  			total += x;
4932  			nodes[node] += x;
4933  
4934  			page = slub_percpu_partial_read_once(c);
4935  			if (page) {
4936  				node = page_to_nid(page);
4937  				if (flags & SO_TOTAL)
4938  					WARN_ON_ONCE(1);
4939  				else if (flags & SO_OBJECTS)
4940  					WARN_ON_ONCE(1);
4941  				else
4942  					x = page->pages;
4943  				total += x;
4944  				nodes[node] += x;
4945  			}
4946  		}
4947  	}
4948  
4949  	/*
4950  	 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
4951  	 * already held which will conflict with an existing lock order:
4952  	 *
4953  	 * mem_hotplug_lock->slab_mutex->kernfs_mutex
4954  	 *
4955  	 * We don't really need mem_hotplug_lock (to hold off
4956  	 * slab_mem_going_offline_callback) here because slab's memory hot
4957  	 * unplug code doesn't destroy the kmem_cache->node[] data.
4958  	 */
4959  
4960  #ifdef CONFIG_SLUB_DEBUG
4961  	if (flags & SO_ALL) {
4962  		struct kmem_cache_node *n;
4963  
4964  		for_each_kmem_cache_node(s, node, n) {
4965  
4966  			if (flags & SO_TOTAL)
4967  				x = atomic_long_read(&n->total_objects);
4968  			else if (flags & SO_OBJECTS)
4969  				x = atomic_long_read(&n->total_objects) -
4970  					count_partial(n, count_free);
4971  			else
4972  				x = atomic_long_read(&n->nr_slabs);
4973  			total += x;
4974  			nodes[node] += x;
4975  		}
4976  
4977  	} else
4978  #endif
4979  	if (flags & SO_PARTIAL) {
4980  		struct kmem_cache_node *n;
4981  
4982  		for_each_kmem_cache_node(s, node, n) {
4983  			if (flags & SO_TOTAL)
4984  				x = count_partial(n, count_total);
4985  			else if (flags & SO_OBJECTS)
4986  				x = count_partial(n, count_inuse);
4987  			else
4988  				x = n->nr_partial;
4989  			total += x;
4990  			nodes[node] += x;
4991  		}
4992  	}
4993  
4994  	len += sysfs_emit_at(buf, len, "%lu", total);
4995  #ifdef CONFIG_NUMA
4996  	for (node = 0; node < nr_node_ids; node++) {
4997  		if (nodes[node])
4998  			len += sysfs_emit_at(buf, len, " N%d=%lu",
4999  					     node, nodes[node]);
5000  	}
5001  #endif
5002  	len += sysfs_emit_at(buf, len, "\n");
5003  	kfree(nodes);
5004  
5005  	return len;
5006  }
5007  
5008  #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
5009  #define to_slab(n) container_of(n, struct kmem_cache, kobj)
5010  
5011  struct slab_attribute {
5012  	struct attribute attr;
5013  	ssize_t (*show)(struct kmem_cache *s, char *buf);
5014  	ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
5015  };
5016  
5017  #define SLAB_ATTR_RO(_name) \
5018  	static struct slab_attribute _name##_attr = \
5019  	__ATTR(_name, 0400, _name##_show, NULL)
5020  
5021  #define SLAB_ATTR(_name) \
5022  	static struct slab_attribute _name##_attr =  \
5023  	__ATTR(_name, 0600, _name##_show, _name##_store)
5024  
5025  static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
5026  {
5027  	return sysfs_emit(buf, "%u\n", s->size);
5028  }
5029  SLAB_ATTR_RO(slab_size);
5030  
5031  static ssize_t align_show(struct kmem_cache *s, char *buf)
5032  {
5033  	return sysfs_emit(buf, "%u\n", s->align);
5034  }
5035  SLAB_ATTR_RO(align);
5036  
5037  static ssize_t object_size_show(struct kmem_cache *s, char *buf)
5038  {
5039  	return sysfs_emit(buf, "%u\n", s->object_size);
5040  }
5041  SLAB_ATTR_RO(object_size);
5042  
5043  static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
5044  {
5045  	return sysfs_emit(buf, "%u\n", oo_objects(s->oo));
5046  }
5047  SLAB_ATTR_RO(objs_per_slab);
5048  
5049  static ssize_t order_show(struct kmem_cache *s, char *buf)
5050  {
5051  	return sysfs_emit(buf, "%u\n", oo_order(s->oo));
5052  }
5053  SLAB_ATTR_RO(order);
5054  
5055  static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
5056  {
5057  	return sysfs_emit(buf, "%lu\n", s->min_partial);
5058  }
5059  
5060  static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
5061  				 size_t length)
5062  {
5063  	unsigned long min;
5064  	int err;
5065  
5066  	err = kstrtoul(buf, 10, &min);
5067  	if (err)
5068  		return err;
5069  
5070  	set_min_partial(s, min);
5071  	return length;
5072  }
5073  SLAB_ATTR(min_partial);
5074  
5075  static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
5076  {
5077  	return sysfs_emit(buf, "%u\n", slub_cpu_partial(s));
5078  }
5079  
5080  static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
5081  				 size_t length)
5082  {
5083  	unsigned int objects;
5084  	int err;
5085  
5086  	err = kstrtouint(buf, 10, &objects);
5087  	if (err)
5088  		return err;
5089  	if (objects && !kmem_cache_has_cpu_partial(s))
5090  		return -EINVAL;
5091  
5092  	slub_set_cpu_partial(s, objects);
5093  	flush_all(s);
5094  	return length;
5095  }
5096  SLAB_ATTR(cpu_partial);
5097  
5098  static ssize_t ctor_show(struct kmem_cache *s, char *buf)
5099  {
5100  	if (!s->ctor)
5101  		return 0;
5102  	return sysfs_emit(buf, "%pS\n", s->ctor);
5103  }
5104  SLAB_ATTR_RO(ctor);
5105  
5106  static ssize_t aliases_show(struct kmem_cache *s, char *buf)
5107  {
5108  	return sysfs_emit(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
5109  }
5110  SLAB_ATTR_RO(aliases);
5111  
5112  static ssize_t partial_show(struct kmem_cache *s, char *buf)
5113  {
5114  	return show_slab_objects(s, buf, SO_PARTIAL);
5115  }
5116  SLAB_ATTR_RO(partial);
5117  
5118  static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
5119  {
5120  	return show_slab_objects(s, buf, SO_CPU);
5121  }
5122  SLAB_ATTR_RO(cpu_slabs);
5123  
5124  static ssize_t objects_show(struct kmem_cache *s, char *buf)
5125  {
5126  	return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5127  }
5128  SLAB_ATTR_RO(objects);
5129  
5130  static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5131  {
5132  	return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5133  }
5134  SLAB_ATTR_RO(objects_partial);
5135  
5136  static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5137  {
5138  	int objects = 0;
5139  	int pages = 0;
5140  	int cpu;
5141  	int len = 0;
5142  
5143  	for_each_online_cpu(cpu) {
5144  		struct page *page;
5145  
5146  		page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5147  
5148  		if (page) {
5149  			pages += page->pages;
5150  			objects += page->pobjects;
5151  		}
5152  	}
5153  
5154  	len += sysfs_emit_at(buf, len, "%d(%d)", objects, pages);
5155  
5156  #ifdef CONFIG_SMP
5157  	for_each_online_cpu(cpu) {
5158  		struct page *page;
5159  
5160  		page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5161  		if (page)
5162  			len += sysfs_emit_at(buf, len, " C%d=%d(%d)",
5163  					     cpu, page->pobjects, page->pages);
5164  	}
5165  #endif
5166  	len += sysfs_emit_at(buf, len, "\n");
5167  
5168  	return len;
5169  }
5170  SLAB_ATTR_RO(slabs_cpu_partial);
5171  
5172  static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5173  {
5174  	return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5175  }
5176  SLAB_ATTR_RO(reclaim_account);
5177  
5178  static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5179  {
5180  	return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5181  }
5182  SLAB_ATTR_RO(hwcache_align);
5183  
5184  #ifdef CONFIG_ZONE_DMA
5185  static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5186  {
5187  	return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5188  }
5189  SLAB_ATTR_RO(cache_dma);
5190  #endif
5191  
5192  static ssize_t usersize_show(struct kmem_cache *s, char *buf)
5193  {
5194  	return sysfs_emit(buf, "%u\n", s->usersize);
5195  }
5196  SLAB_ATTR_RO(usersize);
5197  
5198  static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5199  {
5200  	return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
5201  }
5202  SLAB_ATTR_RO(destroy_by_rcu);
5203  
5204  #ifdef CONFIG_SLUB_DEBUG
5205  static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5206  {
5207  	return show_slab_objects(s, buf, SO_ALL);
5208  }
5209  SLAB_ATTR_RO(slabs);
5210  
5211  static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5212  {
5213  	return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5214  }
5215  SLAB_ATTR_RO(total_objects);
5216  
5217  static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5218  {
5219  	return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5220  }
5221  SLAB_ATTR_RO(sanity_checks);
5222  
5223  static ssize_t trace_show(struct kmem_cache *s, char *buf)
5224  {
5225  	return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5226  }
5227  SLAB_ATTR_RO(trace);
5228  
5229  static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5230  {
5231  	return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5232  }
5233  
5234  SLAB_ATTR_RO(red_zone);
5235  
5236  static ssize_t poison_show(struct kmem_cache *s, char *buf)
5237  {
5238  	return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_POISON));
5239  }
5240  
5241  SLAB_ATTR_RO(poison);
5242  
5243  static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5244  {
5245  	return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5246  }
5247  
5248  SLAB_ATTR_RO(store_user);
5249  
5250  static ssize_t validate_show(struct kmem_cache *s, char *buf)
5251  {
5252  	return 0;
5253  }
5254  
5255  static ssize_t validate_store(struct kmem_cache *s,
5256  			const char *buf, size_t length)
5257  {
5258  	int ret = -EINVAL;
5259  
5260  	if (buf[0] == '1') {
5261  		ret = validate_slab_cache(s);
5262  		if (ret >= 0)
5263  			ret = length;
5264  	}
5265  	return ret;
5266  }
5267  SLAB_ATTR(validate);
5268  
5269  static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
5270  {
5271  	if (!(s->flags & SLAB_STORE_USER))
5272  		return -ENOSYS;
5273  	return list_locations(s, buf, TRACK_ALLOC);
5274  }
5275  SLAB_ATTR_RO(alloc_calls);
5276  
5277  static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
5278  {
5279  	if (!(s->flags & SLAB_STORE_USER))
5280  		return -ENOSYS;
5281  	return list_locations(s, buf, TRACK_FREE);
5282  }
5283  SLAB_ATTR_RO(free_calls);
5284  #endif /* CONFIG_SLUB_DEBUG */
5285  
5286  #ifdef CONFIG_FAILSLAB
5287  static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5288  {
5289  	return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5290  }
5291  SLAB_ATTR_RO(failslab);
5292  #endif
5293  
5294  static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5295  {
5296  	return 0;
5297  }
5298  
5299  static ssize_t shrink_store(struct kmem_cache *s,
5300  			const char *buf, size_t length)
5301  {
5302  	if (buf[0] == '1')
5303  		kmem_cache_shrink(s);
5304  	else
5305  		return -EINVAL;
5306  	return length;
5307  }
5308  SLAB_ATTR(shrink);
5309  
5310  #ifdef CONFIG_NUMA
5311  static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5312  {
5313  	return sysfs_emit(buf, "%u\n", s->remote_node_defrag_ratio / 10);
5314  }
5315  
5316  static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5317  				const char *buf, size_t length)
5318  {
5319  	unsigned int ratio;
5320  	int err;
5321  
5322  	err = kstrtouint(buf, 10, &ratio);
5323  	if (err)
5324  		return err;
5325  	if (ratio > 100)
5326  		return -ERANGE;
5327  
5328  	s->remote_node_defrag_ratio = ratio * 10;
5329  
5330  	return length;
5331  }
5332  SLAB_ATTR(remote_node_defrag_ratio);
5333  #endif
5334  
5335  #ifdef CONFIG_SLUB_STATS
5336  static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5337  {
5338  	unsigned long sum  = 0;
5339  	int cpu;
5340  	int len = 0;
5341  	int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
5342  
5343  	if (!data)
5344  		return -ENOMEM;
5345  
5346  	for_each_online_cpu(cpu) {
5347  		unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5348  
5349  		data[cpu] = x;
5350  		sum += x;
5351  	}
5352  
5353  	len += sysfs_emit_at(buf, len, "%lu", sum);
5354  
5355  #ifdef CONFIG_SMP
5356  	for_each_online_cpu(cpu) {
5357  		if (data[cpu])
5358  			len += sysfs_emit_at(buf, len, " C%d=%u",
5359  					     cpu, data[cpu]);
5360  	}
5361  #endif
5362  	kfree(data);
5363  	len += sysfs_emit_at(buf, len, "\n");
5364  
5365  	return len;
5366  }
5367  
5368  static void clear_stat(struct kmem_cache *s, enum stat_item si)
5369  {
5370  	int cpu;
5371  
5372  	for_each_online_cpu(cpu)
5373  		per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5374  }
5375  
5376  #define STAT_ATTR(si, text) 					\
5377  static ssize_t text##_show(struct kmem_cache *s, char *buf)	\
5378  {								\
5379  	return show_stat(s, buf, si);				\
5380  }								\
5381  static ssize_t text##_store(struct kmem_cache *s,		\
5382  				const char *buf, size_t length)	\
5383  {								\
5384  	if (buf[0] != '0')					\
5385  		return -EINVAL;					\
5386  	clear_stat(s, si);					\
5387  	return length;						\
5388  }								\
5389  SLAB_ATTR(text);						\
5390  
5391  STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5392  STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5393  STAT_ATTR(FREE_FASTPATH, free_fastpath);
5394  STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5395  STAT_ATTR(FREE_FROZEN, free_frozen);
5396  STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5397  STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5398  STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5399  STAT_ATTR(ALLOC_SLAB, alloc_slab);
5400  STAT_ATTR(ALLOC_REFILL, alloc_refill);
5401  STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5402  STAT_ATTR(FREE_SLAB, free_slab);
5403  STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5404  STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5405  STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5406  STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5407  STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5408  STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5409  STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5410  STAT_ATTR(ORDER_FALLBACK, order_fallback);
5411  STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5412  STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5413  STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5414  STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5415  STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5416  STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5417  #endif	/* CONFIG_SLUB_STATS */
5418  
5419  static struct attribute *slab_attrs[] = {
5420  	&slab_size_attr.attr,
5421  	&object_size_attr.attr,
5422  	&objs_per_slab_attr.attr,
5423  	&order_attr.attr,
5424  	&min_partial_attr.attr,
5425  	&cpu_partial_attr.attr,
5426  	&objects_attr.attr,
5427  	&objects_partial_attr.attr,
5428  	&partial_attr.attr,
5429  	&cpu_slabs_attr.attr,
5430  	&ctor_attr.attr,
5431  	&aliases_attr.attr,
5432  	&align_attr.attr,
5433  	&hwcache_align_attr.attr,
5434  	&reclaim_account_attr.attr,
5435  	&destroy_by_rcu_attr.attr,
5436  	&shrink_attr.attr,
5437  	&slabs_cpu_partial_attr.attr,
5438  #ifdef CONFIG_SLUB_DEBUG
5439  	&total_objects_attr.attr,
5440  	&slabs_attr.attr,
5441  	&sanity_checks_attr.attr,
5442  	&trace_attr.attr,
5443  	&red_zone_attr.attr,
5444  	&poison_attr.attr,
5445  	&store_user_attr.attr,
5446  	&validate_attr.attr,
5447  	&alloc_calls_attr.attr,
5448  	&free_calls_attr.attr,
5449  #endif
5450  #ifdef CONFIG_ZONE_DMA
5451  	&cache_dma_attr.attr,
5452  #endif
5453  #ifdef CONFIG_NUMA
5454  	&remote_node_defrag_ratio_attr.attr,
5455  #endif
5456  #ifdef CONFIG_SLUB_STATS
5457  	&alloc_fastpath_attr.attr,
5458  	&alloc_slowpath_attr.attr,
5459  	&free_fastpath_attr.attr,
5460  	&free_slowpath_attr.attr,
5461  	&free_frozen_attr.attr,
5462  	&free_add_partial_attr.attr,
5463  	&free_remove_partial_attr.attr,
5464  	&alloc_from_partial_attr.attr,
5465  	&alloc_slab_attr.attr,
5466  	&alloc_refill_attr.attr,
5467  	&alloc_node_mismatch_attr.attr,
5468  	&free_slab_attr.attr,
5469  	&cpuslab_flush_attr.attr,
5470  	&deactivate_full_attr.attr,
5471  	&deactivate_empty_attr.attr,
5472  	&deactivate_to_head_attr.attr,
5473  	&deactivate_to_tail_attr.attr,
5474  	&deactivate_remote_frees_attr.attr,
5475  	&deactivate_bypass_attr.attr,
5476  	&order_fallback_attr.attr,
5477  	&cmpxchg_double_fail_attr.attr,
5478  	&cmpxchg_double_cpu_fail_attr.attr,
5479  	&cpu_partial_alloc_attr.attr,
5480  	&cpu_partial_free_attr.attr,
5481  	&cpu_partial_node_attr.attr,
5482  	&cpu_partial_drain_attr.attr,
5483  #endif
5484  #ifdef CONFIG_FAILSLAB
5485  	&failslab_attr.attr,
5486  #endif
5487  	&usersize_attr.attr,
5488  
5489  	NULL
5490  };
5491  
5492  static const struct attribute_group slab_attr_group = {
5493  	.attrs = slab_attrs,
5494  };
5495  
5496  static ssize_t slab_attr_show(struct kobject *kobj,
5497  				struct attribute *attr,
5498  				char *buf)
5499  {
5500  	struct slab_attribute *attribute;
5501  	struct kmem_cache *s;
5502  	int err;
5503  
5504  	attribute = to_slab_attr(attr);
5505  	s = to_slab(kobj);
5506  
5507  	if (!attribute->show)
5508  		return -EIO;
5509  
5510  	err = attribute->show(s, buf);
5511  
5512  	return err;
5513  }
5514  
5515  static ssize_t slab_attr_store(struct kobject *kobj,
5516  				struct attribute *attr,
5517  				const char *buf, size_t len)
5518  {
5519  	struct slab_attribute *attribute;
5520  	struct kmem_cache *s;
5521  	int err;
5522  
5523  	attribute = to_slab_attr(attr);
5524  	s = to_slab(kobj);
5525  
5526  	if (!attribute->store)
5527  		return -EIO;
5528  
5529  	err = attribute->store(s, buf, len);
5530  	return err;
5531  }
5532  
5533  static void kmem_cache_release(struct kobject *k)
5534  {
5535  	slab_kmem_cache_release(to_slab(k));
5536  }
5537  
5538  static const struct sysfs_ops slab_sysfs_ops = {
5539  	.show = slab_attr_show,
5540  	.store = slab_attr_store,
5541  };
5542  
5543  static struct kobj_type slab_ktype = {
5544  	.sysfs_ops = &slab_sysfs_ops,
5545  	.release = kmem_cache_release,
5546  };
5547  
5548  static struct kset *slab_kset;
5549  
5550  static inline struct kset *cache_kset(struct kmem_cache *s)
5551  {
5552  	return slab_kset;
5553  }
5554  
5555  #define ID_STR_LENGTH 64
5556  
5557  /* Create a unique string id for a slab cache:
5558   *
5559   * Format	:[flags-]size
5560   */
5561  static char *create_unique_id(struct kmem_cache *s)
5562  {
5563  	char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5564  	char *p = name;
5565  
5566  	BUG_ON(!name);
5567  
5568  	*p++ = ':';
5569  	/*
5570  	 * First flags affecting slabcache operations. We will only
5571  	 * get here for aliasable slabs so we do not need to support
5572  	 * too many flags. The flags here must cover all flags that
5573  	 * are matched during merging to guarantee that the id is
5574  	 * unique.
5575  	 */
5576  	if (s->flags & SLAB_CACHE_DMA)
5577  		*p++ = 'd';
5578  	if (s->flags & SLAB_CACHE_DMA32)
5579  		*p++ = 'D';
5580  	if (s->flags & SLAB_RECLAIM_ACCOUNT)
5581  		*p++ = 'a';
5582  	if (s->flags & SLAB_CONSISTENCY_CHECKS)
5583  		*p++ = 'F';
5584  	if (s->flags & SLAB_ACCOUNT)
5585  		*p++ = 'A';
5586  	if (p != name + 1)
5587  		*p++ = '-';
5588  	p += sprintf(p, "%07u", s->size);
5589  
5590  	BUG_ON(p > name + ID_STR_LENGTH - 1);
5591  	return name;
5592  }
5593  
5594  static int sysfs_slab_add(struct kmem_cache *s)
5595  {
5596  	int err;
5597  	const char *name;
5598  	struct kset *kset = cache_kset(s);
5599  	int unmergeable = slab_unmergeable(s);
5600  
5601  	if (!kset) {
5602  		kobject_init(&s->kobj, &slab_ktype);
5603  		return 0;
5604  	}
5605  
5606  	if (!unmergeable && disable_higher_order_debug &&
5607  			(slub_debug & DEBUG_METADATA_FLAGS))
5608  		unmergeable = 1;
5609  
5610  	if (unmergeable) {
5611  		/*
5612  		 * Slabcache can never be merged so we can use the name proper.
5613  		 * This is typically the case for debug situations. In that
5614  		 * case we can catch duplicate names easily.
5615  		 */
5616  		sysfs_remove_link(&slab_kset->kobj, s->name);
5617  		name = s->name;
5618  	} else {
5619  		/*
5620  		 * Create a unique name for the slab as a target
5621  		 * for the symlinks.
5622  		 */
5623  		name = create_unique_id(s);
5624  	}
5625  
5626  	s->kobj.kset = kset;
5627  	err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5628  	if (err) {
5629  		kobject_put(&s->kobj);
5630  		goto out;
5631  	}
5632  
5633  	err = sysfs_create_group(&s->kobj, &slab_attr_group);
5634  	if (err)
5635  		goto out_del_kobj;
5636  
5637  	if (!unmergeable) {
5638  		/* Setup first alias */
5639  		sysfs_slab_alias(s, s->name);
5640  	}
5641  out:
5642  	if (!unmergeable)
5643  		kfree(name);
5644  	return err;
5645  out_del_kobj:
5646  	kobject_del(&s->kobj);
5647  	goto out;
5648  }
5649  
5650  void sysfs_slab_unlink(struct kmem_cache *s)
5651  {
5652  	if (slab_state >= FULL)
5653  		kobject_del(&s->kobj);
5654  }
5655  
5656  void sysfs_slab_release(struct kmem_cache *s)
5657  {
5658  	if (slab_state >= FULL)
5659  		kobject_put(&s->kobj);
5660  }
5661  
5662  /*
5663   * Need to buffer aliases during bootup until sysfs becomes
5664   * available lest we lose that information.
5665   */
5666  struct saved_alias {
5667  	struct kmem_cache *s;
5668  	const char *name;
5669  	struct saved_alias *next;
5670  };
5671  
5672  static struct saved_alias *alias_list;
5673  
5674  static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5675  {
5676  	struct saved_alias *al;
5677  
5678  	if (slab_state == FULL) {
5679  		/*
5680  		 * If we have a leftover link then remove it.
5681  		 */
5682  		sysfs_remove_link(&slab_kset->kobj, name);
5683  		return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5684  	}
5685  
5686  	al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5687  	if (!al)
5688  		return -ENOMEM;
5689  
5690  	al->s = s;
5691  	al->name = name;
5692  	al->next = alias_list;
5693  	alias_list = al;
5694  	return 0;
5695  }
5696  
5697  static int __init slab_sysfs_init(void)
5698  {
5699  	struct kmem_cache *s;
5700  	int err;
5701  
5702  	mutex_lock(&slab_mutex);
5703  
5704  	slab_kset = kset_create_and_add("slab", NULL, kernel_kobj);
5705  	if (!slab_kset) {
5706  		mutex_unlock(&slab_mutex);
5707  		pr_err("Cannot register slab subsystem.\n");
5708  		return -ENOSYS;
5709  	}
5710  
5711  	slab_state = FULL;
5712  
5713  	list_for_each_entry(s, &slab_caches, list) {
5714  		err = sysfs_slab_add(s);
5715  		if (err)
5716  			pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5717  			       s->name);
5718  	}
5719  
5720  	while (alias_list) {
5721  		struct saved_alias *al = alias_list;
5722  
5723  		alias_list = alias_list->next;
5724  		err = sysfs_slab_alias(al->s, al->name);
5725  		if (err)
5726  			pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5727  			       al->name);
5728  		kfree(al);
5729  	}
5730  
5731  	mutex_unlock(&slab_mutex);
5732  	resiliency_test();
5733  	return 0;
5734  }
5735  
5736  __initcall(slab_sysfs_init);
5737  #endif /* CONFIG_SYSFS */
5738  
5739  /*
5740   * The /proc/slabinfo ABI
5741   */
5742  #ifdef CONFIG_SLUB_DEBUG
5743  void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5744  {
5745  	unsigned long nr_slabs = 0;
5746  	unsigned long nr_objs = 0;
5747  	unsigned long nr_free = 0;
5748  	int node;
5749  	struct kmem_cache_node *n;
5750  
5751  	for_each_kmem_cache_node(s, node, n) {
5752  		nr_slabs += node_nr_slabs(n);
5753  		nr_objs += node_nr_objs(n);
5754  		nr_free += count_partial(n, count_free);
5755  	}
5756  
5757  	sinfo->active_objs = nr_objs - nr_free;
5758  	sinfo->num_objs = nr_objs;
5759  	sinfo->active_slabs = nr_slabs;
5760  	sinfo->num_slabs = nr_slabs;
5761  	sinfo->objects_per_slab = oo_objects(s->oo);
5762  	sinfo->cache_order = oo_order(s->oo);
5763  }
5764  
5765  void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5766  {
5767  }
5768  
5769  ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5770  		       size_t count, loff_t *ppos)
5771  {
5772  	return -EIO;
5773  }
5774  #endif /* CONFIG_SLUB_DEBUG */
5775