Linux Audio

Check our new training course

Loading...
v4.6
 
   1/*
   2 * linux/mm/slab.c
   3 * Written by Mark Hemment, 1996/97.
   4 * (markhe@nextd.demon.co.uk)
   5 *
   6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
   7 *
   8 * Major cleanup, different bufctl logic, per-cpu arrays
   9 *	(c) 2000 Manfred Spraul
  10 *
  11 * Cleanup, make the head arrays unconditional, preparation for NUMA
  12 * 	(c) 2002 Manfred Spraul
  13 *
  14 * An implementation of the Slab Allocator as described in outline in;
  15 *	UNIX Internals: The New Frontiers by Uresh Vahalia
  16 *	Pub: Prentice Hall	ISBN 0-13-101908-2
  17 * or with a little more detail in;
  18 *	The Slab Allocator: An Object-Caching Kernel Memory Allocator
  19 *	Jeff Bonwick (Sun Microsystems).
  20 *	Presented at: USENIX Summer 1994 Technical Conference
  21 *
  22 * The memory is organized in caches, one cache for each object type.
  23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
  24 * Each cache consists out of many slabs (they are small (usually one
  25 * page long) and always contiguous), and each slab contains multiple
  26 * initialized objects.
  27 *
  28 * This means, that your constructor is used only for newly allocated
  29 * slabs and you must pass objects with the same initializations to
  30 * kmem_cache_free.
  31 *
  32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
  33 * normal). If you need a special memory type, then must create a new
  34 * cache for that memory type.
  35 *
  36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
  37 *   full slabs with 0 free objects
  38 *   partial slabs
  39 *   empty slabs with no allocated objects
  40 *
  41 * If partial slabs exist, then new allocations come from these slabs,
  42 * otherwise from empty slabs or new slabs are allocated.
  43 *
  44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
  45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
  46 *
  47 * Each cache has a short per-cpu head array, most allocs
  48 * and frees go into that array, and if that array overflows, then 1/2
  49 * of the entries in the array are given back into the global cache.
  50 * The head array is strictly LIFO and should improve the cache hit rates.
  51 * On SMP, it additionally reduces the spinlock operations.
  52 *
  53 * The c_cpuarray may not be read with enabled local interrupts -
  54 * it's changed with a smp_call_function().
  55 *
  56 * SMP synchronization:
  57 *  constructors and destructors are called without any locking.
  58 *  Several members in struct kmem_cache and struct slab never change, they
  59 *	are accessed without any locking.
  60 *  The per-cpu arrays are never accessed from the wrong cpu, no locking,
  61 *  	and local interrupts are disabled so slab code is preempt-safe.
  62 *  The non-constant members are protected with a per-cache irq spinlock.
  63 *
  64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
  65 * in 2000 - many ideas in the current implementation are derived from
  66 * his patch.
  67 *
  68 * Further notes from the original documentation:
  69 *
  70 * 11 April '97.  Started multi-threading - markhe
  71 *	The global cache-chain is protected by the mutex 'slab_mutex'.
  72 *	The sem is only needed when accessing/extending the cache-chain, which
  73 *	can never happen inside an interrupt (kmem_cache_create(),
  74 *	kmem_cache_shrink() and kmem_cache_reap()).
  75 *
  76 *	At present, each engine can be growing a cache.  This should be blocked.
  77 *
  78 * 15 March 2005. NUMA slab allocator.
  79 *	Shai Fultheim <shai@scalex86.org>.
  80 *	Shobhit Dayal <shobhit@calsoftinc.com>
  81 *	Alok N Kataria <alokk@calsoftinc.com>
  82 *	Christoph Lameter <christoph@lameter.com>
  83 *
  84 *	Modified the slab allocator to be node aware on NUMA systems.
  85 *	Each node has its own list of partial, free and full slabs.
  86 *	All object allocations for a node occur from node specific slab lists.
  87 */
  88
  89#include	<linux/__KEEPIDENTS__B.h>
  90#include	<linux/__KEEPIDENTS__C.h>
  91#include	<linux/__KEEPIDENTS__D.h>
  92#include	<linux/__KEEPIDENTS__E.h>
  93#include	<linux/__KEEPIDENTS__F.h>
  94#include	<linux/__KEEPIDENTS__G.h>
  95#include	<linux/__KEEPIDENTS__H.h>
  96#include	<linux/__KEEPIDENTS__I.h>
  97#include	<linux/__KEEPIDENTS__J.h>
  98#include	<linux/proc_fs.h>
  99#include	<linux/__KEEPIDENTS__BA.h>
 100#include	<linux/__KEEPIDENTS__BB.h>
 101#include	<linux/__KEEPIDENTS__BC.h>
 102#include	<linux/cpu.h>
 103#include	<linux/__KEEPIDENTS__BD.h>
 104#include	<linux/__KEEPIDENTS__BE.h>
 105#include	<linux/rcupdate.h>
 106#include	<linux/__KEEPIDENTS__BF.h>
 107#include	<linux/__KEEPIDENTS__BG.h>
 108#include	<linux/__KEEPIDENTS__BH.h>
 109#include	<linux/kmemleak.h>
 110#include	<linux/__KEEPIDENTS__BI.h>
 111#include	<linux/__KEEPIDENTS__BJ.h>
 112#include	<linux/__KEEPIDENTS__CA-__KEEPIDENTS__CB.h>
 113#include	<linux/__KEEPIDENTS__CC.h>
 114#include	<linux/reciprocal_div.h>
 115#include	<linux/debugobjects.h>
 116#include	<linux/kmemcheck.h>
 117#include	<linux/__KEEPIDENTS__CD.h>
 118#include	<linux/__KEEPIDENTS__CE.h>
 
 119
 120#include	<net/__KEEPIDENTS__CF.h>
 121
 122#include	<asm/cacheflush.h>
 123#include	<asm/tlbflush.h>
 124#include	<asm/page.h>
 125
 126#include <trace/events/kmem.h>
 127
 128#include	"internal.h"
 129
 130#include	"slab.h"
 131
 132/*
 133 * DEBUG	- 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
 134 *		  0 for faster, smaller code (especially in the critical paths).
 135 *
 136 * STATS	- 1 to collect stats for /proc/slabinfo.
 137 *		  0 for faster, smaller code (especially in the critical paths).
 138 *
 139 * FORCED_DEBUG	- 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
 140 */
 141
 142#ifdef CONFIG_DEBUG_SLAB
 143#define	DEBUG		1
 144#define	STATS		1
 145#define	FORCED_DEBUG	1
 146#else
 147#define	DEBUG		0
 148#define	STATS		0
 149#define	FORCED_DEBUG	0
 150#endif
 151
 152/* Shouldn't this be in a header file somewhere? */
 153#define	BYTES_PER_WORD		sizeof(void *)
 154#define	REDZONE_ALIGN		max(BYTES_PER_WORD, __alignof__(unsigned long long))
 155
 156#ifndef ARCH_KMALLOC_FLAGS
 157#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
 158#endif
 159
 160#define FREELIST_BYTE_INDEX (((PAGE_SIZE >> BITS_PER_BYTE) \
 161				<= SLAB_OBJ_MIN_SIZE) ? 1 : 0)
 162
 163#if FREELIST_BYTE_INDEX
 164typedef unsigned char freelist_idx_t;
 165#else
 166typedef unsigned short freelist_idx_t;
 167#endif
 168
 169#define SLAB_OBJ_MAX_NUM ((1 << sizeof(freelist_idx_t) * BITS_PER_BYTE) - 1)
 170
 171/*
 172 * struct array_cache
 173 *
 174 * Purpose:
 175 * - LIFO ordering, to hand out cache-warm objects from _alloc
 176 * - reduce the number of linked list operations
 177 * - reduce spinlock operations
 178 *
 179 * The limit is stored in the per-cpu structure to reduce the data cache
 180 * footprint.
 181 *
 182 */
 183struct array_cache {
 184	unsigned int avail;
 185	unsigned int limit;
 186	unsigned int batchcount;
 187	unsigned int touched;
 188	void *entry[];	/*
 189			 * Must have this definition in here for the proper
 190			 * alignment of array_cache. Also simplifies accessing
 191			 * the entries.
 192			 */
 193};
 194
 195struct alien_cache {
 196	spinlock_t lock;
 197	struct array_cache ac;
 198};
 199
 200/*
 201 * Need this for bootstrapping a per node allocator.
 202 */
 203#define NUM_INIT_LISTS (2 * MAX_NUMNODES)
 204static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
 205#define	CACHE_CACHE 0
 206#define	SIZE_NODE (MAX_NUMNODES)
 207
 208static int drain_freelist(struct kmem_cache *cache,
 209			struct kmem_cache_node *n, int tofree);
 210static void free_block(struct kmem_cache *cachep, void **objpp, int len,
 211			int node, struct list_head *list);
 212static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list);
 213static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
 214static void cache_reap(struct work_struct *unused);
 215
 
 
 
 
 
 216static int slab_early_init = 1;
 217
 218#define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
 219
 220static void kmem_cache_node_init(struct kmem_cache_node *parent)
 221{
 222	INIT_LIST_HEAD(&parent->slabs_full);
 223	INIT_LIST_HEAD(&parent->slabs_partial);
 224	INIT_LIST_HEAD(&parent->slabs_free);
 
 
 225	parent->shared = NULL;
 226	parent->alien = NULL;
 227	parent->colour_next = 0;
 228	spin_lock_init(&parent->list_lock);
 229	parent->free_objects = 0;
 230	parent->free_touched = 0;
 231}
 232
 233#define MAKE_LIST(cachep, listp, slab, nodeid)				\
 234	do {								\
 235		INIT_LIST_HEAD(listp);					\
 236		list_splice(&get_node(cachep, nodeid)->slab, listp);	\
 237	} while (0)
 238
 239#define	MAKE_ALL_LISTS(cachep, ptr, nodeid)				\
 240	do {								\
 241	MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid);	\
 242	MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
 243	MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid);	\
 244	} while (0)
 245
 246#define CFLGS_OBJFREELIST_SLAB	(0x40000000UL)
 247#define CFLGS_OFF_SLAB		(0x80000000UL)
 248#define	OBJFREELIST_SLAB(x)	((x)->flags & CFLGS_OBJFREELIST_SLAB)
 249#define	OFF_SLAB(x)	((x)->flags & CFLGS_OFF_SLAB)
 250
 251#define BATCHREFILL_LIMIT	16
 252/*
 253 * Optimization question: fewer reaps means less probability for unnessary
 254 * cpucache drain/refill cycles.
 255 *
 256 * OTOH the cpuarrays can contain lots of objects,
 257 * which could lock up otherwise freeable slabs.
 258 */
 259#define REAPTIMEOUT_AC		(2*HZ)
 260#define REAPTIMEOUT_NODE	(4*HZ)
 261
 262#if STATS
 263#define	STATS_INC_ACTIVE(x)	((x)->num_active++)
 264#define	STATS_DEC_ACTIVE(x)	((x)->num_active--)
 265#define	STATS_INC_ALLOCED(x)	((x)->num_allocations++)
 266#define	STATS_INC_GROWN(x)	((x)->grown++)
 267#define	STATS_ADD_REAPED(x,y)	((x)->reaped += (y))
 268#define	STATS_SET_HIGH(x)						\
 269	do {								\
 270		if ((x)->num_active > (x)->high_mark)			\
 271			(x)->high_mark = (x)->num_active;		\
 272	} while (0)
 273#define	STATS_INC_ERR(x)	((x)->errors++)
 274#define	STATS_INC_NODEALLOCS(x)	((x)->node_allocs++)
 275#define	STATS_INC_NODEFREES(x)	((x)->node_frees++)
 276#define STATS_INC_ACOVERFLOW(x)   ((x)->node_overflow++)
 277#define	STATS_SET_FREEABLE(x, i)					\
 278	do {								\
 279		if ((x)->max_freeable < i)				\
 280			(x)->max_freeable = i;				\
 281	} while (0)
 282#define STATS_INC_ALLOCHIT(x)	atomic_inc(&(x)->allochit)
 283#define STATS_INC_ALLOCMISS(x)	atomic_inc(&(x)->allocmiss)
 284#define STATS_INC_FREEHIT(x)	atomic_inc(&(x)->freehit)
 285#define STATS_INC_FREEMISS(x)	atomic_inc(&(x)->freemiss)
 286#else
 287#define	STATS_INC_ACTIVE(x)	do { } while (0)
 288#define	STATS_DEC_ACTIVE(x)	do { } while (0)
 289#define	STATS_INC_ALLOCED(x)	do { } while (0)
 290#define	STATS_INC_GROWN(x)	do { } while (0)
 291#define	STATS_ADD_REAPED(x,y)	do { (void)(y); } while (0)
 292#define	STATS_SET_HIGH(x)	do { } while (0)
 293#define	STATS_INC_ERR(x)	do { } while (0)
 294#define	STATS_INC_NODEALLOCS(x)	do { } while (0)
 295#define	STATS_INC_NODEFREES(x)	do { } while (0)
 296#define STATS_INC_ACOVERFLOW(x)   do { } while (0)
 297#define	STATS_SET_FREEABLE(x, i) do { } while (0)
 298#define STATS_INC_ALLOCHIT(x)	do { } while (0)
 299#define STATS_INC_ALLOCMISS(x)	do { } while (0)
 300#define STATS_INC_FREEHIT(x)	do { } while (0)
 301#define STATS_INC_FREEMISS(x)	do { } while (0)
 302#endif
 303
 304#if DEBUG
 305
 306/*
 307 * memory layout of objects:
 308 * 0		: objp
 309 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
 310 * 		the end of an object is aligned with the end of the real
 311 * 		allocation. Catches writes behind the end of the allocation.
 312 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
 313 * 		redzone word.
 314 * cachep->obj_offset: The real object.
 315 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
 316 * cachep->size - 1* BYTES_PER_WORD: last caller address
 317 *					[BYTES_PER_WORD long]
 318 */
 319static int obj_offset(struct kmem_cache *cachep)
 320{
 321	return cachep->obj_offset;
 322}
 323
 324static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
 325{
 326	BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
 327	return (unsigned long long*) (objp + obj_offset(cachep) -
 328				      sizeof(unsigned long long));
 329}
 330
 331static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
 332{
 333	BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
 334	if (cachep->flags & SLAB_STORE_USER)
 335		return (unsigned long long *)(objp + cachep->size -
 336					      sizeof(unsigned long long) -
 337					      REDZONE_ALIGN);
 338	return (unsigned long long *) (objp + cachep->size -
 339				       sizeof(unsigned long long));
 340}
 341
 342static void **dbg_userword(struct kmem_cache *cachep, void *objp)
 343{
 344	BUG_ON(!(cachep->flags & SLAB_STORE_USER));
 345	return (void **)(objp + cachep->size - BYTES_PER_WORD);
 346}
 347
 348#else
 349
 350#define obj_offset(x)			0
 351#define dbg_redzone1(cachep, objp)	({BUG(); (unsigned long long *)NULL;})
 352#define dbg_redzone2(cachep, objp)	({BUG(); (unsigned long long *)NULL;})
 353#define dbg_userword(cachep, objp)	({BUG(); (void **)NULL;})
 354
 355#endif
 356
 357#ifdef CONFIG_DEBUG_SLAB_LEAK
 358
 359static inline bool is_store_user_clean(struct kmem_cache *cachep)
 360{
 361	return atomic_read(&cachep->store_user_clean) == 1;
 362}
 363
 364static inline void set_store_user_clean(struct kmem_cache *cachep)
 365{
 366	atomic_set(&cachep->store_user_clean, 1);
 367}
 368
 369static inline void set_store_user_dirty(struct kmem_cache *cachep)
 370{
 371	if (is_store_user_clean(cachep))
 372		atomic_set(&cachep->store_user_clean, 0);
 373}
 374
 375#else
 376static inline void set_store_user_dirty(struct kmem_cache *cachep) {}
 377
 378#endif
 379
 380/*
 381 * Do not go above this order unless 0 objects fit into the slab or
 382 * overridden on the command line.
 383 */
 384#define	SLAB_MAX_ORDER_HI	1
 385#define	SLAB_MAX_ORDER_LO	0
 386static int slab_max_order = SLAB_MAX_ORDER_LO;
 387static bool slab_max_order_set __initdata;
 388
 389static inline struct kmem_cache *virt_to_cache(const void *obj)
 390{
 391	struct page *page = virt_to_head_page(obj);
 392	return page->slab_cache;
 393}
 394
 395static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
 396				 unsigned int idx)
 397{
 398	return page->s_mem + cache->size * idx;
 399}
 400
 401/*
 402 * We want to avoid an expensive divide : (offset / cache->size)
 403 *   Using the fact that size is a constant for a particular cache,
 404 *   we can replace (offset / cache->size) by
 405 *   reciprocal_divide(offset, cache->reciprocal_buffer_size)
 406 */
 407static inline unsigned int obj_to_index(const struct kmem_cache *cache,
 408					const struct page *page, void *obj)
 409{
 410	u32 offset = (obj - page->s_mem);
 411	return reciprocal_divide(offset, cache->reciprocal_buffer_size);
 412}
 413
 414#define BOOT_CPUCACHE_ENTRIES	1
 415/* internal cache of cache description objs */
 416static struct kmem_cache kmem_cache_boot = {
 417	.batchcount = 1,
 418	.limit = BOOT_CPUCACHE_ENTRIES,
 419	.shared = 1,
 420	.size = sizeof(struct kmem_cache),
 421	.name = "kmem_cache",
 422};
 423
 424#define BAD_ALIEN_MAGIC 0x01020304ul
 425
 426static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
 427
 428static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
 429{
 430	return this_cpu_ptr(cachep->cpu_cache);
 431}
 432
 433/*
 434 * Calculate the number of objects and left-over bytes for a given buffer size.
 435 */
 436static unsigned int cache_estimate(unsigned long gfporder, size_t buffer_size,
 437		unsigned long flags, size_t *left_over)
 438{
 439	unsigned int num;
 440	size_t slab_size = PAGE_SIZE << gfporder;
 441
 442	/*
 443	 * The slab management structure can be either off the slab or
 444	 * on it. For the latter case, the memory allocated for a
 445	 * slab is used for:
 446	 *
 447	 * - @buffer_size bytes for each object
 448	 * - One freelist_idx_t for each object
 449	 *
 450	 * We don't need to consider alignment of freelist because
 451	 * freelist will be at the end of slab page. The objects will be
 452	 * at the correct alignment.
 453	 *
 454	 * If the slab management structure is off the slab, then the
 455	 * alignment will already be calculated into the size. Because
 456	 * the slabs are all pages aligned, the objects will be at the
 457	 * correct alignment when allocated.
 458	 */
 459	if (flags & (CFLGS_OBJFREELIST_SLAB | CFLGS_OFF_SLAB)) {
 460		num = slab_size / buffer_size;
 461		*left_over = slab_size % buffer_size;
 462	} else {
 463		num = slab_size / (buffer_size + sizeof(freelist_idx_t));
 464		*left_over = slab_size %
 465			(buffer_size + sizeof(freelist_idx_t));
 466	}
 467
 468	return num;
 469}
 470
 471#if DEBUG
 472#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
 473
 474static void __slab_error(const char *function, struct kmem_cache *cachep,
 475			char *msg)
 476{
 477	pr_err("slab error in %s(): cache `%s': %s\n",
 478	       function, cachep->name, msg);
 479	dump_stack();
 480	add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
 481}
 482#endif
 483
 484/*
 485 * By default on NUMA we use alien caches to stage the freeing of
 486 * objects allocated from other nodes. This causes massive memory
 487 * inefficiencies when using fake NUMA setup to split memory into a
 488 * large number of small nodes, so it can be disabled on the command
 489 * line
 490  */
 491
 492static int use_alien_caches __read_mostly = 1;
 493static int __init noaliencache_setup(char *s)
 494{
 495	use_alien_caches = 0;
 496	return 1;
 497}
 498__setup("noaliencache", noaliencache_setup);
 499
 500static int __init slab_max_order_setup(char *str)
 501{
 502	get_option(&str, &slab_max_order);
 503	slab_max_order = slab_max_order < 0 ? 0 :
 504				min(slab_max_order, MAX_ORDER - 1);
 505	slab_max_order_set = true;
 506
 507	return 1;
 508}
 509__setup("slab_max_order=", slab_max_order_setup);
 510
 511#ifdef CONFIG_NUMA
 512/*
 513 * Special reaping functions for NUMA systems called from cache_reap().
 514 * These take care of doing round robin flushing of alien caches (containing
 515 * objects freed on different nodes from which they were allocated) and the
 516 * flushing of remote pcps by calling drain_node_pages.
 517 */
 518static DEFINE_PER_CPU(unsigned long, slab_reap_node);
 519
 520static void init_reap_node(int cpu)
 521{
 522	int node;
 523
 524	node = next_node(cpu_to_mem(cpu), node_online_map);
 525	if (node == MAX_NUMNODES)
 526		node = first_node(node_online_map);
 527
 528	per_cpu(slab_reap_node, cpu) = node;
 529}
 530
 531static void next_reap_node(void)
 532{
 533	int node = __this_cpu_read(slab_reap_node);
 534
 535	node = next_node(node, node_online_map);
 536	if (unlikely(node >= MAX_NUMNODES))
 537		node = first_node(node_online_map);
 538	__this_cpu_write(slab_reap_node, node);
 539}
 540
 541#else
 542#define init_reap_node(cpu) do { } while (0)
 543#define next_reap_node(void) do { } while (0)
 544#endif
 545
 546/*
 547 * Initiate the reap timer running on the target CPU.  We run at around 1 to 2Hz
 548 * via the workqueue/eventd.
 549 * Add the CPU number into the expiration time to minimize the possibility of
 550 * the CPUs getting into lockstep and contending for the global cache chain
 551 * lock.
 552 */
 553static void start_cpu_timer(int cpu)
 554{
 555	struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
 556
 557	/*
 558	 * When this gets called from do_initcalls via cpucache_init(),
 559	 * init_workqueues() has already run, so keventd will be setup
 560	 * at that time.
 561	 */
 562	if (keventd_up() && reap_work->work.func == NULL) {
 563		init_reap_node(cpu);
 564		INIT_DEFERRABLE_WORK(reap_work, cache_reap);
 565		schedule_delayed_work_on(cpu, reap_work,
 566					__round_jiffies_relative(HZ, cpu));
 567	}
 568}
 569
 570static void init_arraycache(struct array_cache *ac, int limit, int batch)
 571{
 572	/*
 573	 * The array_cache structures contain pointers to free object.
 574	 * However, when such objects are allocated or transferred to another
 575	 * cache the pointers are not cleared and they could be counted as
 576	 * valid references during a kmemleak scan. Therefore, kmemleak must
 577	 * not scan such objects.
 578	 */
 579	kmemleak_no_scan(ac);
 580	if (ac) {
 581		ac->avail = 0;
 582		ac->limit = limit;
 583		ac->batchcount = batch;
 584		ac->touched = 0;
 585	}
 586}
 587
 588static struct array_cache *alloc_arraycache(int node, int entries,
 589					    int batchcount, gfp_t gfp)
 590{
 591	size_t memsize = sizeof(void *) * entries + sizeof(struct array_cache);
 592	struct array_cache *ac = NULL;
 593
 594	ac = kmalloc_node(memsize, gfp, node);
 595	init_arraycache(ac, entries, batchcount);
 596	return ac;
 597}
 598
 599static noinline void cache_free_pfmemalloc(struct kmem_cache *cachep,
 600					struct page *page, void *objp)
 601{
 602	struct kmem_cache_node *n;
 603	int page_node;
 604	LIST_HEAD(list);
 605
 606	page_node = page_to_nid(page);
 607	n = get_node(cachep, page_node);
 608
 609	spin_lock(&n->list_lock);
 610	free_block(cachep, &objp, 1, page_node, &list);
 611	spin_unlock(&n->list_lock);
 612
 613	slabs_destroy(cachep, &list);
 614}
 615
 616/*
 617 * Transfer objects in one arraycache to another.
 618 * Locking must be handled by the caller.
 619 *
 620 * Return the number of entries transferred.
 621 */
 622static int transfer_objects(struct array_cache *to,
 623		struct array_cache *from, unsigned int max)
 624{
 625	/* Figure out how many entries to transfer */
 626	int nr = min3(from->avail, max, to->limit - to->avail);
 627
 628	if (!nr)
 629		return 0;
 630
 631	memcpy(to->entry + to->avail, from->entry + from->avail -nr,
 632			sizeof(void *) *nr);
 633
 634	from->avail -= nr;
 635	to->avail += nr;
 636	return nr;
 637}
 638
 639#ifndef CONFIG_NUMA
 640
 641#define drain_alien_cache(cachep, alien) do { } while (0)
 642#define reap_alien(cachep, n) do { } while (0)
 643
 644static inline struct alien_cache **alloc_alien_cache(int node,
 645						int limit, gfp_t gfp)
 646{
 647	return (struct alien_cache **)BAD_ALIEN_MAGIC;
 648}
 649
 650static inline void free_alien_cache(struct alien_cache **ac_ptr)
 651{
 652}
 653
 654static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
 655{
 656	return 0;
 657}
 658
 659static inline void *alternate_node_alloc(struct kmem_cache *cachep,
 660		gfp_t flags)
 661{
 662	return NULL;
 663}
 664
 665static inline void *____cache_alloc_node(struct kmem_cache *cachep,
 666		 gfp_t flags, int nodeid)
 667{
 668	return NULL;
 669}
 670
 671static inline gfp_t gfp_exact_node(gfp_t flags)
 672{
 673	return flags & ~__GFP_NOFAIL;
 674}
 675
 676#else	/* CONFIG_NUMA */
 677
 678static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
 679static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
 680
 681static struct alien_cache *__alloc_alien_cache(int node, int entries,
 682						int batch, gfp_t gfp)
 683{
 684	size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache);
 685	struct alien_cache *alc = NULL;
 686
 687	alc = kmalloc_node(memsize, gfp, node);
 688	init_arraycache(&alc->ac, entries, batch);
 689	spin_lock_init(&alc->lock);
 690	return alc;
 691}
 692
 693static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
 694{
 695	struct alien_cache **alc_ptr;
 696	size_t memsize = sizeof(void *) * nr_node_ids;
 697	int i;
 698
 699	if (limit > 1)
 700		limit = 12;
 701	alc_ptr = kzalloc_node(memsize, gfp, node);
 702	if (!alc_ptr)
 703		return NULL;
 704
 705	for_each_node(i) {
 706		if (i == node || !node_online(i))
 707			continue;
 708		alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp);
 709		if (!alc_ptr[i]) {
 710			for (i--; i >= 0; i--)
 711				kfree(alc_ptr[i]);
 712			kfree(alc_ptr);
 713			return NULL;
 714		}
 715	}
 716	return alc_ptr;
 717}
 718
 719static void free_alien_cache(struct alien_cache **alc_ptr)
 720{
 721	int i;
 722
 723	if (!alc_ptr)
 724		return;
 725	for_each_node(i)
 726	    kfree(alc_ptr[i]);
 727	kfree(alc_ptr);
 728}
 729
 730static void __drain_alien_cache(struct kmem_cache *cachep,
 731				struct array_cache *ac, int node,
 732				struct list_head *list)
 733{
 734	struct kmem_cache_node *n = get_node(cachep, node);
 735
 736	if (ac->avail) {
 737		spin_lock(&n->list_lock);
 738		/*
 739		 * Stuff objects into the remote nodes shared array first.
 740		 * That way we could avoid the overhead of putting the objects
 741		 * into the free lists and getting them back later.
 742		 */
 743		if (n->shared)
 744			transfer_objects(n->shared, ac, ac->limit);
 745
 746		free_block(cachep, ac->entry, ac->avail, node, list);
 747		ac->avail = 0;
 748		spin_unlock(&n->list_lock);
 749	}
 750}
 751
 752/*
 753 * Called from cache_reap() to regularly drain alien caches round robin.
 754 */
 755static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
 756{
 757	int node = __this_cpu_read(slab_reap_node);
 758
 759	if (n->alien) {
 760		struct alien_cache *alc = n->alien[node];
 761		struct array_cache *ac;
 762
 763		if (alc) {
 764			ac = &alc->ac;
 765			if (ac->avail && spin_trylock_irq(&alc->lock)) {
 766				LIST_HEAD(list);
 767
 768				__drain_alien_cache(cachep, ac, node, &list);
 769				spin_unlock_irq(&alc->lock);
 770				slabs_destroy(cachep, &list);
 771			}
 772		}
 773	}
 774}
 775
 776static void drain_alien_cache(struct kmem_cache *cachep,
 777				struct alien_cache **alien)
 778{
 779	int i = 0;
 780	struct alien_cache *alc;
 781	struct array_cache *ac;
 782	unsigned long flags;
 783
 784	for_each_online_node(i) {
 785		alc = alien[i];
 786		if (alc) {
 787			LIST_HEAD(list);
 788
 789			ac = &alc->ac;
 790			spin_lock_irqsave(&alc->lock, flags);
 791			__drain_alien_cache(cachep, ac, i, &list);
 792			spin_unlock_irqrestore(&alc->lock, flags);
 793			slabs_destroy(cachep, &list);
 794		}
 795	}
 796}
 797
 798static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
 799				int node, int page_node)
 800{
 801	struct kmem_cache_node *n;
 802	struct alien_cache *alien = NULL;
 803	struct array_cache *ac;
 804	LIST_HEAD(list);
 805
 806	n = get_node(cachep, node);
 807	STATS_INC_NODEFREES(cachep);
 808	if (n->alien && n->alien[page_node]) {
 809		alien = n->alien[page_node];
 810		ac = &alien->ac;
 811		spin_lock(&alien->lock);
 812		if (unlikely(ac->avail == ac->limit)) {
 813			STATS_INC_ACOVERFLOW(cachep);
 814			__drain_alien_cache(cachep, ac, page_node, &list);
 815		}
 816		ac->entry[ac->avail++] = objp;
 817		spin_unlock(&alien->lock);
 818		slabs_destroy(cachep, &list);
 819	} else {
 820		n = get_node(cachep, page_node);
 821		spin_lock(&n->list_lock);
 822		free_block(cachep, &objp, 1, page_node, &list);
 823		spin_unlock(&n->list_lock);
 824		slabs_destroy(cachep, &list);
 825	}
 826	return 1;
 827}
 828
 829static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
 830{
 831	int page_node = page_to_nid(virt_to_page(objp));
 832	int node = numa_mem_id();
 833	/*
 834	 * Make sure we are not freeing a object from another node to the array
 835	 * cache on this cpu.
 836	 */
 837	if (likely(node == page_node))
 838		return 0;
 839
 840	return __cache_free_alien(cachep, objp, node, page_node);
 841}
 842
 843/*
 844 * Construct gfp mask to allocate from a specific node but do not reclaim or
 845 * warn about failures.
 846 */
 847static inline gfp_t gfp_exact_node(gfp_t flags)
 848{
 849	return (flags | __GFP_THISNODE | __GFP_NOWARN) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
 850}
 851#endif
 852
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 853/*
 854 * Allocates and initializes node for a node on each slab cache, used for
 855 * either memory or cpu hotplug.  If memory is being hot-added, the kmem_cache_node
 856 * will be allocated off-node since memory is not yet online for the new node.
 857 * When hotplugging memory or a cpu, existing node are not replaced if
 858 * already in use.
 859 *
 860 * Must hold slab_mutex.
 861 */
 862static int init_cache_node_node(int node)
 863{
 
 864	struct kmem_cache *cachep;
 865	struct kmem_cache_node *n;
 866	const size_t memsize = sizeof(struct kmem_cache_node);
 867
 868	list_for_each_entry(cachep, &slab_caches, list) {
 869		/*
 870		 * Set up the kmem_cache_node for cpu before we can
 871		 * begin anything. Make sure some other cpu on this
 872		 * node has not already allocated this
 873		 */
 874		n = get_node(cachep, node);
 875		if (!n) {
 876			n = kmalloc_node(memsize, GFP_KERNEL, node);
 877			if (!n)
 878				return -ENOMEM;
 879			kmem_cache_node_init(n);
 880			n->next_reap = jiffies + REAPTIMEOUT_NODE +
 881			    ((unsigned long)cachep) % REAPTIMEOUT_NODE;
 882
 883			/*
 884			 * The kmem_cache_nodes don't come and go as CPUs
 885			 * come and go.  slab_mutex is sufficient
 886			 * protection here.
 887			 */
 888			cachep->node[node] = n;
 889		}
 890
 891		spin_lock_irq(&n->list_lock);
 892		n->free_limit =
 893			(1 + nr_cpus_node(node)) *
 894			cachep->batchcount + cachep->num;
 895		spin_unlock_irq(&n->list_lock);
 896	}
 
 897	return 0;
 898}
 
 899
 900static inline int slabs_tofree(struct kmem_cache *cachep,
 901						struct kmem_cache_node *n)
 902{
 903	return (n->free_objects + cachep->num - 1) / cachep->num;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 904}
 905
 
 
 906static void cpuup_canceled(long cpu)
 907{
 908	struct kmem_cache *cachep;
 909	struct kmem_cache_node *n = NULL;
 910	int node = cpu_to_mem(cpu);
 911	const struct cpumask *mask = cpumask_of_node(node);
 912
 913	list_for_each_entry(cachep, &slab_caches, list) {
 914		struct array_cache *nc;
 915		struct array_cache *shared;
 916		struct alien_cache **alien;
 917		LIST_HEAD(list);
 918
 919		n = get_node(cachep, node);
 920		if (!n)
 921			continue;
 922
 923		spin_lock_irq(&n->list_lock);
 924
 925		/* Free limit for this kmem_cache_node */
 926		n->free_limit -= cachep->batchcount;
 927
 928		/* cpu is dead; no one can alloc from it. */
 929		nc = per_cpu_ptr(cachep->cpu_cache, cpu);
 930		if (nc) {
 931			free_block(cachep, nc->entry, nc->avail, node, &list);
 932			nc->avail = 0;
 933		}
 934
 935		if (!cpumask_empty(mask)) {
 936			spin_unlock_irq(&n->list_lock);
 937			goto free_slab;
 938		}
 939
 940		shared = n->shared;
 941		if (shared) {
 942			free_block(cachep, shared->entry,
 943				   shared->avail, node, &list);
 944			n->shared = NULL;
 945		}
 946
 947		alien = n->alien;
 948		n->alien = NULL;
 949
 950		spin_unlock_irq(&n->list_lock);
 951
 952		kfree(shared);
 953		if (alien) {
 954			drain_alien_cache(cachep, alien);
 955			free_alien_cache(alien);
 956		}
 957
 958free_slab:
 959		slabs_destroy(cachep, &list);
 960	}
 961	/*
 962	 * In the previous loop, all the objects were freed to
 963	 * the respective cache's slabs,  now we can go ahead and
 964	 * shrink each nodelist to its limit.
 965	 */
 966	list_for_each_entry(cachep, &slab_caches, list) {
 967		n = get_node(cachep, node);
 968		if (!n)
 969			continue;
 970		drain_freelist(cachep, n, slabs_tofree(cachep, n));
 971	}
 972}
 973
 974static int cpuup_prepare(long cpu)
 975{
 976	struct kmem_cache *cachep;
 977	struct kmem_cache_node *n = NULL;
 978	int node = cpu_to_mem(cpu);
 979	int err;
 980
 981	/*
 982	 * We need to do this right in the beginning since
 983	 * alloc_arraycache's are going to use this list.
 984	 * kmalloc_node allows us to add the slab to the right
 985	 * kmem_cache_node and not this cpu's kmem_cache_node
 986	 */
 987	err = init_cache_node_node(node);
 988	if (err < 0)
 989		goto bad;
 990
 991	/*
 992	 * Now we can go ahead with allocating the shared arrays and
 993	 * array caches
 994	 */
 995	list_for_each_entry(cachep, &slab_caches, list) {
 996		struct array_cache *shared = NULL;
 997		struct alien_cache **alien = NULL;
 998
 999		if (cachep->shared) {
1000			shared = alloc_arraycache(node,
1001				cachep->shared * cachep->batchcount,
1002				0xbaadf00d, GFP_KERNEL);
1003			if (!shared)
1004				goto bad;
1005		}
1006		if (use_alien_caches) {
1007			alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1008			if (!alien) {
1009				kfree(shared);
1010				goto bad;
1011			}
1012		}
1013		n = get_node(cachep, node);
1014		BUG_ON(!n);
1015
1016		spin_lock_irq(&n->list_lock);
1017		if (!n->shared) {
1018			/*
1019			 * We are serialised from CPU_DEAD or
1020			 * CPU_UP_CANCELLED by the cpucontrol lock
1021			 */
1022			n->shared = shared;
1023			shared = NULL;
1024		}
1025#ifdef CONFIG_NUMA
1026		if (!n->alien) {
1027			n->alien = alien;
1028			alien = NULL;
1029		}
1030#endif
1031		spin_unlock_irq(&n->list_lock);
1032		kfree(shared);
1033		free_alien_cache(alien);
1034	}
1035
1036	return 0;
1037bad:
1038	cpuup_canceled(cpu);
1039	return -ENOMEM;
1040}
1041
1042static int cpuup_callback(struct notifier_block *nfb,
1043				    unsigned long action, void *hcpu)
1044{
1045	long cpu = (long)hcpu;
1046	int err = 0;
1047
1048	switch (action) {
1049	case CPU_UP_PREPARE:
1050	case CPU_UP_PREPARE_FROZEN:
1051		mutex_lock(&slab_mutex);
1052		err = cpuup_prepare(cpu);
1053		mutex_unlock(&slab_mutex);
1054		break;
1055	case CPU_ONLINE:
1056	case CPU_ONLINE_FROZEN:
1057		start_cpu_timer(cpu);
1058		break;
1059#ifdef CONFIG_HOTPLUG_CPU
1060  	case CPU_DOWN_PREPARE:
1061  	case CPU_DOWN_PREPARE_FROZEN:
1062		/*
1063		 * Shutdown cache reaper. Note that the slab_mutex is
1064		 * held so that if cache_reap() is invoked it cannot do
1065		 * anything expensive but will only modify reap_work
1066		 * and reschedule the timer.
1067		*/
1068		cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1069		/* Now the cache_reaper is guaranteed to be not running. */
1070		per_cpu(slab_reap_work, cpu).work.func = NULL;
1071  		break;
1072  	case CPU_DOWN_FAILED:
1073  	case CPU_DOWN_FAILED_FROZEN:
1074		start_cpu_timer(cpu);
1075  		break;
1076	case CPU_DEAD:
1077	case CPU_DEAD_FROZEN:
1078		/*
1079		 * Even if all the cpus of a node are down, we don't free the
1080		 * kmem_cache_node of any cache. This to avoid a race between
1081		 * cpu_down, and a kmalloc allocation from another cpu for
1082		 * memory from the node of the cpu going down.  The node
1083		 * structure is usually allocated from kmem_cache_create() and
1084		 * gets destroyed at kmem_cache_destroy().
1085		 */
1086		/* fall through */
1087#endif
1088	case CPU_UP_CANCELED:
1089	case CPU_UP_CANCELED_FROZEN:
1090		mutex_lock(&slab_mutex);
1091		cpuup_canceled(cpu);
1092		mutex_unlock(&slab_mutex);
1093		break;
1094	}
1095	return notifier_from_errno(err);
1096}
1097
1098static struct notifier_block cpucache_notifier = {
1099	&cpuup_callback, NULL, 0
1100};
 
 
 
 
 
 
 
 
 
 
1101
1102#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1103/*
1104 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1105 * Returns -EBUSY if all objects cannot be drained so that the node is not
1106 * removed.
1107 *
1108 * Must hold slab_mutex.
1109 */
1110static int __meminit drain_cache_node_node(int node)
1111{
1112	struct kmem_cache *cachep;
1113	int ret = 0;
1114
1115	list_for_each_entry(cachep, &slab_caches, list) {
1116		struct kmem_cache_node *n;
1117
1118		n = get_node(cachep, node);
1119		if (!n)
1120			continue;
1121
1122		drain_freelist(cachep, n, slabs_tofree(cachep, n));
1123
1124		if (!list_empty(&n->slabs_full) ||
1125		    !list_empty(&n->slabs_partial)) {
1126			ret = -EBUSY;
1127			break;
1128		}
1129	}
1130	return ret;
1131}
1132
1133static int __meminit slab_memory_callback(struct notifier_block *self,
1134					unsigned long action, void *arg)
1135{
1136	struct memory_notify *mnb = arg;
1137	int ret = 0;
1138	int nid;
1139
1140	nid = mnb->status_change_nid;
1141	if (nid < 0)
1142		goto out;
1143
1144	switch (action) {
1145	case MEM_GOING_ONLINE:
1146		mutex_lock(&slab_mutex);
1147		ret = init_cache_node_node(nid);
1148		mutex_unlock(&slab_mutex);
1149		break;
1150	case MEM_GOING_OFFLINE:
1151		mutex_lock(&slab_mutex);
1152		ret = drain_cache_node_node(nid);
1153		mutex_unlock(&slab_mutex);
1154		break;
1155	case MEM_ONLINE:
1156	case MEM_OFFLINE:
1157	case MEM_CANCEL_ONLINE:
1158	case MEM_CANCEL_OFFLINE:
1159		break;
1160	}
1161out:
1162	return notifier_from_errno(ret);
1163}
1164#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1165
1166/*
1167 * swap the static kmem_cache_node with kmalloced memory
1168 */
1169static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1170				int nodeid)
1171{
1172	struct kmem_cache_node *ptr;
1173
1174	ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1175	BUG_ON(!ptr);
1176
1177	memcpy(ptr, list, sizeof(struct kmem_cache_node));
1178	/*
1179	 * Do not assume that spinlocks can be initialized via memcpy:
1180	 */
1181	spin_lock_init(&ptr->list_lock);
1182
1183	MAKE_ALL_LISTS(cachep, ptr, nodeid);
1184	cachep->node[nodeid] = ptr;
1185}
1186
1187/*
1188 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1189 * size of kmem_cache_node.
1190 */
1191static void __init set_up_node(struct kmem_cache *cachep, int index)
1192{
1193	int node;
1194
1195	for_each_online_node(node) {
1196		cachep->node[node] = &init_kmem_cache_node[index + node];
1197		cachep->node[node]->next_reap = jiffies +
1198		    REAPTIMEOUT_NODE +
1199		    ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1200	}
1201}
1202
1203/*
1204 * Initialisation.  Called after the page allocator have been initialised and
1205 * before smp_init().
1206 */
1207void __init kmem_cache_init(void)
1208{
1209	int i;
1210
1211	BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1212					sizeof(struct rcu_head));
1213	kmem_cache = &kmem_cache_boot;
1214
1215	if (num_possible_nodes() == 1)
1216		use_alien_caches = 0;
1217
1218	for (i = 0; i < NUM_INIT_LISTS; i++)
1219		kmem_cache_node_init(&init_kmem_cache_node[i]);
1220
1221	/*
1222	 * Fragmentation resistance on low memory - only use bigger
1223	 * page orders on machines with more than 32MB of memory if
1224	 * not overridden on the command line.
1225	 */
1226	if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1227		slab_max_order = SLAB_MAX_ORDER_HI;
1228
1229	/* Bootstrap is tricky, because several objects are allocated
1230	 * from caches that do not exist yet:
1231	 * 1) initialize the kmem_cache cache: it contains the struct
1232	 *    kmem_cache structures of all caches, except kmem_cache itself:
1233	 *    kmem_cache is statically allocated.
1234	 *    Initially an __init data area is used for the head array and the
1235	 *    kmem_cache_node structures, it's replaced with a kmalloc allocated
1236	 *    array at the end of the bootstrap.
1237	 * 2) Create the first kmalloc cache.
1238	 *    The struct kmem_cache for the new cache is allocated normally.
1239	 *    An __init data area is used for the head array.
1240	 * 3) Create the remaining kmalloc caches, with minimally sized
1241	 *    head arrays.
1242	 * 4) Replace the __init data head arrays for kmem_cache and the first
1243	 *    kmalloc cache with kmalloc allocated arrays.
1244	 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1245	 *    the other cache's with kmalloc allocated memory.
1246	 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1247	 */
1248
1249	/* 1) create the kmem_cache */
1250
1251	/*
1252	 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1253	 */
1254	create_boot_cache(kmem_cache, "kmem_cache",
1255		offsetof(struct kmem_cache, node) +
1256				  nr_node_ids * sizeof(struct kmem_cache_node *),
1257				  SLAB_HWCACHE_ALIGN);
1258	list_add(&kmem_cache->list, &slab_caches);
 
1259	slab_state = PARTIAL;
1260
1261	/*
1262	 * Initialize the caches that provide memory for the  kmem_cache_node
1263	 * structures first.  Without this, further allocations will bug.
1264	 */
1265	kmalloc_caches[INDEX_NODE] = create_kmalloc_cache("kmalloc-node",
1266				kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
 
 
1267	slab_state = PARTIAL_NODE;
1268	setup_kmalloc_cache_index_table();
1269
1270	slab_early_init = 0;
1271
1272	/* 5) Replace the bootstrap kmem_cache_node */
1273	{
1274		int nid;
1275
1276		for_each_online_node(nid) {
1277			init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
1278
1279			init_list(kmalloc_caches[INDEX_NODE],
1280					  &init_kmem_cache_node[SIZE_NODE + nid], nid);
1281		}
1282	}
1283
1284	create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
1285}
1286
1287void __init kmem_cache_init_late(void)
1288{
1289	struct kmem_cache *cachep;
1290
1291	slab_state = UP;
1292
1293	/* 6) resize the head arrays to their final sizes */
1294	mutex_lock(&slab_mutex);
1295	list_for_each_entry(cachep, &slab_caches, list)
1296		if (enable_cpucache(cachep, GFP_NOWAIT))
1297			BUG();
1298	mutex_unlock(&slab_mutex);
1299
1300	/* Done! */
1301	slab_state = FULL;
1302
1303	/*
1304	 * Register a cpu startup notifier callback that initializes
1305	 * cpu_cache_get for all new cpus
1306	 */
1307	register_cpu_notifier(&cpucache_notifier);
1308
1309#ifdef CONFIG_NUMA
1310	/*
1311	 * Register a memory hotplug callback that initializes and frees
1312	 * node.
1313	 */
1314	hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1315#endif
1316
1317	/*
1318	 * The reap timers are started later, with a module init call: That part
1319	 * of the kernel is not yet operational.
1320	 */
1321}
1322
1323static int __init cpucache_init(void)
1324{
1325	int cpu;
1326
1327	/*
1328	 * Register the timers that return unneeded pages to the page allocator
1329	 */
1330	for_each_online_cpu(cpu)
1331		start_cpu_timer(cpu);
 
1332
1333	/* Done! */
1334	slab_state = FULL;
1335	return 0;
1336}
1337__initcall(cpucache_init);
1338
1339static noinline void
1340slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1341{
1342#if DEBUG
1343	struct kmem_cache_node *n;
1344	struct page *page;
1345	unsigned long flags;
1346	int node;
1347	static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1348				      DEFAULT_RATELIMIT_BURST);
1349
1350	if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1351		return;
1352
1353	pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1354		nodeid, gfpflags, &gfpflags);
1355	pr_warn("  cache: %s, object size: %d, order: %d\n",
1356		cachep->name, cachep->size, cachep->gfporder);
1357
1358	for_each_kmem_cache_node(cachep, node, n) {
1359		unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1360		unsigned long active_slabs = 0, num_slabs = 0;
1361
1362		spin_lock_irqsave(&n->list_lock, flags);
1363		list_for_each_entry(page, &n->slabs_full, lru) {
1364			active_objs += cachep->num;
1365			active_slabs++;
1366		}
1367		list_for_each_entry(page, &n->slabs_partial, lru) {
1368			active_objs += page->active;
1369			active_slabs++;
1370		}
1371		list_for_each_entry(page, &n->slabs_free, lru)
1372			num_slabs++;
1373
1374		free_objects += n->free_objects;
1375		spin_unlock_irqrestore(&n->list_lock, flags);
1376
1377		num_slabs += active_slabs;
1378		num_objs = num_slabs * cachep->num;
1379		pr_warn("  node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1380			node, active_slabs, num_slabs, active_objs, num_objs,
1381			free_objects);
1382	}
1383#endif
1384}
1385
1386/*
1387 * Interface to system's page allocator. No need to hold the
1388 * kmem_cache_node ->list_lock.
1389 *
1390 * If we requested dmaable memory, we will get it. Even if we
1391 * did not request dmaable memory, we might get it, but that
1392 * would be relatively rare and ignorable.
1393 */
1394static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1395								int nodeid)
1396{
1397	struct page *page;
1398	int nr_pages;
1399
1400	flags |= cachep->allocflags;
1401	if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1402		flags |= __GFP_RECLAIMABLE;
1403
1404	page = __alloc_pages_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1405	if (!page) {
1406		slab_out_of_memory(cachep, flags, nodeid);
1407		return NULL;
1408	}
1409
1410	if (memcg_charge_slab(page, flags, cachep->gfporder, cachep)) {
1411		__free_pages(page, cachep->gfporder);
1412		return NULL;
1413	}
1414
1415	nr_pages = (1 << cachep->gfporder);
1416	if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1417		add_zone_page_state(page_zone(page),
1418			NR_SLAB_RECLAIMABLE, nr_pages);
1419	else
1420		add_zone_page_state(page_zone(page),
1421			NR_SLAB_UNRECLAIMABLE, nr_pages);
1422
1423	__SetPageSlab(page);
1424	/* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1425	if (sk_memalloc_socks() && page_is_pfmemalloc(page))
1426		SetPageSlabPfmemalloc(page);
1427
1428	if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1429		kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1430
1431		if (cachep->ctor)
1432			kmemcheck_mark_uninitialized_pages(page, nr_pages);
1433		else
1434			kmemcheck_mark_unallocated_pages(page, nr_pages);
1435	}
1436
1437	return page;
1438}
1439
1440/*
1441 * Interface to system's page release.
1442 */
1443static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
1444{
1445	int order = cachep->gfporder;
1446	unsigned long nr_freed = (1 << order);
1447
1448	kmemcheck_free_shadow(page, order);
1449
1450	if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1451		sub_zone_page_state(page_zone(page),
1452				NR_SLAB_RECLAIMABLE, nr_freed);
1453	else
1454		sub_zone_page_state(page_zone(page),
1455				NR_SLAB_UNRECLAIMABLE, nr_freed);
1456
1457	BUG_ON(!PageSlab(page));
1458	__ClearPageSlabPfmemalloc(page);
1459	__ClearPageSlab(page);
1460	page_mapcount_reset(page);
1461	page->mapping = NULL;
1462
1463	if (current->reclaim_state)
1464		current->reclaim_state->reclaimed_slab += nr_freed;
1465	memcg_uncharge_slab(page, order, cachep);
1466	__free_pages(page, order);
1467}
1468
1469static void kmem_rcu_free(struct rcu_head *head)
1470{
1471	struct kmem_cache *cachep;
1472	struct page *page;
1473
1474	page = container_of(head, struct page, rcu_head);
1475	cachep = page->slab_cache;
1476
1477	kmem_freepages(cachep, page);
1478}
1479
1480#if DEBUG
1481static bool is_debug_pagealloc_cache(struct kmem_cache *cachep)
1482{
1483	if (debug_pagealloc_enabled() && OFF_SLAB(cachep) &&
1484		(cachep->size % PAGE_SIZE) == 0)
1485		return true;
1486
1487	return false;
1488}
1489
1490#ifdef CONFIG_DEBUG_PAGEALLOC
1491static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1492			    unsigned long caller)
1493{
1494	int size = cachep->object_size;
1495
1496	addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1497
1498	if (size < 5 * sizeof(unsigned long))
1499		return;
1500
1501	*addr++ = 0x12345678;
1502	*addr++ = caller;
1503	*addr++ = smp_processor_id();
1504	size -= 3 * sizeof(unsigned long);
1505	{
1506		unsigned long *sptr = &caller;
1507		unsigned long svalue;
1508
1509		while (!kstack_end(sptr)) {
1510			svalue = *sptr++;
1511			if (kernel_text_address(svalue)) {
1512				*addr++ = svalue;
1513				size -= sizeof(unsigned long);
1514				if (size <= sizeof(unsigned long))
1515					break;
1516			}
1517		}
1518
1519	}
1520	*addr++ = 0x87654321;
1521}
1522
1523static void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1524				int map, unsigned long caller)
1525{
1526	if (!is_debug_pagealloc_cache(cachep))
1527		return;
1528
1529	if (caller)
1530		store_stackinfo(cachep, objp, caller);
1531
1532	kernel_map_pages(virt_to_page(objp), cachep->size / PAGE_SIZE, map);
1533}
1534
1535#else
1536static inline void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1537				int map, unsigned long caller) {}
1538
1539#endif
1540
1541static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1542{
1543	int size = cachep->object_size;
1544	addr = &((char *)addr)[obj_offset(cachep)];
1545
1546	memset(addr, val, size);
1547	*(unsigned char *)(addr + size - 1) = POISON_END;
1548}
1549
1550static void dump_line(char *data, int offset, int limit)
1551{
1552	int i;
1553	unsigned char error = 0;
1554	int bad_count = 0;
1555
1556	pr_err("%03x: ", offset);
1557	for (i = 0; i < limit; i++) {
1558		if (data[offset + i] != POISON_FREE) {
1559			error = data[offset + i];
1560			bad_count++;
1561		}
1562	}
1563	print_hex_dump(KERN_CONT, "", 0, 16, 1,
1564			&data[offset], limit, 1);
1565
1566	if (bad_count == 1) {
1567		error ^= POISON_FREE;
1568		if (!(error & (error - 1))) {
1569			pr_err("Single bit error detected. Probably bad RAM.\n");
1570#ifdef CONFIG_X86
1571			pr_err("Run memtest86+ or a similar memory test tool.\n");
1572#else
1573			pr_err("Run a memory test tool.\n");
1574#endif
1575		}
1576	}
1577}
1578#endif
1579
1580#if DEBUG
1581
1582static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1583{
1584	int i, size;
1585	char *realobj;
1586
1587	if (cachep->flags & SLAB_RED_ZONE) {
1588		pr_err("Redzone: 0x%llx/0x%llx\n",
1589		       *dbg_redzone1(cachep, objp),
1590		       *dbg_redzone2(cachep, objp));
1591	}
1592
1593	if (cachep->flags & SLAB_STORE_USER) {
1594		pr_err("Last user: [<%p>](%pSR)\n",
1595		       *dbg_userword(cachep, objp),
1596		       *dbg_userword(cachep, objp));
1597	}
1598	realobj = (char *)objp + obj_offset(cachep);
1599	size = cachep->object_size;
1600	for (i = 0; i < size && lines; i += 16, lines--) {
1601		int limit;
1602		limit = 16;
1603		if (i + limit > size)
1604			limit = size - i;
1605		dump_line(realobj, i, limit);
1606	}
1607}
1608
1609static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1610{
1611	char *realobj;
1612	int size, i;
1613	int lines = 0;
1614
1615	if (is_debug_pagealloc_cache(cachep))
1616		return;
1617
1618	realobj = (char *)objp + obj_offset(cachep);
1619	size = cachep->object_size;
1620
1621	for (i = 0; i < size; i++) {
1622		char exp = POISON_FREE;
1623		if (i == size - 1)
1624			exp = POISON_END;
1625		if (realobj[i] != exp) {
1626			int limit;
1627			/* Mismatch ! */
1628			/* Print header */
1629			if (lines == 0) {
1630				pr_err("Slab corruption (%s): %s start=%p, len=%d\n",
1631				       print_tainted(), cachep->name,
1632				       realobj, size);
1633				print_objinfo(cachep, objp, 0);
1634			}
1635			/* Hexdump the affected line */
1636			i = (i / 16) * 16;
1637			limit = 16;
1638			if (i + limit > size)
1639				limit = size - i;
1640			dump_line(realobj, i, limit);
1641			i += 16;
1642			lines++;
1643			/* Limit to 5 lines */
1644			if (lines > 5)
1645				break;
1646		}
1647	}
1648	if (lines != 0) {
1649		/* Print some data about the neighboring objects, if they
1650		 * exist:
1651		 */
1652		struct page *page = virt_to_head_page(objp);
1653		unsigned int objnr;
1654
1655		objnr = obj_to_index(cachep, page, objp);
1656		if (objnr) {
1657			objp = index_to_obj(cachep, page, objnr - 1);
1658			realobj = (char *)objp + obj_offset(cachep);
1659			pr_err("Prev obj: start=%p, len=%d\n", realobj, size);
1660			print_objinfo(cachep, objp, 2);
1661		}
1662		if (objnr + 1 < cachep->num) {
1663			objp = index_to_obj(cachep, page, objnr + 1);
1664			realobj = (char *)objp + obj_offset(cachep);
1665			pr_err("Next obj: start=%p, len=%d\n", realobj, size);
1666			print_objinfo(cachep, objp, 2);
1667		}
1668	}
1669}
1670#endif
1671
1672#if DEBUG
1673static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1674						struct page *page)
1675{
1676	int i;
1677
1678	if (OBJFREELIST_SLAB(cachep) && cachep->flags & SLAB_POISON) {
1679		poison_obj(cachep, page->freelist - obj_offset(cachep),
1680			POISON_FREE);
1681	}
1682
1683	for (i = 0; i < cachep->num; i++) {
1684		void *objp = index_to_obj(cachep, page, i);
1685
1686		if (cachep->flags & SLAB_POISON) {
1687			check_poison_obj(cachep, objp);
1688			slab_kernel_map(cachep, objp, 1, 0);
1689		}
1690		if (cachep->flags & SLAB_RED_ZONE) {
1691			if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1692				slab_error(cachep, "start of a freed object was overwritten");
1693			if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1694				slab_error(cachep, "end of a freed object was overwritten");
1695		}
1696	}
1697}
1698#else
1699static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1700						struct page *page)
1701{
1702}
1703#endif
1704
1705/**
1706 * slab_destroy - destroy and release all objects in a slab
1707 * @cachep: cache pointer being destroyed
1708 * @page: page pointer being destroyed
1709 *
1710 * Destroy all the objs in a slab page, and release the mem back to the system.
1711 * Before calling the slab page must have been unlinked from the cache. The
1712 * kmem_cache_node ->list_lock is not held/needed.
1713 */
1714static void slab_destroy(struct kmem_cache *cachep, struct page *page)
1715{
1716	void *freelist;
1717
1718	freelist = page->freelist;
1719	slab_destroy_debugcheck(cachep, page);
1720	if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
1721		call_rcu(&page->rcu_head, kmem_rcu_free);
1722	else
1723		kmem_freepages(cachep, page);
1724
1725	/*
1726	 * From now on, we don't use freelist
1727	 * although actual page can be freed in rcu context
1728	 */
1729	if (OFF_SLAB(cachep))
1730		kmem_cache_free(cachep->freelist_cache, freelist);
1731}
1732
1733static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
1734{
1735	struct page *page, *n;
1736
1737	list_for_each_entry_safe(page, n, list, lru) {
1738		list_del(&page->lru);
1739		slab_destroy(cachep, page);
1740	}
1741}
1742
1743/**
1744 * calculate_slab_order - calculate size (page order) of slabs
1745 * @cachep: pointer to the cache that is being created
1746 * @size: size of objects to be created in this cache.
1747 * @flags: slab allocation flags
1748 *
1749 * Also calculates the number of objects per slab.
1750 *
1751 * This could be made much more intelligent.  For now, try to avoid using
1752 * high order pages for slabs.  When the gfp() functions are more friendly
1753 * towards high-order requests, this should be changed.
1754 */
1755static size_t calculate_slab_order(struct kmem_cache *cachep,
1756				size_t size, unsigned long flags)
1757{
1758	size_t left_over = 0;
1759	int gfporder;
1760
1761	for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
1762		unsigned int num;
1763		size_t remainder;
1764
1765		num = cache_estimate(gfporder, size, flags, &remainder);
1766		if (!num)
1767			continue;
1768
1769		/* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1770		if (num > SLAB_OBJ_MAX_NUM)
1771			break;
1772
1773		if (flags & CFLGS_OFF_SLAB) {
1774			struct kmem_cache *freelist_cache;
1775			size_t freelist_size;
1776
1777			freelist_size = num * sizeof(freelist_idx_t);
1778			freelist_cache = kmalloc_slab(freelist_size, 0u);
1779			if (!freelist_cache)
1780				continue;
1781
1782			/*
1783			 * Needed to avoid possible looping condition
1784			 * in cache_grow()
1785			 */
1786			if (OFF_SLAB(freelist_cache))
1787				continue;
1788
1789			/* check if off slab has enough benefit */
1790			if (freelist_cache->size > cachep->size / 2)
1791				continue;
1792		}
1793
1794		/* Found something acceptable - save it away */
1795		cachep->num = num;
1796		cachep->gfporder = gfporder;
1797		left_over = remainder;
1798
1799		/*
1800		 * A VFS-reclaimable slab tends to have most allocations
1801		 * as GFP_NOFS and we really don't want to have to be allocating
1802		 * higher-order pages when we are unable to shrink dcache.
1803		 */
1804		if (flags & SLAB_RECLAIM_ACCOUNT)
1805			break;
1806
1807		/*
1808		 * Large number of objects is good, but very large slabs are
1809		 * currently bad for the gfp()s.
1810		 */
1811		if (gfporder >= slab_max_order)
1812			break;
1813
1814		/*
1815		 * Acceptable internal fragmentation?
1816		 */
1817		if (left_over * 8 <= (PAGE_SIZE << gfporder))
1818			break;
1819	}
1820	return left_over;
1821}
1822
1823static struct array_cache __percpu *alloc_kmem_cache_cpus(
1824		struct kmem_cache *cachep, int entries, int batchcount)
1825{
1826	int cpu;
1827	size_t size;
1828	struct array_cache __percpu *cpu_cache;
1829
1830	size = sizeof(void *) * entries + sizeof(struct array_cache);
1831	cpu_cache = __alloc_percpu(size, sizeof(void *));
1832
1833	if (!cpu_cache)
1834		return NULL;
1835
1836	for_each_possible_cpu(cpu) {
1837		init_arraycache(per_cpu_ptr(cpu_cache, cpu),
1838				entries, batchcount);
1839	}
1840
1841	return cpu_cache;
1842}
1843
1844static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
1845{
1846	if (slab_state >= FULL)
1847		return enable_cpucache(cachep, gfp);
1848
1849	cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1);
1850	if (!cachep->cpu_cache)
1851		return 1;
1852
1853	if (slab_state == DOWN) {
1854		/* Creation of first cache (kmem_cache). */
1855		set_up_node(kmem_cache, CACHE_CACHE);
1856	} else if (slab_state == PARTIAL) {
1857		/* For kmem_cache_node */
1858		set_up_node(cachep, SIZE_NODE);
1859	} else {
1860		int node;
1861
1862		for_each_online_node(node) {
1863			cachep->node[node] = kmalloc_node(
1864				sizeof(struct kmem_cache_node), gfp, node);
1865			BUG_ON(!cachep->node[node]);
1866			kmem_cache_node_init(cachep->node[node]);
1867		}
1868	}
1869
1870	cachep->node[numa_mem_id()]->next_reap =
1871			jiffies + REAPTIMEOUT_NODE +
1872			((unsigned long)cachep) % REAPTIMEOUT_NODE;
1873
1874	cpu_cache_get(cachep)->avail = 0;
1875	cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1876	cpu_cache_get(cachep)->batchcount = 1;
1877	cpu_cache_get(cachep)->touched = 0;
1878	cachep->batchcount = 1;
1879	cachep->limit = BOOT_CPUCACHE_ENTRIES;
1880	return 0;
1881}
1882
1883unsigned long kmem_cache_flags(unsigned long object_size,
1884	unsigned long flags, const char *name,
1885	void (*ctor)(void *))
1886{
1887	return flags;
1888}
1889
1890struct kmem_cache *
1891__kmem_cache_alias(const char *name, size_t size, size_t align,
1892		   unsigned long flags, void (*ctor)(void *))
1893{
1894	struct kmem_cache *cachep;
1895
1896	cachep = find_mergeable(size, align, flags, name, ctor);
1897	if (cachep) {
1898		cachep->refcount++;
1899
1900		/*
1901		 * Adjust the object sizes so that we clear
1902		 * the complete object on kzalloc.
1903		 */
1904		cachep->object_size = max_t(int, cachep->object_size, size);
1905	}
1906	return cachep;
1907}
1908
1909static bool set_objfreelist_slab_cache(struct kmem_cache *cachep,
1910			size_t size, unsigned long flags)
1911{
1912	size_t left;
1913
1914	cachep->num = 0;
1915
1916	if (cachep->ctor || flags & SLAB_DESTROY_BY_RCU)
1917		return false;
1918
1919	left = calculate_slab_order(cachep, size,
1920			flags | CFLGS_OBJFREELIST_SLAB);
1921	if (!cachep->num)
1922		return false;
1923
1924	if (cachep->num * sizeof(freelist_idx_t) > cachep->object_size)
1925		return false;
1926
1927	cachep->colour = left / cachep->colour_off;
1928
1929	return true;
1930}
1931
1932static bool set_off_slab_cache(struct kmem_cache *cachep,
1933			size_t size, unsigned long flags)
1934{
1935	size_t left;
1936
1937	cachep->num = 0;
1938
1939	/*
1940	 * Always use on-slab management when SLAB_NOLEAKTRACE
1941	 * to avoid recursive calls into kmemleak.
1942	 */
1943	if (flags & SLAB_NOLEAKTRACE)
1944		return false;
1945
1946	/*
1947	 * Size is large, assume best to place the slab management obj
1948	 * off-slab (should allow better packing of objs).
1949	 */
1950	left = calculate_slab_order(cachep, size, flags | CFLGS_OFF_SLAB);
1951	if (!cachep->num)
1952		return false;
1953
1954	/*
1955	 * If the slab has been placed off-slab, and we have enough space then
1956	 * move it on-slab. This is at the expense of any extra colouring.
1957	 */
1958	if (left >= cachep->num * sizeof(freelist_idx_t))
1959		return false;
1960
1961	cachep->colour = left / cachep->colour_off;
1962
1963	return true;
1964}
1965
1966static bool set_on_slab_cache(struct kmem_cache *cachep,
1967			size_t size, unsigned long flags)
1968{
1969	size_t left;
1970
1971	cachep->num = 0;
1972
1973	left = calculate_slab_order(cachep, size, flags);
1974	if (!cachep->num)
1975		return false;
1976
1977	cachep->colour = left / cachep->colour_off;
1978
1979	return true;
1980}
1981
1982/**
1983 * __kmem_cache_create - Create a cache.
1984 * @cachep: cache management descriptor
1985 * @flags: SLAB flags
1986 *
1987 * Returns a ptr to the cache on success, NULL on failure.
1988 * Cannot be called within a int, but can be interrupted.
1989 * The @ctor is run when new pages are allocated by the cache.
1990 *
1991 * The flags are
1992 *
1993 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1994 * to catch references to uninitialised memory.
1995 *
1996 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1997 * for buffer overruns.
1998 *
1999 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2000 * cacheline.  This can be beneficial if you're counting cycles as closely
2001 * as davem.
2002 */
2003int
2004__kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
2005{
2006	size_t ralign = BYTES_PER_WORD;
2007	gfp_t gfp;
2008	int err;
2009	size_t size = cachep->size;
2010
2011#if DEBUG
2012#if FORCED_DEBUG
2013	/*
2014	 * Enable redzoning and last user accounting, except for caches with
2015	 * large objects, if the increased size would increase the object size
2016	 * above the next power of two: caches with object sizes just above a
2017	 * power of two have a significant amount of internal fragmentation.
2018	 */
2019	if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2020						2 * sizeof(unsigned long long)))
2021		flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2022	if (!(flags & SLAB_DESTROY_BY_RCU))
2023		flags |= SLAB_POISON;
2024#endif
2025#endif
2026
2027	/*
2028	 * Check that size is in terms of words.  This is needed to avoid
2029	 * unaligned accesses for some archs when redzoning is used, and makes
2030	 * sure any on-slab bufctl's are also correctly aligned.
2031	 */
2032	if (size & (BYTES_PER_WORD - 1)) {
2033		size += (BYTES_PER_WORD - 1);
2034		size &= ~(BYTES_PER_WORD - 1);
2035	}
2036
2037	if (flags & SLAB_RED_ZONE) {
2038		ralign = REDZONE_ALIGN;
2039		/* If redzoning, ensure that the second redzone is suitably
2040		 * aligned, by adjusting the object size accordingly. */
2041		size += REDZONE_ALIGN - 1;
2042		size &= ~(REDZONE_ALIGN - 1);
2043	}
2044
2045	/* 3) caller mandated alignment */
2046	if (ralign < cachep->align) {
2047		ralign = cachep->align;
2048	}
2049	/* disable debug if necessary */
2050	if (ralign > __alignof__(unsigned long long))
2051		flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2052	/*
2053	 * 4) Store it.
2054	 */
2055	cachep->align = ralign;
2056	cachep->colour_off = cache_line_size();
2057	/* Offset must be a multiple of the alignment. */
2058	if (cachep->colour_off < cachep->align)
2059		cachep->colour_off = cachep->align;
2060
2061	if (slab_is_available())
2062		gfp = GFP_KERNEL;
2063	else
2064		gfp = GFP_NOWAIT;
2065
2066#if DEBUG
2067
2068	/*
2069	 * Both debugging options require word-alignment which is calculated
2070	 * into align above.
2071	 */
2072	if (flags & SLAB_RED_ZONE) {
2073		/* add space for red zone words */
2074		cachep->obj_offset += sizeof(unsigned long long);
2075		size += 2 * sizeof(unsigned long long);
2076	}
2077	if (flags & SLAB_STORE_USER) {
2078		/* user store requires one word storage behind the end of
2079		 * the real object. But if the second red zone needs to be
2080		 * aligned to 64 bits, we must allow that much space.
2081		 */
2082		if (flags & SLAB_RED_ZONE)
2083			size += REDZONE_ALIGN;
2084		else
2085			size += BYTES_PER_WORD;
2086	}
2087#endif
2088
2089	kasan_cache_create(cachep, &size, &flags);
2090
2091	size = ALIGN(size, cachep->align);
2092	/*
2093	 * We should restrict the number of objects in a slab to implement
2094	 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2095	 */
2096	if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2097		size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
2098
2099#if DEBUG
2100	/*
2101	 * To activate debug pagealloc, off-slab management is necessary
2102	 * requirement. In early phase of initialization, small sized slab
2103	 * doesn't get initialized so it would not be possible. So, we need
2104	 * to check size >= 256. It guarantees that all necessary small
2105	 * sized slab is initialized in current slab initialization sequence.
2106	 */
2107	if (debug_pagealloc_enabled() && (flags & SLAB_POISON) &&
2108		size >= 256 && cachep->object_size > cache_line_size()) {
2109		if (size < PAGE_SIZE || size % PAGE_SIZE == 0) {
2110			size_t tmp_size = ALIGN(size, PAGE_SIZE);
2111
2112			if (set_off_slab_cache(cachep, tmp_size, flags)) {
2113				flags |= CFLGS_OFF_SLAB;
2114				cachep->obj_offset += tmp_size - size;
2115				size = tmp_size;
2116				goto done;
2117			}
2118		}
2119	}
2120#endif
2121
2122	if (set_objfreelist_slab_cache(cachep, size, flags)) {
2123		flags |= CFLGS_OBJFREELIST_SLAB;
2124		goto done;
2125	}
2126
2127	if (set_off_slab_cache(cachep, size, flags)) {
2128		flags |= CFLGS_OFF_SLAB;
2129		goto done;
2130	}
2131
2132	if (set_on_slab_cache(cachep, size, flags))
2133		goto done;
2134
2135	return -E2BIG;
2136
2137done:
2138	cachep->freelist_size = cachep->num * sizeof(freelist_idx_t);
2139	cachep->flags = flags;
2140	cachep->allocflags = __GFP_COMP;
2141	if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2142		cachep->allocflags |= GFP_DMA;
 
 
2143	cachep->size = size;
2144	cachep->reciprocal_buffer_size = reciprocal_value(size);
2145
2146#if DEBUG
2147	/*
2148	 * If we're going to use the generic kernel_map_pages()
2149	 * poisoning, then it's going to smash the contents of
2150	 * the redzone and userword anyhow, so switch them off.
2151	 */
2152	if (IS_ENABLED(CONFIG_PAGE_POISONING) &&
2153		(cachep->flags & SLAB_POISON) &&
2154		is_debug_pagealloc_cache(cachep))
2155		cachep->flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2156#endif
2157
2158	if (OFF_SLAB(cachep)) {
2159		cachep->freelist_cache =
2160			kmalloc_slab(cachep->freelist_size, 0u);
2161	}
2162
2163	err = setup_cpu_cache(cachep, gfp);
2164	if (err) {
2165		__kmem_cache_release(cachep);
2166		return err;
2167	}
2168
2169	return 0;
2170}
2171
2172#if DEBUG
2173static void check_irq_off(void)
2174{
2175	BUG_ON(!irqs_disabled());
2176}
2177
2178static void check_irq_on(void)
2179{
2180	BUG_ON(irqs_disabled());
2181}
2182
 
 
 
 
 
2183static void check_spinlock_acquired(struct kmem_cache *cachep)
2184{
2185#ifdef CONFIG_SMP
2186	check_irq_off();
2187	assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
2188#endif
2189}
2190
2191static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2192{
2193#ifdef CONFIG_SMP
2194	check_irq_off();
2195	assert_spin_locked(&get_node(cachep, node)->list_lock);
2196#endif
2197}
2198
2199#else
2200#define check_irq_off()	do { } while(0)
2201#define check_irq_on()	do { } while(0)
 
2202#define check_spinlock_acquired(x) do { } while(0)
2203#define check_spinlock_acquired_node(x, y) do { } while(0)
2204#endif
2205
2206static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
2207			struct array_cache *ac,
2208			int force, int node);
 
 
 
 
 
 
 
 
 
 
 
 
 
2209
2210static void do_drain(void *arg)
2211{
2212	struct kmem_cache *cachep = arg;
2213	struct array_cache *ac;
2214	int node = numa_mem_id();
2215	struct kmem_cache_node *n;
2216	LIST_HEAD(list);
2217
2218	check_irq_off();
2219	ac = cpu_cache_get(cachep);
2220	n = get_node(cachep, node);
2221	spin_lock(&n->list_lock);
2222	free_block(cachep, ac->entry, ac->avail, node, &list);
2223	spin_unlock(&n->list_lock);
2224	slabs_destroy(cachep, &list);
2225	ac->avail = 0;
2226}
2227
2228static void drain_cpu_caches(struct kmem_cache *cachep)
2229{
2230	struct kmem_cache_node *n;
2231	int node;
 
2232
2233	on_each_cpu(do_drain, cachep, 1);
2234	check_irq_on();
2235	for_each_kmem_cache_node(cachep, node, n)
2236		if (n->alien)
2237			drain_alien_cache(cachep, n->alien);
2238
2239	for_each_kmem_cache_node(cachep, node, n)
2240		drain_array(cachep, n, n->shared, 1, node);
 
 
 
 
 
2241}
2242
2243/*
2244 * Remove slabs from the list of free slabs.
2245 * Specify the number of slabs to drain in tofree.
2246 *
2247 * Returns the actual number of slabs released.
2248 */
2249static int drain_freelist(struct kmem_cache *cache,
2250			struct kmem_cache_node *n, int tofree)
2251{
2252	struct list_head *p;
2253	int nr_freed;
2254	struct page *page;
2255
2256	nr_freed = 0;
2257	while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
2258
2259		spin_lock_irq(&n->list_lock);
2260		p = n->slabs_free.prev;
2261		if (p == &n->slabs_free) {
2262			spin_unlock_irq(&n->list_lock);
2263			goto out;
2264		}
2265
2266		page = list_entry(p, struct page, lru);
2267		list_del(&page->lru);
 
 
2268		/*
2269		 * Safe to drop the lock. The slab is no longer linked
2270		 * to the cache.
2271		 */
2272		n->free_objects -= cache->num;
2273		spin_unlock_irq(&n->list_lock);
2274		slab_destroy(cache, page);
2275		nr_freed++;
2276	}
2277out:
2278	return nr_freed;
2279}
2280
2281int __kmem_cache_shrink(struct kmem_cache *cachep, bool deactivate)
 
 
 
 
 
 
 
 
 
 
 
 
2282{
2283	int ret = 0;
2284	int node;
2285	struct kmem_cache_node *n;
2286
2287	drain_cpu_caches(cachep);
2288
2289	check_irq_on();
2290	for_each_kmem_cache_node(cachep, node, n) {
2291		drain_freelist(cachep, n, slabs_tofree(cachep, n));
2292
2293		ret += !list_empty(&n->slabs_full) ||
2294			!list_empty(&n->slabs_partial);
2295	}
2296	return (ret ? 1 : 0);
2297}
2298
 
 
 
 
 
 
 
2299int __kmem_cache_shutdown(struct kmem_cache *cachep)
2300{
2301	return __kmem_cache_shrink(cachep, false);
2302}
2303
2304void __kmem_cache_release(struct kmem_cache *cachep)
2305{
2306	int i;
2307	struct kmem_cache_node *n;
2308
 
 
2309	free_percpu(cachep->cpu_cache);
2310
2311	/* NUMA: free the node structures */
2312	for_each_kmem_cache_node(cachep, i, n) {
2313		kfree(n->shared);
2314		free_alien_cache(n->alien);
2315		kfree(n);
2316		cachep->node[i] = NULL;
2317	}
2318}
2319
2320/*
2321 * Get the memory for a slab management obj.
2322 *
2323 * For a slab cache when the slab descriptor is off-slab, the
2324 * slab descriptor can't come from the same cache which is being created,
2325 * Because if it is the case, that means we defer the creation of
2326 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2327 * And we eventually call down to __kmem_cache_create(), which
2328 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2329 * This is a "chicken-and-egg" problem.
2330 *
2331 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2332 * which are all initialized during kmem_cache_init().
2333 */
2334static void *alloc_slabmgmt(struct kmem_cache *cachep,
2335				   struct page *page, int colour_off,
2336				   gfp_t local_flags, int nodeid)
2337{
2338	void *freelist;
2339	void *addr = page_address(page);
2340
2341	page->s_mem = addr + colour_off;
2342	page->active = 0;
2343
2344	if (OBJFREELIST_SLAB(cachep))
2345		freelist = NULL;
2346	else if (OFF_SLAB(cachep)) {
2347		/* Slab management obj is off-slab. */
2348		freelist = kmem_cache_alloc_node(cachep->freelist_cache,
2349					      local_flags, nodeid);
2350		if (!freelist)
2351			return NULL;
2352	} else {
2353		/* We will use last bytes at the slab for freelist */
2354		freelist = addr + (PAGE_SIZE << cachep->gfporder) -
2355				cachep->freelist_size;
2356	}
2357
2358	return freelist;
2359}
2360
2361static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
2362{
2363	return ((freelist_idx_t *)page->freelist)[idx];
2364}
2365
2366static inline void set_free_obj(struct page *page,
2367					unsigned int idx, freelist_idx_t val)
2368{
2369	((freelist_idx_t *)(page->freelist))[idx] = val;
2370}
2371
2372static void cache_init_objs_debug(struct kmem_cache *cachep, struct page *page)
2373{
2374#if DEBUG
2375	int i;
2376
2377	for (i = 0; i < cachep->num; i++) {
2378		void *objp = index_to_obj(cachep, page, i);
2379
2380		if (cachep->flags & SLAB_STORE_USER)
2381			*dbg_userword(cachep, objp) = NULL;
2382
2383		if (cachep->flags & SLAB_RED_ZONE) {
2384			*dbg_redzone1(cachep, objp) = RED_INACTIVE;
2385			*dbg_redzone2(cachep, objp) = RED_INACTIVE;
2386		}
2387		/*
2388		 * Constructors are not allowed to allocate memory from the same
2389		 * cache which they are a constructor for.  Otherwise, deadlock.
2390		 * They must also be threaded.
2391		 */
2392		if (cachep->ctor && !(cachep->flags & SLAB_POISON)) {
2393			kasan_unpoison_object_data(cachep,
2394						   objp + obj_offset(cachep));
2395			cachep->ctor(objp + obj_offset(cachep));
2396			kasan_poison_object_data(
2397				cachep, objp + obj_offset(cachep));
2398		}
2399
2400		if (cachep->flags & SLAB_RED_ZONE) {
2401			if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2402				slab_error(cachep, "constructor overwrote the end of an object");
2403			if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2404				slab_error(cachep, "constructor overwrote the start of an object");
2405		}
2406		/* need to poison the objs? */
2407		if (cachep->flags & SLAB_POISON) {
2408			poison_obj(cachep, objp, POISON_FREE);
2409			slab_kernel_map(cachep, objp, 0, 0);
2410		}
2411	}
2412#endif
2413}
2414
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2415static void cache_init_objs(struct kmem_cache *cachep,
2416			    struct page *page)
2417{
2418	int i;
2419	void *objp;
 
2420
2421	cache_init_objs_debug(cachep, page);
2422
2423	if (OBJFREELIST_SLAB(cachep)) {
 
 
 
2424		page->freelist = index_to_obj(cachep, page, cachep->num - 1) +
2425						obj_offset(cachep);
2426	}
2427
2428	for (i = 0; i < cachep->num; i++) {
 
 
 
2429		/* constructor could break poison info */
2430		if (DEBUG == 0 && cachep->ctor) {
2431			objp = index_to_obj(cachep, page, i);
2432			kasan_unpoison_object_data(cachep, objp);
2433			cachep->ctor(objp);
2434			kasan_poison_object_data(cachep, objp);
2435		}
2436
2437		set_free_obj(page, i, i);
2438	}
2439}
2440
2441static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2442{
2443	if (CONFIG_ZONE_DMA_FLAG) {
2444		if (flags & GFP_DMA)
2445			BUG_ON(!(cachep->allocflags & GFP_DMA));
2446		else
2447			BUG_ON(cachep->allocflags & GFP_DMA);
2448	}
2449}
2450
2451static void *slab_get_obj(struct kmem_cache *cachep, struct page *page)
2452{
2453	void *objp;
2454
2455	objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
2456	page->active++;
2457
2458#if DEBUG
2459	if (cachep->flags & SLAB_STORE_USER)
2460		set_store_user_dirty(cachep);
2461#endif
2462
2463	return objp;
2464}
2465
2466static void slab_put_obj(struct kmem_cache *cachep,
2467			struct page *page, void *objp)
2468{
2469	unsigned int objnr = obj_to_index(cachep, page, objp);
2470#if DEBUG
2471	unsigned int i;
2472
2473	/* Verify double free bug */
2474	for (i = page->active; i < cachep->num; i++) {
2475		if (get_free_obj(page, i) == objnr) {
2476			pr_err("slab: double free detected in cache '%s', objp %p\n",
2477			       cachep->name, objp);
2478			BUG();
2479		}
2480	}
2481#endif
2482	page->active--;
2483	if (!page->freelist)
2484		page->freelist = objp + obj_offset(cachep);
2485
2486	set_free_obj(page, page->active, objnr);
2487}
2488
2489/*
2490 * Map pages beginning at addr to the given cache and slab. This is required
2491 * for the slab allocator to be able to lookup the cache and slab of a
2492 * virtual address for kfree, ksize, and slab debugging.
2493 */
2494static void slab_map_pages(struct kmem_cache *cache, struct page *page,
2495			   void *freelist)
2496{
2497	page->slab_cache = cache;
2498	page->freelist = freelist;
2499}
2500
2501/*
2502 * Grow (by 1) the number of slabs within a cache.  This is called by
2503 * kmem_cache_alloc() when there are no active objs left in a cache.
2504 */
2505static int cache_grow(struct kmem_cache *cachep,
2506		gfp_t flags, int nodeid, struct page *page)
2507{
2508	void *freelist;
2509	size_t offset;
2510	gfp_t local_flags;
 
2511	struct kmem_cache_node *n;
 
2512
2513	/*
2514	 * Be lazy and only check for valid flags here,  keeping it out of the
2515	 * critical path in kmem_cache_alloc().
2516	 */
2517	if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
2518		pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
2519		BUG();
 
 
 
2520	}
2521	local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2522
2523	/* Take the node list lock to change the colour_next on this node */
2524	check_irq_off();
2525	n = get_node(cachep, nodeid);
2526	spin_lock(&n->list_lock);
2527
2528	/* Get colour for the slab, and cal the next value. */
2529	offset = n->colour_next;
2530	n->colour_next++;
2531	if (n->colour_next >= cachep->colour)
2532		n->colour_next = 0;
2533	spin_unlock(&n->list_lock);
2534
2535	offset *= cachep->colour_off;
2536
2537	if (gfpflags_allow_blocking(local_flags))
2538		local_irq_enable();
2539
2540	/*
2541	 * The test for missing atomic flag is performed here, rather than
2542	 * the more obvious place, simply to reduce the critical path length
2543	 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2544	 * will eventually be caught here (where it matters).
2545	 */
2546	kmem_flagcheck(cachep, flags);
2547
2548	/*
2549	 * Get mem for the objs.  Attempt to allocate a physical page from
2550	 * 'nodeid'.
2551	 */
2552	if (!page)
2553		page = kmem_getpages(cachep, local_flags, nodeid);
2554	if (!page)
2555		goto failed;
2556
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2557	/* Get slab management. */
2558	freelist = alloc_slabmgmt(cachep, page, offset,
2559			local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2560	if (OFF_SLAB(cachep) && !freelist)
2561		goto opps1;
2562
2563	slab_map_pages(cachep, page, freelist);
2564
2565	kasan_poison_slab(page);
2566	cache_init_objs(cachep, page);
2567
2568	if (gfpflags_allow_blocking(local_flags))
2569		local_irq_disable();
2570	check_irq_off();
2571	spin_lock(&n->list_lock);
2572
2573	/* Make slab active. */
2574	list_add_tail(&page->lru, &(n->slabs_free));
2575	STATS_INC_GROWN(cachep);
2576	n->free_objects += cachep->num;
2577	spin_unlock(&n->list_lock);
2578	return 1;
2579opps1:
2580	kmem_freepages(cachep, page);
2581failed:
2582	if (gfpflags_allow_blocking(local_flags))
2583		local_irq_disable();
2584	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2585}
2586
2587#if DEBUG
2588
2589/*
2590 * Perform extra freeing checks:
2591 * - detect bad pointers.
2592 * - POISON/RED_ZONE checking
2593 */
2594static void kfree_debugcheck(const void *objp)
2595{
2596	if (!virt_addr_valid(objp)) {
2597		pr_err("kfree_debugcheck: out of range ptr %lxh\n",
2598		       (unsigned long)objp);
2599		BUG();
2600	}
2601}
2602
2603static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2604{
2605	unsigned long long redzone1, redzone2;
2606
2607	redzone1 = *dbg_redzone1(cache, obj);
2608	redzone2 = *dbg_redzone2(cache, obj);
2609
2610	/*
2611	 * Redzone is ok.
2612	 */
2613	if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2614		return;
2615
2616	if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2617		slab_error(cache, "double free detected");
2618	else
2619		slab_error(cache, "memory outside object was overwritten");
2620
2621	pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2622	       obj, redzone1, redzone2);
2623}
2624
2625static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2626				   unsigned long caller)
2627{
2628	unsigned int objnr;
2629	struct page *page;
2630
2631	BUG_ON(virt_to_cache(objp) != cachep);
2632
2633	objp -= obj_offset(cachep);
2634	kfree_debugcheck(objp);
2635	page = virt_to_head_page(objp);
2636
2637	if (cachep->flags & SLAB_RED_ZONE) {
2638		verify_redzone_free(cachep, objp);
2639		*dbg_redzone1(cachep, objp) = RED_INACTIVE;
2640		*dbg_redzone2(cachep, objp) = RED_INACTIVE;
2641	}
2642	if (cachep->flags & SLAB_STORE_USER) {
2643		set_store_user_dirty(cachep);
2644		*dbg_userword(cachep, objp) = (void *)caller;
2645	}
2646
2647	objnr = obj_to_index(cachep, page, objp);
2648
2649	BUG_ON(objnr >= cachep->num);
2650	BUG_ON(objp != index_to_obj(cachep, page, objnr));
2651
2652	if (cachep->flags & SLAB_POISON) {
2653		poison_obj(cachep, objp, POISON_FREE);
2654		slab_kernel_map(cachep, objp, 0, caller);
2655	}
2656	return objp;
2657}
2658
2659#else
2660#define kfree_debugcheck(x) do { } while(0)
2661#define cache_free_debugcheck(x,objp,z) (objp)
2662#endif
2663
2664static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
2665						void **list)
2666{
2667#if DEBUG
2668	void *next = *list;
2669	void *objp;
2670
2671	while (next) {
2672		objp = next - obj_offset(cachep);
2673		next = *(void **)next;
2674		poison_obj(cachep, objp, POISON_FREE);
2675	}
2676#endif
2677}
2678
2679static inline void fixup_slab_list(struct kmem_cache *cachep,
2680				struct kmem_cache_node *n, struct page *page,
2681				void **list)
2682{
2683	/* move slabp to correct slabp list: */
2684	list_del(&page->lru);
2685	if (page->active == cachep->num) {
2686		list_add(&page->lru, &n->slabs_full);
2687		if (OBJFREELIST_SLAB(cachep)) {
2688#if DEBUG
2689			/* Poisoning will be done without holding the lock */
2690			if (cachep->flags & SLAB_POISON) {
2691				void **objp = page->freelist;
2692
2693				*objp = *list;
2694				*list = objp;
2695			}
2696#endif
2697			page->freelist = NULL;
2698		}
2699	} else
2700		list_add(&page->lru, &n->slabs_partial);
2701}
2702
2703/* Try to find non-pfmemalloc slab if needed */
2704static noinline struct page *get_valid_first_slab(struct kmem_cache_node *n,
2705					struct page *page, bool pfmemalloc)
2706{
2707	if (!page)
2708		return NULL;
2709
2710	if (pfmemalloc)
2711		return page;
2712
2713	if (!PageSlabPfmemalloc(page))
2714		return page;
2715
2716	/* No need to keep pfmemalloc slab if we have enough free objects */
2717	if (n->free_objects > n->free_limit) {
2718		ClearPageSlabPfmemalloc(page);
2719		return page;
2720	}
2721
2722	/* Move pfmemalloc slab to the end of list to speed up next search */
2723	list_del(&page->lru);
2724	if (!page->active)
2725		list_add_tail(&page->lru, &n->slabs_free);
2726	else
 
2727		list_add_tail(&page->lru, &n->slabs_partial);
2728
2729	list_for_each_entry(page, &n->slabs_partial, lru) {
2730		if (!PageSlabPfmemalloc(page))
2731			return page;
2732	}
2733
 
2734	list_for_each_entry(page, &n->slabs_free, lru) {
2735		if (!PageSlabPfmemalloc(page))
 
2736			return page;
 
2737	}
2738
2739	return NULL;
2740}
2741
2742static struct page *get_first_slab(struct kmem_cache_node *n, bool pfmemalloc)
2743{
2744	struct page *page;
2745
2746	page = list_first_entry_or_null(&n->slabs_partial,
2747			struct page, lru);
2748	if (!page) {
2749		n->free_touched = 1;
2750		page = list_first_entry_or_null(&n->slabs_free,
2751				struct page, lru);
 
 
2752	}
2753
2754	if (sk_memalloc_socks())
2755		return get_valid_first_slab(n, page, pfmemalloc);
2756
2757	return page;
2758}
2759
2760static noinline void *cache_alloc_pfmemalloc(struct kmem_cache *cachep,
2761				struct kmem_cache_node *n, gfp_t flags)
2762{
2763	struct page *page;
2764	void *obj;
2765	void *list = NULL;
2766
2767	if (!gfp_pfmemalloc_allowed(flags))
2768		return NULL;
2769
2770	spin_lock(&n->list_lock);
2771	page = get_first_slab(n, true);
2772	if (!page) {
2773		spin_unlock(&n->list_lock);
2774		return NULL;
2775	}
2776
2777	obj = slab_get_obj(cachep, page);
2778	n->free_objects--;
2779
2780	fixup_slab_list(cachep, n, page, &list);
2781
2782	spin_unlock(&n->list_lock);
2783	fixup_objfreelist_debug(cachep, &list);
2784
2785	return obj;
2786}
2787
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2788static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2789{
2790	int batchcount;
2791	struct kmem_cache_node *n;
2792	struct array_cache *ac;
2793	int node;
2794	void *list = NULL;
 
2795
2796	check_irq_off();
2797	node = numa_mem_id();
2798
2799retry:
2800	ac = cpu_cache_get(cachep);
2801	batchcount = ac->batchcount;
2802	if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2803		/*
2804		 * If there was little recent activity on this cache, then
2805		 * perform only a partial refill.  Otherwise we could generate
2806		 * refill bouncing.
2807		 */
2808		batchcount = BATCHREFILL_LIMIT;
2809	}
2810	n = get_node(cachep, node);
2811
2812	BUG_ON(ac->avail > 0 || !n);
 
 
 
 
2813	spin_lock(&n->list_lock);
 
2814
2815	/* See if we can refill from the shared array */
2816	if (n->shared && transfer_objects(ac, n->shared, batchcount)) {
2817		n->shared->touched = 1;
2818		goto alloc_done;
2819	}
2820
2821	while (batchcount > 0) {
2822		struct page *page;
2823		/* Get slab alloc is to come from. */
2824		page = get_first_slab(n, false);
2825		if (!page)
2826			goto must_grow;
2827
2828		check_spinlock_acquired(cachep);
2829
2830		/*
2831		 * The slab was either on partial or free list so
2832		 * there must be at least one object available for
2833		 * allocation.
2834		 */
2835		BUG_ON(page->active >= cachep->num);
2836
2837		while (page->active < cachep->num && batchcount--) {
2838			STATS_INC_ALLOCED(cachep);
2839			STATS_INC_ACTIVE(cachep);
2840			STATS_SET_HIGH(cachep);
2841
2842			ac->entry[ac->avail++] = slab_get_obj(cachep, page);
2843		}
2844
2845		fixup_slab_list(cachep, n, page, &list);
2846	}
2847
2848must_grow:
2849	n->free_objects -= ac->avail;
2850alloc_done:
2851	spin_unlock(&n->list_lock);
2852	fixup_objfreelist_debug(cachep, &list);
2853
 
2854	if (unlikely(!ac->avail)) {
2855		int x;
2856
2857		/* Check if we can use obj in pfmemalloc slab */
2858		if (sk_memalloc_socks()) {
2859			void *obj = cache_alloc_pfmemalloc(cachep, n, flags);
2860
2861			if (obj)
2862				return obj;
2863		}
2864
2865		x = cache_grow(cachep, gfp_exact_node(flags), node, NULL);
2866
2867		/* cache_grow can reenable interrupts, then ac could change. */
 
 
 
2868		ac = cpu_cache_get(cachep);
2869		node = numa_mem_id();
 
 
2870
2871		/* no objects in sight? abort */
2872		if (!x && ac->avail == 0)
2873			return NULL;
2874
2875		if (!ac->avail)		/* objects refilled by interrupt? */
2876			goto retry;
2877	}
2878	ac->touched = 1;
2879
2880	return ac->entry[--ac->avail];
2881}
2882
2883static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2884						gfp_t flags)
2885{
2886	might_sleep_if(gfpflags_allow_blocking(flags));
2887#if DEBUG
2888	kmem_flagcheck(cachep, flags);
2889#endif
2890}
2891
2892#if DEBUG
2893static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2894				gfp_t flags, void *objp, unsigned long caller)
2895{
2896	if (!objp)
2897		return objp;
2898	if (cachep->flags & SLAB_POISON) {
2899		check_poison_obj(cachep, objp);
2900		slab_kernel_map(cachep, objp, 1, 0);
2901		poison_obj(cachep, objp, POISON_INUSE);
2902	}
2903	if (cachep->flags & SLAB_STORE_USER)
2904		*dbg_userword(cachep, objp) = (void *)caller;
2905
2906	if (cachep->flags & SLAB_RED_ZONE) {
2907		if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2908				*dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2909			slab_error(cachep, "double free, or memory outside object was overwritten");
2910			pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2911			       objp, *dbg_redzone1(cachep, objp),
2912			       *dbg_redzone2(cachep, objp));
2913		}
2914		*dbg_redzone1(cachep, objp) = RED_ACTIVE;
2915		*dbg_redzone2(cachep, objp) = RED_ACTIVE;
2916	}
2917
2918	objp += obj_offset(cachep);
2919	if (cachep->ctor && cachep->flags & SLAB_POISON)
2920		cachep->ctor(objp);
2921	if (ARCH_SLAB_MINALIGN &&
2922	    ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
2923		pr_err("0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
2924		       objp, (int)ARCH_SLAB_MINALIGN);
2925	}
2926	return objp;
2927}
2928#else
2929#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2930#endif
2931
2932static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
2933{
2934	void *objp;
2935	struct array_cache *ac;
2936
2937	check_irq_off();
2938
2939	ac = cpu_cache_get(cachep);
2940	if (likely(ac->avail)) {
2941		ac->touched = 1;
2942		objp = ac->entry[--ac->avail];
2943
2944		STATS_INC_ALLOCHIT(cachep);
2945		goto out;
2946	}
2947
2948	STATS_INC_ALLOCMISS(cachep);
2949	objp = cache_alloc_refill(cachep, flags);
2950	/*
2951	 * the 'ac' may be updated by cache_alloc_refill(),
2952	 * and kmemleak_erase() requires its correct value.
2953	 */
2954	ac = cpu_cache_get(cachep);
2955
2956out:
2957	/*
2958	 * To avoid a false negative, if an object that is in one of the
2959	 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
2960	 * treat the array pointers as a reference to the object.
2961	 */
2962	if (objp)
2963		kmemleak_erase(&ac->entry[ac->avail]);
2964	return objp;
2965}
2966
2967#ifdef CONFIG_NUMA
2968/*
2969 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
2970 *
2971 * If we are in_interrupt, then process context, including cpusets and
2972 * mempolicy, may not apply and should not be used for allocation policy.
2973 */
2974static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
2975{
2976	int nid_alloc, nid_here;
2977
2978	if (in_interrupt() || (flags & __GFP_THISNODE))
2979		return NULL;
2980	nid_alloc = nid_here = numa_mem_id();
2981	if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
2982		nid_alloc = cpuset_slab_spread_node();
2983	else if (current->mempolicy)
2984		nid_alloc = mempolicy_slab_node();
2985	if (nid_alloc != nid_here)
2986		return ____cache_alloc_node(cachep, flags, nid_alloc);
2987	return NULL;
2988}
2989
2990/*
2991 * Fallback function if there was no memory available and no objects on a
2992 * certain node and fall back is permitted. First we scan all the
2993 * available node for available objects. If that fails then we
2994 * perform an allocation without specifying a node. This allows the page
2995 * allocator to do its reclaim / fallback magic. We then insert the
2996 * slab into the proper nodelist and then allocate from it.
2997 */
2998static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
2999{
3000	struct zonelist *zonelist;
3001	gfp_t local_flags;
3002	struct zoneref *z;
3003	struct zone *zone;
3004	enum zone_type high_zoneidx = gfp_zone(flags);
3005	void *obj = NULL;
 
3006	int nid;
3007	unsigned int cpuset_mems_cookie;
3008
3009	if (flags & __GFP_THISNODE)
3010		return NULL;
3011
3012	local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3013
3014retry_cpuset:
3015	cpuset_mems_cookie = read_mems_allowed_begin();
3016	zonelist = node_zonelist(mempolicy_slab_node(), flags);
3017
3018retry:
3019	/*
3020	 * Look through allowed nodes for objects available
3021	 * from existing per node queues.
3022	 */
3023	for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3024		nid = zone_to_nid(zone);
3025
3026		if (cpuset_zone_allowed(zone, flags) &&
3027			get_node(cache, nid) &&
3028			get_node(cache, nid)->free_objects) {
3029				obj = ____cache_alloc_node(cache,
3030					gfp_exact_node(flags), nid);
3031				if (obj)
3032					break;
3033		}
3034	}
3035
3036	if (!obj) {
3037		/*
3038		 * This allocation will be performed within the constraints
3039		 * of the current cpuset / memory policy requirements.
3040		 * We may trigger various forms of reclaim on the allowed
3041		 * set and go into memory reserves if necessary.
3042		 */
3043		struct page *page;
3044
3045		if (gfpflags_allow_blocking(local_flags))
3046			local_irq_enable();
3047		kmem_flagcheck(cache, flags);
3048		page = kmem_getpages(cache, local_flags, numa_mem_id());
3049		if (gfpflags_allow_blocking(local_flags))
3050			local_irq_disable();
3051		if (page) {
 
 
 
 
3052			/*
3053			 * Insert into the appropriate per node queues
 
3054			 */
3055			nid = page_to_nid(page);
3056			if (cache_grow(cache, flags, nid, page)) {
3057				obj = ____cache_alloc_node(cache,
3058					gfp_exact_node(flags), nid);
3059				if (!obj)
3060					/*
3061					 * Another processor may allocate the
3062					 * objects in the slab since we are
3063					 * not holding any locks.
3064					 */
3065					goto retry;
3066			} else {
3067				/* cache_grow already freed obj */
3068				obj = NULL;
3069			}
3070		}
3071	}
3072
3073	if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
3074		goto retry_cpuset;
3075	return obj;
3076}
3077
3078/*
3079 * A interface to enable slab creation on nodeid
3080 */
3081static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3082				int nodeid)
3083{
3084	struct page *page;
3085	struct kmem_cache_node *n;
3086	void *obj;
3087	void *list = NULL;
3088	int x;
3089
3090	VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES);
3091	n = get_node(cachep, nodeid);
3092	BUG_ON(!n);
3093
3094retry:
3095	check_irq_off();
3096	spin_lock(&n->list_lock);
3097	page = get_first_slab(n, false);
3098	if (!page)
3099		goto must_grow;
3100
3101	check_spinlock_acquired_node(cachep, nodeid);
3102
3103	STATS_INC_NODEALLOCS(cachep);
3104	STATS_INC_ACTIVE(cachep);
3105	STATS_SET_HIGH(cachep);
3106
3107	BUG_ON(page->active == cachep->num);
3108
3109	obj = slab_get_obj(cachep, page);
3110	n->free_objects--;
3111
3112	fixup_slab_list(cachep, n, page, &list);
3113
3114	spin_unlock(&n->list_lock);
3115	fixup_objfreelist_debug(cachep, &list);
3116	goto done;
3117
3118must_grow:
3119	spin_unlock(&n->list_lock);
3120	x = cache_grow(cachep, gfp_exact_node(flags), nodeid, NULL);
3121	if (x)
3122		goto retry;
3123
3124	return fallback_alloc(cachep, flags);
 
3125
3126done:
3127	return obj;
3128}
3129
3130static __always_inline void *
3131slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3132		   unsigned long caller)
3133{
3134	unsigned long save_flags;
3135	void *ptr;
3136	int slab_node = numa_mem_id();
3137
3138	flags &= gfp_allowed_mask;
3139	cachep = slab_pre_alloc_hook(cachep, flags);
3140	if (unlikely(!cachep))
3141		return NULL;
3142
3143	cache_alloc_debugcheck_before(cachep, flags);
3144	local_irq_save(save_flags);
3145
3146	if (nodeid == NUMA_NO_NODE)
3147		nodeid = slab_node;
3148
3149	if (unlikely(!get_node(cachep, nodeid))) {
3150		/* Node not bootstrapped yet */
3151		ptr = fallback_alloc(cachep, flags);
3152		goto out;
3153	}
3154
3155	if (nodeid == slab_node) {
3156		/*
3157		 * Use the locally cached objects if possible.
3158		 * However ____cache_alloc does not allow fallback
3159		 * to other nodes. It may fail while we still have
3160		 * objects on other nodes available.
3161		 */
3162		ptr = ____cache_alloc(cachep, flags);
3163		if (ptr)
3164			goto out;
3165	}
3166	/* ___cache_alloc_node can fall back to other nodes */
3167	ptr = ____cache_alloc_node(cachep, flags, nodeid);
3168  out:
3169	local_irq_restore(save_flags);
3170	ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3171
3172	if (unlikely(flags & __GFP_ZERO) && ptr)
3173		memset(ptr, 0, cachep->object_size);
3174
3175	slab_post_alloc_hook(cachep, flags, 1, &ptr);
3176	return ptr;
3177}
3178
3179static __always_inline void *
3180__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3181{
3182	void *objp;
3183
3184	if (current->mempolicy || cpuset_do_slab_mem_spread()) {
3185		objp = alternate_node_alloc(cache, flags);
3186		if (objp)
3187			goto out;
3188	}
3189	objp = ____cache_alloc(cache, flags);
3190
3191	/*
3192	 * We may just have run out of memory on the local node.
3193	 * ____cache_alloc_node() knows how to locate memory on other nodes
3194	 */
3195	if (!objp)
3196		objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3197
3198  out:
3199	return objp;
3200}
3201#else
3202
3203static __always_inline void *
3204__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3205{
3206	return ____cache_alloc(cachep, flags);
3207}
3208
3209#endif /* CONFIG_NUMA */
3210
3211static __always_inline void *
3212slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3213{
3214	unsigned long save_flags;
3215	void *objp;
3216
3217	flags &= gfp_allowed_mask;
3218	cachep = slab_pre_alloc_hook(cachep, flags);
3219	if (unlikely(!cachep))
3220		return NULL;
3221
3222	cache_alloc_debugcheck_before(cachep, flags);
3223	local_irq_save(save_flags);
3224	objp = __do_cache_alloc(cachep, flags);
3225	local_irq_restore(save_flags);
3226	objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3227	prefetchw(objp);
3228
3229	if (unlikely(flags & __GFP_ZERO) && objp)
3230		memset(objp, 0, cachep->object_size);
3231
3232	slab_post_alloc_hook(cachep, flags, 1, &objp);
3233	return objp;
3234}
3235
3236/*
3237 * Caller needs to acquire correct kmem_cache_node's list_lock
3238 * @list: List of detached free slabs should be freed by caller
3239 */
3240static void free_block(struct kmem_cache *cachep, void **objpp,
3241			int nr_objects, int node, struct list_head *list)
3242{
3243	int i;
3244	struct kmem_cache_node *n = get_node(cachep, node);
 
 
 
3245
3246	for (i = 0; i < nr_objects; i++) {
3247		void *objp;
3248		struct page *page;
3249
3250		objp = objpp[i];
3251
3252		page = virt_to_head_page(objp);
3253		list_del(&page->lru);
3254		check_spinlock_acquired_node(cachep, node);
3255		slab_put_obj(cachep, page, objp);
3256		STATS_DEC_ACTIVE(cachep);
3257		n->free_objects++;
3258
3259		/* fixup slab chains */
3260		if (page->active == 0) {
3261			if (n->free_objects > n->free_limit) {
3262				n->free_objects -= cachep->num;
3263				list_add_tail(&page->lru, list);
3264			} else {
3265				list_add(&page->lru, &n->slabs_free);
3266			}
3267		} else {
3268			/* Unconditionally move a slab to the end of the
3269			 * partial list on free - maximum time for the
3270			 * other objects to be freed, too.
3271			 */
3272			list_add_tail(&page->lru, &n->slabs_partial);
3273		}
3274	}
 
 
 
 
 
 
 
 
 
3275}
3276
3277static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3278{
3279	int batchcount;
3280	struct kmem_cache_node *n;
3281	int node = numa_mem_id();
3282	LIST_HEAD(list);
3283
3284	batchcount = ac->batchcount;
3285
3286	check_irq_off();
3287	n = get_node(cachep, node);
3288	spin_lock(&n->list_lock);
3289	if (n->shared) {
3290		struct array_cache *shared_array = n->shared;
3291		int max = shared_array->limit - shared_array->avail;
3292		if (max) {
3293			if (batchcount > max)
3294				batchcount = max;
3295			memcpy(&(shared_array->entry[shared_array->avail]),
3296			       ac->entry, sizeof(void *) * batchcount);
3297			shared_array->avail += batchcount;
3298			goto free_done;
3299		}
3300	}
3301
3302	free_block(cachep, ac->entry, batchcount, node, &list);
3303free_done:
3304#if STATS
3305	{
3306		int i = 0;
3307		struct page *page;
3308
3309		list_for_each_entry(page, &n->slabs_free, lru) {
3310			BUG_ON(page->active);
3311
3312			i++;
3313		}
3314		STATS_SET_FREEABLE(cachep, i);
3315	}
3316#endif
3317	spin_unlock(&n->list_lock);
3318	slabs_destroy(cachep, &list);
3319	ac->avail -= batchcount;
3320	memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3321}
3322
3323/*
3324 * Release an obj back to its cache. If the obj has a constructed state, it must
3325 * be in this state _before_ it is released.  Called with disabled ints.
3326 */
3327static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3328				unsigned long caller)
3329{
3330	struct array_cache *ac = cpu_cache_get(cachep);
 
 
3331
3332	kasan_slab_free(cachep, objp);
 
 
 
 
 
 
3333
3334	check_irq_off();
3335	kmemleak_free_recursive(objp, cachep->flags);
3336	objp = cache_free_debugcheck(cachep, objp, caller);
3337
3338	kmemcheck_slab_free(cachep, objp, cachep->object_size);
3339
3340	/*
3341	 * Skip calling cache_free_alien() when the platform is not numa.
3342	 * This will avoid cache misses that happen while accessing slabp (which
3343	 * is per page memory  reference) to get nodeid. Instead use a global
3344	 * variable to skip the call, which is mostly likely to be present in
3345	 * the cache.
3346	 */
3347	if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3348		return;
3349
3350	if (ac->avail < ac->limit) {
3351		STATS_INC_FREEHIT(cachep);
3352	} else {
3353		STATS_INC_FREEMISS(cachep);
3354		cache_flusharray(cachep, ac);
3355	}
3356
3357	if (sk_memalloc_socks()) {
3358		struct page *page = virt_to_head_page(objp);
3359
3360		if (unlikely(PageSlabPfmemalloc(page))) {
3361			cache_free_pfmemalloc(cachep, page, objp);
3362			return;
3363		}
3364	}
3365
3366	ac->entry[ac->avail++] = objp;
3367}
3368
3369/**
3370 * kmem_cache_alloc - Allocate an object
3371 * @cachep: The cache to allocate from.
3372 * @flags: See kmalloc().
3373 *
3374 * Allocate an object from this cache.  The flags are only relevant
3375 * if the cache has no available objects.
3376 */
3377void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3378{
3379	void *ret = slab_alloc(cachep, flags, _RET_IP_);
3380
3381	kasan_slab_alloc(cachep, ret, flags);
3382	trace_kmem_cache_alloc(_RET_IP_, ret,
3383			       cachep->object_size, cachep->size, flags);
3384
3385	return ret;
3386}
3387EXPORT_SYMBOL(kmem_cache_alloc);
3388
3389static __always_inline void
3390cache_alloc_debugcheck_after_bulk(struct kmem_cache *s, gfp_t flags,
3391				  size_t size, void **p, unsigned long caller)
3392{
3393	size_t i;
3394
3395	for (i = 0; i < size; i++)
3396		p[i] = cache_alloc_debugcheck_after(s, flags, p[i], caller);
3397}
3398
3399int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3400			  void **p)
3401{
3402	size_t i;
3403
3404	s = slab_pre_alloc_hook(s, flags);
3405	if (!s)
3406		return 0;
3407
3408	cache_alloc_debugcheck_before(s, flags);
3409
3410	local_irq_disable();
3411	for (i = 0; i < size; i++) {
3412		void *objp = __do_cache_alloc(s, flags);
3413
3414		if (unlikely(!objp))
3415			goto error;
3416		p[i] = objp;
3417	}
3418	local_irq_enable();
3419
3420	cache_alloc_debugcheck_after_bulk(s, flags, size, p, _RET_IP_);
3421
3422	/* Clear memory outside IRQ disabled section */
3423	if (unlikely(flags & __GFP_ZERO))
3424		for (i = 0; i < size; i++)
3425			memset(p[i], 0, s->object_size);
3426
3427	slab_post_alloc_hook(s, flags, size, p);
3428	/* FIXME: Trace call missing. Christoph would like a bulk variant */
3429	return size;
3430error:
3431	local_irq_enable();
3432	cache_alloc_debugcheck_after_bulk(s, flags, i, p, _RET_IP_);
3433	slab_post_alloc_hook(s, flags, i, p);
3434	__kmem_cache_free_bulk(s, i, p);
3435	return 0;
3436}
3437EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3438
3439#ifdef CONFIG_TRACING
3440void *
3441kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3442{
3443	void *ret;
3444
3445	ret = slab_alloc(cachep, flags, _RET_IP_);
3446
3447	kasan_kmalloc(cachep, ret, size, flags);
3448	trace_kmalloc(_RET_IP_, ret,
3449		      size, cachep->size, flags);
3450	return ret;
3451}
3452EXPORT_SYMBOL(kmem_cache_alloc_trace);
3453#endif
3454
3455#ifdef CONFIG_NUMA
3456/**
3457 * kmem_cache_alloc_node - Allocate an object on the specified node
3458 * @cachep: The cache to allocate from.
3459 * @flags: See kmalloc().
3460 * @nodeid: node number of the target node.
3461 *
3462 * Identical to kmem_cache_alloc but it will allocate memory on the given
3463 * node, which can improve the performance for cpu bound structures.
3464 *
3465 * Fallback to other node is possible if __GFP_THISNODE is not set.
3466 */
3467void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3468{
3469	void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3470
3471	kasan_slab_alloc(cachep, ret, flags);
3472	trace_kmem_cache_alloc_node(_RET_IP_, ret,
3473				    cachep->object_size, cachep->size,
3474				    flags, nodeid);
3475
3476	return ret;
3477}
3478EXPORT_SYMBOL(kmem_cache_alloc_node);
3479
3480#ifdef CONFIG_TRACING
3481void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3482				  gfp_t flags,
3483				  int nodeid,
3484				  size_t size)
3485{
3486	void *ret;
3487
3488	ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3489
3490	kasan_kmalloc(cachep, ret, size, flags);
3491	trace_kmalloc_node(_RET_IP_, ret,
3492			   size, cachep->size,
3493			   flags, nodeid);
3494	return ret;
3495}
3496EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3497#endif
3498
3499static __always_inline void *
3500__do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3501{
3502	struct kmem_cache *cachep;
3503	void *ret;
3504
3505	cachep = kmalloc_slab(size, flags);
3506	if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3507		return cachep;
3508	ret = kmem_cache_alloc_node_trace(cachep, flags, node, size);
3509	kasan_kmalloc(cachep, ret, size, flags);
3510
3511	return ret;
3512}
3513
3514void *__kmalloc_node(size_t size, gfp_t flags, int node)
3515{
3516	return __do_kmalloc_node(size, flags, node, _RET_IP_);
3517}
3518EXPORT_SYMBOL(__kmalloc_node);
3519
3520void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3521		int node, unsigned long caller)
3522{
3523	return __do_kmalloc_node(size, flags, node, caller);
3524}
3525EXPORT_SYMBOL(__kmalloc_node_track_caller);
3526#endif /* CONFIG_NUMA */
3527
3528/**
3529 * __do_kmalloc - allocate memory
3530 * @size: how many bytes of memory are required.
3531 * @flags: the type of memory to allocate (see kmalloc).
3532 * @caller: function caller for debug tracking of the caller
3533 */
3534static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3535					  unsigned long caller)
3536{
3537	struct kmem_cache *cachep;
3538	void *ret;
3539
3540	cachep = kmalloc_slab(size, flags);
3541	if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3542		return cachep;
3543	ret = slab_alloc(cachep, flags, caller);
3544
3545	kasan_kmalloc(cachep, ret, size, flags);
3546	trace_kmalloc(caller, ret,
3547		      size, cachep->size, flags);
3548
3549	return ret;
3550}
3551
3552void *__kmalloc(size_t size, gfp_t flags)
3553{
3554	return __do_kmalloc(size, flags, _RET_IP_);
3555}
3556EXPORT_SYMBOL(__kmalloc);
3557
3558void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3559{
3560	return __do_kmalloc(size, flags, caller);
3561}
3562EXPORT_SYMBOL(__kmalloc_track_caller);
3563
3564/**
3565 * kmem_cache_free - Deallocate an object
3566 * @cachep: The cache the allocation was from.
3567 * @objp: The previously allocated object.
3568 *
3569 * Free an object which was previously allocated from this
3570 * cache.
3571 */
3572void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3573{
3574	unsigned long flags;
3575	cachep = cache_from_obj(cachep, objp);
3576	if (!cachep)
3577		return;
3578
3579	local_irq_save(flags);
3580	debug_check_no_locks_freed(objp, cachep->object_size);
3581	if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3582		debug_check_no_obj_freed(objp, cachep->object_size);
3583	__cache_free(cachep, objp, _RET_IP_);
3584	local_irq_restore(flags);
3585
3586	trace_kmem_cache_free(_RET_IP_, objp);
3587}
3588EXPORT_SYMBOL(kmem_cache_free);
3589
3590void kmem_cache_free_bulk(struct kmem_cache *orig_s, size_t size, void **p)
3591{
3592	struct kmem_cache *s;
3593	size_t i;
3594
3595	local_irq_disable();
3596	for (i = 0; i < size; i++) {
3597		void *objp = p[i];
3598
3599		if (!orig_s) /* called via kfree_bulk */
3600			s = virt_to_cache(objp);
3601		else
3602			s = cache_from_obj(orig_s, objp);
3603
3604		debug_check_no_locks_freed(objp, s->object_size);
3605		if (!(s->flags & SLAB_DEBUG_OBJECTS))
3606			debug_check_no_obj_freed(objp, s->object_size);
3607
3608		__cache_free(s, objp, _RET_IP_);
3609	}
3610	local_irq_enable();
3611
3612	/* FIXME: add tracing */
3613}
3614EXPORT_SYMBOL(kmem_cache_free_bulk);
3615
3616/**
3617 * kfree - free previously allocated memory
3618 * @objp: pointer returned by kmalloc.
3619 *
3620 * If @objp is NULL, no operation is performed.
3621 *
3622 * Don't free memory not originally allocated by kmalloc()
3623 * or you will run into trouble.
3624 */
3625void kfree(const void *objp)
3626{
3627	struct kmem_cache *c;
3628	unsigned long flags;
3629
3630	trace_kfree(_RET_IP_, objp);
3631
3632	if (unlikely(ZERO_OR_NULL_PTR(objp)))
3633		return;
3634	local_irq_save(flags);
3635	kfree_debugcheck(objp);
3636	c = virt_to_cache(objp);
3637	debug_check_no_locks_freed(objp, c->object_size);
3638
3639	debug_check_no_obj_freed(objp, c->object_size);
3640	__cache_free(c, (void *)objp, _RET_IP_);
3641	local_irq_restore(flags);
3642}
3643EXPORT_SYMBOL(kfree);
3644
3645/*
3646 * This initializes kmem_cache_node or resizes various caches for all nodes.
3647 */
3648static int alloc_kmem_cache_node(struct kmem_cache *cachep, gfp_t gfp)
3649{
 
3650	int node;
3651	struct kmem_cache_node *n;
3652	struct array_cache *new_shared;
3653	struct alien_cache **new_alien = NULL;
3654
3655	for_each_online_node(node) {
3656
3657		if (use_alien_caches) {
3658			new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3659			if (!new_alien)
3660				goto fail;
3661		}
3662
3663		new_shared = NULL;
3664		if (cachep->shared) {
3665			new_shared = alloc_arraycache(node,
3666				cachep->shared*cachep->batchcount,
3667					0xbaadf00d, gfp);
3668			if (!new_shared) {
3669				free_alien_cache(new_alien);
3670				goto fail;
3671			}
3672		}
3673
3674		n = get_node(cachep, node);
3675		if (n) {
3676			struct array_cache *shared = n->shared;
3677			LIST_HEAD(list);
3678
3679			spin_lock_irq(&n->list_lock);
3680
3681			if (shared)
3682				free_block(cachep, shared->entry,
3683						shared->avail, node, &list);
3684
3685			n->shared = new_shared;
3686			if (!n->alien) {
3687				n->alien = new_alien;
3688				new_alien = NULL;
3689			}
3690			n->free_limit = (1 + nr_cpus_node(node)) *
3691					cachep->batchcount + cachep->num;
3692			spin_unlock_irq(&n->list_lock);
3693			slabs_destroy(cachep, &list);
3694			kfree(shared);
3695			free_alien_cache(new_alien);
3696			continue;
3697		}
3698		n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3699		if (!n) {
3700			free_alien_cache(new_alien);
3701			kfree(new_shared);
3702			goto fail;
3703		}
3704
3705		kmem_cache_node_init(n);
3706		n->next_reap = jiffies + REAPTIMEOUT_NODE +
3707				((unsigned long)cachep) % REAPTIMEOUT_NODE;
3708		n->shared = new_shared;
3709		n->alien = new_alien;
3710		n->free_limit = (1 + nr_cpus_node(node)) *
3711					cachep->batchcount + cachep->num;
3712		cachep->node[node] = n;
3713	}
 
3714	return 0;
3715
3716fail:
3717	if (!cachep->list.next) {
3718		/* Cache is not active yet. Roll back what we did */
3719		node--;
3720		while (node >= 0) {
3721			n = get_node(cachep, node);
3722			if (n) {
3723				kfree(n->shared);
3724				free_alien_cache(n->alien);
3725				kfree(n);
3726				cachep->node[node] = NULL;
3727			}
3728			node--;
3729		}
3730	}
3731	return -ENOMEM;
3732}
3733
3734/* Always called with the slab_mutex held */
3735static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
3736				int batchcount, int shared, gfp_t gfp)
3737{
3738	struct array_cache __percpu *cpu_cache, *prev;
3739	int cpu;
3740
3741	cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount);
3742	if (!cpu_cache)
3743		return -ENOMEM;
3744
3745	prev = cachep->cpu_cache;
3746	cachep->cpu_cache = cpu_cache;
3747	kick_all_cpus_sync();
 
 
 
 
 
3748
3749	check_irq_on();
3750	cachep->batchcount = batchcount;
3751	cachep->limit = limit;
3752	cachep->shared = shared;
3753
3754	if (!prev)
3755		goto alloc_node;
3756
3757	for_each_online_cpu(cpu) {
3758		LIST_HEAD(list);
3759		int node;
3760		struct kmem_cache_node *n;
3761		struct array_cache *ac = per_cpu_ptr(prev, cpu);
3762
3763		node = cpu_to_mem(cpu);
3764		n = get_node(cachep, node);
3765		spin_lock_irq(&n->list_lock);
3766		free_block(cachep, ac->entry, ac->avail, node, &list);
3767		spin_unlock_irq(&n->list_lock);
3768		slabs_destroy(cachep, &list);
3769	}
3770	free_percpu(prev);
3771
3772alloc_node:
3773	return alloc_kmem_cache_node(cachep, gfp);
3774}
3775
3776static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3777				int batchcount, int shared, gfp_t gfp)
3778{
3779	int ret;
3780	struct kmem_cache *c;
3781
3782	ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3783
3784	if (slab_state < FULL)
3785		return ret;
3786
3787	if ((ret < 0) || !is_root_cache(cachep))
3788		return ret;
3789
3790	lockdep_assert_held(&slab_mutex);
3791	for_each_memcg_cache(c, cachep) {
3792		/* return value determined by the root cache only */
3793		__do_tune_cpucache(c, limit, batchcount, shared, gfp);
3794	}
3795
3796	return ret;
3797}
3798
3799/* Called with slab_mutex held always */
3800static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3801{
3802	int err;
3803	int limit = 0;
3804	int shared = 0;
3805	int batchcount = 0;
3806
 
 
 
 
3807	if (!is_root_cache(cachep)) {
3808		struct kmem_cache *root = memcg_root_cache(cachep);
3809		limit = root->limit;
3810		shared = root->shared;
3811		batchcount = root->batchcount;
3812	}
3813
3814	if (limit && shared && batchcount)
3815		goto skip_setup;
3816	/*
3817	 * The head array serves three purposes:
3818	 * - create a LIFO ordering, i.e. return objects that are cache-warm
3819	 * - reduce the number of spinlock operations.
3820	 * - reduce the number of linked list operations on the slab and
3821	 *   bufctl chains: array operations are cheaper.
3822	 * The numbers are guessed, we should auto-tune as described by
3823	 * Bonwick.
3824	 */
3825	if (cachep->size > 131072)
3826		limit = 1;
3827	else if (cachep->size > PAGE_SIZE)
3828		limit = 8;
3829	else if (cachep->size > 1024)
3830		limit = 24;
3831	else if (cachep->size > 256)
3832		limit = 54;
3833	else
3834		limit = 120;
3835
3836	/*
3837	 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3838	 * allocation behaviour: Most allocs on one cpu, most free operations
3839	 * on another cpu. For these cases, an efficient object passing between
3840	 * cpus is necessary. This is provided by a shared array. The array
3841	 * replaces Bonwick's magazine layer.
3842	 * On uniprocessor, it's functionally equivalent (but less efficient)
3843	 * to a larger limit. Thus disabled by default.
3844	 */
3845	shared = 0;
3846	if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
3847		shared = 8;
3848
3849#if DEBUG
3850	/*
3851	 * With debugging enabled, large batchcount lead to excessively long
3852	 * periods with disabled local interrupts. Limit the batchcount
3853	 */
3854	if (limit > 32)
3855		limit = 32;
3856#endif
3857	batchcount = (limit + 1) / 2;
3858skip_setup:
3859	err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
 
3860	if (err)
3861		pr_err("enable_cpucache failed for %s, error %d\n",
3862		       cachep->name, -err);
3863	return err;
3864}
3865
3866/*
3867 * Drain an array if it contains any elements taking the node lock only if
3868 * necessary. Note that the node listlock also protects the array_cache
3869 * if drain_array() is used on the shared array.
3870 */
3871static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
3872			 struct array_cache *ac, int force, int node)
3873{
3874	LIST_HEAD(list);
3875	int tofree;
 
 
3876
3877	if (!ac || !ac->avail)
3878		return;
3879	if (ac->touched && !force) {
 
3880		ac->touched = 0;
3881	} else {
3882		spin_lock_irq(&n->list_lock);
3883		if (ac->avail) {
3884			tofree = force ? ac->avail : (ac->limit + 4) / 5;
3885			if (tofree > ac->avail)
3886				tofree = (ac->avail + 1) / 2;
3887			free_block(cachep, ac->entry, tofree, node, &list);
3888			ac->avail -= tofree;
3889			memmove(ac->entry, &(ac->entry[tofree]),
3890				sizeof(void *) * ac->avail);
3891		}
3892		spin_unlock_irq(&n->list_lock);
3893		slabs_destroy(cachep, &list);
3894	}
 
 
 
 
 
 
3895}
3896
3897/**
3898 * cache_reap - Reclaim memory from caches.
3899 * @w: work descriptor
3900 *
3901 * Called from workqueue/eventd every few seconds.
3902 * Purpose:
3903 * - clear the per-cpu caches for this CPU.
3904 * - return freeable pages to the main free memory pool.
3905 *
3906 * If we cannot acquire the cache chain mutex then just give up - we'll try
3907 * again on the next iteration.
3908 */
3909static void cache_reap(struct work_struct *w)
3910{
3911	struct kmem_cache *searchp;
3912	struct kmem_cache_node *n;
3913	int node = numa_mem_id();
3914	struct delayed_work *work = to_delayed_work(w);
3915
3916	if (!mutex_trylock(&slab_mutex))
3917		/* Give up. Setup the next iteration. */
3918		goto out;
3919
3920	list_for_each_entry(searchp, &slab_caches, list) {
3921		check_irq_on();
3922
3923		/*
3924		 * We only take the node lock if absolutely necessary and we
3925		 * have established with reasonable certainty that
3926		 * we can do some work if the lock was obtained.
3927		 */
3928		n = get_node(searchp, node);
3929
3930		reap_alien(searchp, n);
3931
3932		drain_array(searchp, n, cpu_cache_get(searchp), 0, node);
3933
3934		/*
3935		 * These are racy checks but it does not matter
3936		 * if we skip one check or scan twice.
3937		 */
3938		if (time_after(n->next_reap, jiffies))
3939			goto next;
3940
3941		n->next_reap = jiffies + REAPTIMEOUT_NODE;
3942
3943		drain_array(searchp, n, n->shared, 0, node);
3944
3945		if (n->free_touched)
3946			n->free_touched = 0;
3947		else {
3948			int freed;
3949
3950			freed = drain_freelist(searchp, n, (n->free_limit +
3951				5 * searchp->num - 1) / (5 * searchp->num));
3952			STATS_ADD_REAPED(searchp, freed);
3953		}
3954next:
3955		cond_resched();
3956	}
3957	check_irq_on();
3958	mutex_unlock(&slab_mutex);
3959	next_reap_node();
3960out:
3961	/* Set up the next iteration */
3962	schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_AC));
 
3963}
3964
3965#ifdef CONFIG_SLABINFO
3966void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
3967{
3968	struct page *page;
3969	unsigned long active_objs;
3970	unsigned long num_objs;
3971	unsigned long active_slabs = 0;
3972	unsigned long num_slabs, free_objects = 0, shared_avail = 0;
3973	const char *name;
3974	char *error = NULL;
3975	int node;
3976	struct kmem_cache_node *n;
3977
3978	active_objs = 0;
3979	num_slabs = 0;
3980	for_each_kmem_cache_node(cachep, node, n) {
3981
3982		check_irq_on();
3983		spin_lock_irq(&n->list_lock);
3984
3985		list_for_each_entry(page, &n->slabs_full, lru) {
3986			if (page->active != cachep->num && !error)
3987				error = "slabs_full accounting error";
3988			active_objs += cachep->num;
3989			active_slabs++;
3990		}
3991		list_for_each_entry(page, &n->slabs_partial, lru) {
3992			if (page->active == cachep->num && !error)
3993				error = "slabs_partial accounting error";
3994			if (!page->active && !error)
3995				error = "slabs_partial accounting error";
3996			active_objs += page->active;
3997			active_slabs++;
3998		}
3999		list_for_each_entry(page, &n->slabs_free, lru) {
4000			if (page->active && !error)
4001				error = "slabs_free accounting error";
4002			num_slabs++;
4003		}
4004		free_objects += n->free_objects;
4005		if (n->shared)
4006			shared_avail += n->shared->avail;
4007
4008		spin_unlock_irq(&n->list_lock);
4009	}
4010	num_slabs += active_slabs;
4011	num_objs = num_slabs * cachep->num;
4012	if (num_objs - active_objs != free_objects && !error)
4013		error = "free_objects accounting error";
4014
4015	name = cachep->name;
4016	if (error)
4017		pr_err("slab: cache %s error: %s\n", name, error);
4018
4019	sinfo->active_objs = active_objs;
4020	sinfo->num_objs = num_objs;
4021	sinfo->active_slabs = active_slabs;
4022	sinfo->num_slabs = num_slabs;
4023	sinfo->shared_avail = shared_avail;
4024	sinfo->limit = cachep->limit;
4025	sinfo->batchcount = cachep->batchcount;
4026	sinfo->shared = cachep->shared;
4027	sinfo->objects_per_slab = cachep->num;
4028	sinfo->cache_order = cachep->gfporder;
4029}
4030
4031void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4032{
4033#if STATS
4034	{			/* node stats */
4035		unsigned long high = cachep->high_mark;
4036		unsigned long allocs = cachep->num_allocations;
4037		unsigned long grown = cachep->grown;
4038		unsigned long reaped = cachep->reaped;
4039		unsigned long errors = cachep->errors;
4040		unsigned long max_freeable = cachep->max_freeable;
4041		unsigned long node_allocs = cachep->node_allocs;
4042		unsigned long node_frees = cachep->node_frees;
4043		unsigned long overflows = cachep->node_overflow;
4044
4045		seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
4046			   allocs, high, grown,
4047			   reaped, errors, max_freeable, node_allocs,
4048			   node_frees, overflows);
4049	}
4050	/* cpu stats */
4051	{
4052		unsigned long allochit = atomic_read(&cachep->allochit);
4053		unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4054		unsigned long freehit = atomic_read(&cachep->freehit);
4055		unsigned long freemiss = atomic_read(&cachep->freemiss);
4056
4057		seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4058			   allochit, allocmiss, freehit, freemiss);
4059	}
4060#endif
4061}
4062
4063#define MAX_SLABINFO_WRITE 128
4064/**
4065 * slabinfo_write - Tuning for the slab allocator
4066 * @file: unused
4067 * @buffer: user buffer
4068 * @count: data length
4069 * @ppos: unused
4070 */
4071ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4072		       size_t count, loff_t *ppos)
4073{
4074	char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4075	int limit, batchcount, shared, res;
4076	struct kmem_cache *cachep;
4077
4078	if (count > MAX_SLABINFO_WRITE)
4079		return -EINVAL;
4080	if (copy_from_user(&kbuf, buffer, count))
4081		return -EFAULT;
4082	kbuf[MAX_SLABINFO_WRITE] = '\0';
4083
4084	tmp = strchr(kbuf, ' ');
4085	if (!tmp)
4086		return -EINVAL;
4087	*tmp = '\0';
4088	tmp++;
4089	if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4090		return -EINVAL;
4091
4092	/* Find the cache in the chain of caches. */
4093	mutex_lock(&slab_mutex);
4094	res = -EINVAL;
4095	list_for_each_entry(cachep, &slab_caches, list) {
4096		if (!strcmp(cachep->name, kbuf)) {
4097			if (limit < 1 || batchcount < 1 ||
4098					batchcount > limit || shared < 0) {
4099				res = 0;
4100			} else {
4101				res = do_tune_cpucache(cachep, limit,
4102						       batchcount, shared,
4103						       GFP_KERNEL);
4104			}
4105			break;
4106		}
4107	}
4108	mutex_unlock(&slab_mutex);
4109	if (res >= 0)
4110		res = count;
4111	return res;
4112}
4113
4114#ifdef CONFIG_DEBUG_SLAB_LEAK
4115
4116static inline int add_caller(unsigned long *n, unsigned long v)
4117{
4118	unsigned long *p;
4119	int l;
4120	if (!v)
4121		return 1;
4122	l = n[1];
4123	p = n + 2;
4124	while (l) {
4125		int i = l/2;
4126		unsigned long *q = p + 2 * i;
4127		if (*q == v) {
4128			q[1]++;
4129			return 1;
4130		}
4131		if (*q > v) {
4132			l = i;
4133		} else {
4134			p = q + 2;
4135			l -= i + 1;
4136		}
4137	}
4138	if (++n[1] == n[0])
4139		return 0;
4140	memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4141	p[0] = v;
4142	p[1] = 1;
4143	return 1;
4144}
4145
4146static void handle_slab(unsigned long *n, struct kmem_cache *c,
4147						struct page *page)
4148{
4149	void *p;
4150	int i, j;
4151	unsigned long v;
4152
4153	if (n[0] == n[1])
4154		return;
4155	for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
4156		bool active = true;
4157
4158		for (j = page->active; j < c->num; j++) {
4159			if (get_free_obj(page, j) == i) {
4160				active = false;
4161				break;
4162			}
4163		}
4164
4165		if (!active)
4166			continue;
4167
4168		/*
4169		 * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table
4170		 * mapping is established when actual object allocation and
4171		 * we could mistakenly access the unmapped object in the cpu
4172		 * cache.
4173		 */
4174		if (probe_kernel_read(&v, dbg_userword(c, p), sizeof(v)))
4175			continue;
4176
4177		if (!add_caller(n, v))
4178			return;
4179	}
4180}
4181
4182static void show_symbol(struct seq_file *m, unsigned long address)
4183{
4184#ifdef CONFIG_KALLSYMS
4185	unsigned long offset, size;
4186	char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4187
4188	if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4189		seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4190		if (modname[0])
4191			seq_printf(m, " [%s]", modname);
4192		return;
4193	}
4194#endif
4195	seq_printf(m, "%p", (void *)address);
4196}
4197
4198static int leaks_show(struct seq_file *m, void *p)
4199{
4200	struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4201	struct page *page;
4202	struct kmem_cache_node *n;
4203	const char *name;
4204	unsigned long *x = m->private;
4205	int node;
4206	int i;
4207
4208	if (!(cachep->flags & SLAB_STORE_USER))
4209		return 0;
4210	if (!(cachep->flags & SLAB_RED_ZONE))
4211		return 0;
4212
4213	/*
4214	 * Set store_user_clean and start to grab stored user information
4215	 * for all objects on this cache. If some alloc/free requests comes
4216	 * during the processing, information would be wrong so restart
4217	 * whole processing.
4218	 */
4219	do {
4220		set_store_user_clean(cachep);
4221		drain_cpu_caches(cachep);
4222
4223		x[1] = 0;
4224
4225		for_each_kmem_cache_node(cachep, node, n) {
4226
4227			check_irq_on();
4228			spin_lock_irq(&n->list_lock);
4229
4230			list_for_each_entry(page, &n->slabs_full, lru)
4231				handle_slab(x, cachep, page);
4232			list_for_each_entry(page, &n->slabs_partial, lru)
4233				handle_slab(x, cachep, page);
4234			spin_unlock_irq(&n->list_lock);
4235		}
4236	} while (!is_store_user_clean(cachep));
4237
4238	name = cachep->name;
4239	if (x[0] == x[1]) {
4240		/* Increase the buffer size */
4241		mutex_unlock(&slab_mutex);
4242		m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4243		if (!m->private) {
4244			/* Too bad, we are really out */
4245			m->private = x;
4246			mutex_lock(&slab_mutex);
4247			return -ENOMEM;
4248		}
4249		*(unsigned long *)m->private = x[0] * 2;
4250		kfree(x);
4251		mutex_lock(&slab_mutex);
4252		/* Now make sure this entry will be retried */
4253		m->count = m->size;
4254		return 0;
4255	}
4256	for (i = 0; i < x[1]; i++) {
4257		seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4258		show_symbol(m, x[2*i+2]);
4259		seq_putc(m, '\n');
4260	}
4261
4262	return 0;
4263}
4264
4265static const struct seq_operations slabstats_op = {
4266	.start = slab_start,
4267	.next = slab_next,
4268	.stop = slab_stop,
4269	.show = leaks_show,
4270};
4271
4272static int slabstats_open(struct inode *inode, struct file *file)
4273{
4274	unsigned long *n;
4275
4276	n = __seq_open_private(file, &slabstats_op, PAGE_SIZE);
4277	if (!n)
4278		return -ENOMEM;
4279
4280	*n = PAGE_SIZE / (2 * sizeof(unsigned long));
4281
4282	return 0;
4283}
4284
4285static const struct file_operations proc_slabstats_operations = {
4286	.open		= slabstats_open,
4287	.read		= seq_read,
4288	.llseek		= seq_lseek,
4289	.release	= seq_release_private,
4290};
4291#endif
4292
4293static int __init slab_proc_init(void)
4294{
4295#ifdef CONFIG_DEBUG_SLAB_LEAK
4296	proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4297#endif
4298	return 0;
4299}
4300module_init(slab_proc_init);
4301#endif
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4302
4303/**
4304 * ksize - get the actual amount of memory allocated for a given object
4305 * @objp: Pointer to the object
4306 *
4307 * kmalloc may internally round up allocations and return more memory
4308 * than requested. ksize() can be used to determine the actual amount of
4309 * memory allocated. The caller may use this additional memory, even though
4310 * a smaller amount of memory was initially specified with the kmalloc call.
4311 * The caller must guarantee that objp points to a valid object previously
4312 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4313 * must not be freed during the duration of the call.
4314 */
4315size_t ksize(const void *objp)
4316{
4317	size_t size;
4318
4319	BUG_ON(!objp);
4320	if (unlikely(objp == ZERO_SIZE_PTR))
4321		return 0;
4322
4323	size = virt_to_cache(objp)->object_size;
4324	/* We assume that ksize callers could use the whole allocated area,
4325	 * so we need to unpoison this area.
4326	 */
4327	kasan_krealloc(objp, size, GFP_NOWAIT);
4328
4329	return size;
4330}
4331EXPORT_SYMBOL(ksize);
v4.17
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * linux/mm/slab.c
   4 * Written by Mark Hemment, 1996/97.
   5 * (markhe@nextd.demon.co.uk)
   6 *
   7 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
   8 *
   9 * Major cleanup, different bufctl logic, per-cpu arrays
  10 *	(c) 2000 Manfred Spraul
  11 *
  12 * Cleanup, make the head arrays unconditional, preparation for NUMA
  13 * 	(c) 2002 Manfred Spraul
  14 *
  15 * An implementation of the Slab Allocator as described in outline in;
  16 *	UNIX Internals: The New Frontiers by Uresh Vahalia
  17 *	Pub: Prentice Hall	ISBN 0-13-101908-2
  18 * or with a little more detail in;
  19 *	The Slab Allocator: An Object-Caching Kernel Memory Allocator
  20 *	Jeff Bonwick (Sun Microsystems).
  21 *	Presented at: USENIX Summer 1994 Technical Conference
  22 *
  23 * The memory is organized in caches, one cache for each object type.
  24 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
  25 * Each cache consists out of many slabs (they are small (usually one
  26 * page long) and always contiguous), and each slab contains multiple
  27 * initialized objects.
  28 *
  29 * This means, that your constructor is used only for newly allocated
  30 * slabs and you must pass objects with the same initializations to
  31 * kmem_cache_free.
  32 *
  33 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
  34 * normal). If you need a special memory type, then must create a new
  35 * cache for that memory type.
  36 *
  37 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
  38 *   full slabs with 0 free objects
  39 *   partial slabs
  40 *   empty slabs with no allocated objects
  41 *
  42 * If partial slabs exist, then new allocations come from these slabs,
  43 * otherwise from empty slabs or new slabs are allocated.
  44 *
  45 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
  46 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
  47 *
  48 * Each cache has a short per-cpu head array, most allocs
  49 * and frees go into that array, and if that array overflows, then 1/2
  50 * of the entries in the array are given back into the global cache.
  51 * The head array is strictly LIFO and should improve the cache hit rates.
  52 * On SMP, it additionally reduces the spinlock operations.
  53 *
  54 * The c_cpuarray may not be read with enabled local interrupts -
  55 * it's changed with a smp_call_function().
  56 *
  57 * SMP synchronization:
  58 *  constructors and destructors are called without any locking.
  59 *  Several members in struct kmem_cache and struct slab never change, they
  60 *	are accessed without any locking.
  61 *  The per-cpu arrays are never accessed from the wrong cpu, no locking,
  62 *  	and local interrupts are disabled so slab code is preempt-safe.
  63 *  The non-constant members are protected with a per-cache irq spinlock.
  64 *
  65 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
  66 * in 2000 - many ideas in the current implementation are derived from
  67 * his patch.
  68 *
  69 * Further notes from the original documentation:
  70 *
  71 * 11 April '97.  Started multi-threading - markhe
  72 *	The global cache-chain is protected by the mutex 'slab_mutex'.
  73 *	The sem is only needed when accessing/extending the cache-chain, which
  74 *	can never happen inside an interrupt (kmem_cache_create(),
  75 *	kmem_cache_shrink() and kmem_cache_reap()).
  76 *
  77 *	At present, each engine can be growing a cache.  This should be blocked.
  78 *
  79 * 15 March 2005. NUMA slab allocator.
  80 *	Shai Fultheim <shai@scalex86.org>.
  81 *	Shobhit Dayal <shobhit@calsoftinc.com>
  82 *	Alok N Kataria <alokk@calsoftinc.com>
  83 *	Christoph Lameter <christoph@lameter.com>
  84 *
  85 *	Modified the slab allocator to be node aware on NUMA systems.
  86 *	Each node has its own list of partial, free and full slabs.
  87 *	All object allocations for a node occur from node specific slab lists.
  88 */
  89
  90#include	<linux/__KEEPIDENTS__B.h>
  91#include	<linux/__KEEPIDENTS__C.h>
  92#include	<linux/__KEEPIDENTS__D.h>
  93#include	<linux/__KEEPIDENTS__E.h>
  94#include	<linux/__KEEPIDENTS__F.h>
  95#include	<linux/__KEEPIDENTS__G.h>
  96#include	<linux/__KEEPIDENTS__H.h>
  97#include	<linux/__KEEPIDENTS__I.h>
  98#include	<linux/__KEEPIDENTS__J.h>
  99#include	<linux/proc_fs.h>
 100#include	<linux/__KEEPIDENTS__BA.h>
 101#include	<linux/__KEEPIDENTS__BB.h>
 102#include	<linux/__KEEPIDENTS__BC.h>
 103#include	<linux/cpu.h>
 104#include	<linux/__KEEPIDENTS__BD.h>
 105#include	<linux/__KEEPIDENTS__BE.h>
 106#include	<linux/rcupdate.h>
 107#include	<linux/__KEEPIDENTS__BF.h>
 108#include	<linux/__KEEPIDENTS__BG.h>
 109#include	<linux/__KEEPIDENTS__BH.h>
 110#include	<linux/kmemleak.h>
 111#include	<linux/__KEEPIDENTS__BI.h>
 112#include	<linux/__KEEPIDENTS__BJ.h>
 113#include	<linux/__KEEPIDENTS__CA-__KEEPIDENTS__CB.h>
 114#include	<linux/__KEEPIDENTS__CC.h>
 115#include	<linux/reciprocal_div.h>
 116#include	<linux/debugobjects.h>
 
 117#include	<linux/__KEEPIDENTS__CD.h>
 118#include	<linux/__KEEPIDENTS__CE.h>
 119#include	<linux/__KEEPIDENTS__CF/task_stack.h>
 120
 121#include	<net/__KEEPIDENTS__CG.h>
 122
 123#include	<asm/cacheflush.h>
 124#include	<asm/tlbflush.h>
 125#include	<asm/page.h>
 126
 127#include <trace/events/kmem.h>
 128
 129#include	"internal.h"
 130
 131#include	"slab.h"
 132
 133/*
 134 * DEBUG	- 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
 135 *		  0 for faster, smaller code (especially in the critical paths).
 136 *
 137 * STATS	- 1 to collect stats for /proc/slabinfo.
 138 *		  0 for faster, smaller code (especially in the critical paths).
 139 *
 140 * FORCED_DEBUG	- 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
 141 */
 142
 143#ifdef CONFIG_DEBUG_SLAB
 144#define	DEBUG		1
 145#define	STATS		1
 146#define	FORCED_DEBUG	1
 147#else
 148#define	DEBUG		0
 149#define	STATS		0
 150#define	FORCED_DEBUG	0
 151#endif
 152
 153/* Shouldn't this be in a header file somewhere? */
 154#define	BYTES_PER_WORD		sizeof(void *)
 155#define	REDZONE_ALIGN		max(BYTES_PER_WORD, __alignof__(unsigned long long))
 156
 157#ifndef ARCH_KMALLOC_FLAGS
 158#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
 159#endif
 160
 161#define FREELIST_BYTE_INDEX (((PAGE_SIZE >> BITS_PER_BYTE) \
 162				<= SLAB_OBJ_MIN_SIZE) ? 1 : 0)
 163
 164#if FREELIST_BYTE_INDEX
 165typedef unsigned char freelist_idx_t;
 166#else
 167typedef unsigned short freelist_idx_t;
 168#endif
 169
 170#define SLAB_OBJ_MAX_NUM ((1 << sizeof(freelist_idx_t) * BITS_PER_BYTE) - 1)
 171
 172/*
 173 * struct array_cache
 174 *
 175 * Purpose:
 176 * - LIFO ordering, to hand out cache-warm objects from _alloc
 177 * - reduce the number of linked list operations
 178 * - reduce spinlock operations
 179 *
 180 * The limit is stored in the per-cpu structure to reduce the data cache
 181 * footprint.
 182 *
 183 */
 184struct array_cache {
 185	unsigned int avail;
 186	unsigned int limit;
 187	unsigned int batchcount;
 188	unsigned int touched;
 189	void *entry[];	/*
 190			 * Must have this definition in here for the proper
 191			 * alignment of array_cache. Also simplifies accessing
 192			 * the entries.
 193			 */
 194};
 195
 196struct alien_cache {
 197	spinlock_t lock;
 198	struct array_cache ac;
 199};
 200
 201/*
 202 * Need this for bootstrapping a per node allocator.
 203 */
 204#define NUM_INIT_LISTS (2 * MAX_NUMNODES)
 205static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
 206#define	CACHE_CACHE 0
 207#define	SIZE_NODE (MAX_NUMNODES)
 208
 209static int drain_freelist(struct kmem_cache *cache,
 210			struct kmem_cache_node *n, int tofree);
 211static void free_block(struct kmem_cache *cachep, void **objpp, int len,
 212			int node, struct list_head *list);
 213static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list);
 214static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
 215static void cache_reap(struct work_struct *unused);
 216
 217static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
 218						void **list);
 219static inline void fixup_slab_list(struct kmem_cache *cachep,
 220				struct kmem_cache_node *n, struct page *page,
 221				void **list);
 222static int slab_early_init = 1;
 223
 224#define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
 225
 226static void kmem_cache_node_init(struct kmem_cache_node *parent)
 227{
 228	INIT_LIST_HEAD(&parent->slabs_full);
 229	INIT_LIST_HEAD(&parent->slabs_partial);
 230	INIT_LIST_HEAD(&parent->slabs_free);
 231	parent->total_slabs = 0;
 232	parent->free_slabs = 0;
 233	parent->shared = NULL;
 234	parent->alien = NULL;
 235	parent->colour_next = 0;
 236	spin_lock_init(&parent->list_lock);
 237	parent->free_objects = 0;
 238	parent->free_touched = 0;
 239}
 240
 241#define MAKE_LIST(cachep, listp, slab, nodeid)				\
 242	do {								\
 243		INIT_LIST_HEAD(listp);					\
 244		list_splice(&get_node(cachep, nodeid)->slab, listp);	\
 245	} while (0)
 246
 247#define	MAKE_ALL_LISTS(cachep, ptr, nodeid)				\
 248	do {								\
 249	MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid);	\
 250	MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
 251	MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid);	\
 252	} while (0)
 253
 254#define CFLGS_OBJFREELIST_SLAB	((slab_flags_t __force)0x40000000U)
 255#define CFLGS_OFF_SLAB		((slab_flags_t __force)0x80000000U)
 256#define	OBJFREELIST_SLAB(x)	((x)->flags & CFLGS_OBJFREELIST_SLAB)
 257#define	OFF_SLAB(x)	((x)->flags & CFLGS_OFF_SLAB)
 258
 259#define BATCHREFILL_LIMIT	16
 260/*
 261 * Optimization question: fewer reaps means less probability for unnessary
 262 * cpucache drain/refill cycles.
 263 *
 264 * OTOH the cpuarrays can contain lots of objects,
 265 * which could lock up otherwise freeable slabs.
 266 */
 267#define REAPTIMEOUT_AC		(2*HZ)
 268#define REAPTIMEOUT_NODE	(4*HZ)
 269
 270#if STATS
 271#define	STATS_INC_ACTIVE(x)	((x)->num_active++)
 272#define	STATS_DEC_ACTIVE(x)	((x)->num_active--)
 273#define	STATS_INC_ALLOCED(x)	((x)->num_allocations++)
 274#define	STATS_INC_GROWN(x)	((x)->grown++)
 275#define	STATS_ADD_REAPED(x,y)	((x)->reaped += (y))
 276#define	STATS_SET_HIGH(x)						\
 277	do {								\
 278		if ((x)->num_active > (x)->high_mark)			\
 279			(x)->high_mark = (x)->num_active;		\
 280	} while (0)
 281#define	STATS_INC_ERR(x)	((x)->errors++)
 282#define	STATS_INC_NODEALLOCS(x)	((x)->node_allocs++)
 283#define	STATS_INC_NODEFREES(x)	((x)->node_frees++)
 284#define STATS_INC_ACOVERFLOW(x)   ((x)->node_overflow++)
 285#define	STATS_SET_FREEABLE(x, i)					\
 286	do {								\
 287		if ((x)->max_freeable < i)				\
 288			(x)->max_freeable = i;				\
 289	} while (0)
 290#define STATS_INC_ALLOCHIT(x)	atomic_inc(&(x)->allochit)
 291#define STATS_INC_ALLOCMISS(x)	atomic_inc(&(x)->allocmiss)
 292#define STATS_INC_FREEHIT(x)	atomic_inc(&(x)->freehit)
 293#define STATS_INC_FREEMISS(x)	atomic_inc(&(x)->freemiss)
 294#else
 295#define	STATS_INC_ACTIVE(x)	do { } while (0)
 296#define	STATS_DEC_ACTIVE(x)	do { } while (0)
 297#define	STATS_INC_ALLOCED(x)	do { } while (0)
 298#define	STATS_INC_GROWN(x)	do { } while (0)
 299#define	STATS_ADD_REAPED(x,y)	do { (void)(y); } while (0)
 300#define	STATS_SET_HIGH(x)	do { } while (0)
 301#define	STATS_INC_ERR(x)	do { } while (0)
 302#define	STATS_INC_NODEALLOCS(x)	do { } while (0)
 303#define	STATS_INC_NODEFREES(x)	do { } while (0)
 304#define STATS_INC_ACOVERFLOW(x)   do { } while (0)
 305#define	STATS_SET_FREEABLE(x, i) do { } while (0)
 306#define STATS_INC_ALLOCHIT(x)	do { } while (0)
 307#define STATS_INC_ALLOCMISS(x)	do { } while (0)
 308#define STATS_INC_FREEHIT(x)	do { } while (0)
 309#define STATS_INC_FREEMISS(x)	do { } while (0)
 310#endif
 311
 312#if DEBUG
 313
 314/*
 315 * memory layout of objects:
 316 * 0		: objp
 317 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
 318 * 		the end of an object is aligned with the end of the real
 319 * 		allocation. Catches writes behind the end of the allocation.
 320 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
 321 * 		redzone word.
 322 * cachep->obj_offset: The real object.
 323 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
 324 * cachep->size - 1* BYTES_PER_WORD: last caller address
 325 *					[BYTES_PER_WORD long]
 326 */
 327static int obj_offset(struct kmem_cache *cachep)
 328{
 329	return cachep->obj_offset;
 330}
 331
 332static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
 333{
 334	BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
 335	return (unsigned long long*) (objp + obj_offset(cachep) -
 336				      sizeof(unsigned long long));
 337}
 338
 339static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
 340{
 341	BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
 342	if (cachep->flags & SLAB_STORE_USER)
 343		return (unsigned long long *)(objp + cachep->size -
 344					      sizeof(unsigned long long) -
 345					      REDZONE_ALIGN);
 346	return (unsigned long long *) (objp + cachep->size -
 347				       sizeof(unsigned long long));
 348}
 349
 350static void **dbg_userword(struct kmem_cache *cachep, void *objp)
 351{
 352	BUG_ON(!(cachep->flags & SLAB_STORE_USER));
 353	return (void **)(objp + cachep->size - BYTES_PER_WORD);
 354}
 355
 356#else
 357
 358#define obj_offset(x)			0
 359#define dbg_redzone1(cachep, objp)	({BUG(); (unsigned long long *)NULL;})
 360#define dbg_redzone2(cachep, objp)	({BUG(); (unsigned long long *)NULL;})
 361#define dbg_userword(cachep, objp)	({BUG(); (void **)NULL;})
 362
 363#endif
 364
 365#ifdef CONFIG_DEBUG_SLAB_LEAK
 366
 367static inline bool is_store_user_clean(struct kmem_cache *cachep)
 368{
 369	return atomic_read(&cachep->store_user_clean) == 1;
 370}
 371
 372static inline void set_store_user_clean(struct kmem_cache *cachep)
 373{
 374	atomic_set(&cachep->store_user_clean, 1);
 375}
 376
 377static inline void set_store_user_dirty(struct kmem_cache *cachep)
 378{
 379	if (is_store_user_clean(cachep))
 380		atomic_set(&cachep->store_user_clean, 0);
 381}
 382
 383#else
 384static inline void set_store_user_dirty(struct kmem_cache *cachep) {}
 385
 386#endif
 387
 388/*
 389 * Do not go above this order unless 0 objects fit into the slab or
 390 * overridden on the command line.
 391 */
 392#define	SLAB_MAX_ORDER_HI	1
 393#define	SLAB_MAX_ORDER_LO	0
 394static int slab_max_order = SLAB_MAX_ORDER_LO;
 395static bool slab_max_order_set __initdata;
 396
 397static inline struct kmem_cache *virt_to_cache(const void *obj)
 398{
 399	struct page *page = virt_to_head_page(obj);
 400	return page->slab_cache;
 401}
 402
 403static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
 404				 unsigned int idx)
 405{
 406	return page->s_mem + cache->size * idx;
 407}
 408
 409/*
 410 * We want to avoid an expensive divide : (offset / cache->size)
 411 *   Using the fact that size is a constant for a particular cache,
 412 *   we can replace (offset / cache->size) by
 413 *   reciprocal_divide(offset, cache->reciprocal_buffer_size)
 414 */
 415static inline unsigned int obj_to_index(const struct kmem_cache *cache,
 416					const struct page *page, void *obj)
 417{
 418	u32 offset = (obj - page->s_mem);
 419	return reciprocal_divide(offset, cache->reciprocal_buffer_size);
 420}
 421
 422#define BOOT_CPUCACHE_ENTRIES	1
 423/* internal cache of cache description objs */
 424static struct kmem_cache kmem_cache_boot = {
 425	.batchcount = 1,
 426	.limit = BOOT_CPUCACHE_ENTRIES,
 427	.shared = 1,
 428	.size = sizeof(struct kmem_cache),
 429	.name = "kmem_cache",
 430};
 431
 
 
 432static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
 433
 434static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
 435{
 436	return this_cpu_ptr(cachep->cpu_cache);
 437}
 438
 439/*
 440 * Calculate the number of objects and left-over bytes for a given buffer size.
 441 */
 442static unsigned int cache_estimate(unsigned long gfporder, size_t buffer_size,
 443		slab_flags_t flags, size_t *left_over)
 444{
 445	unsigned int num;
 446	size_t slab_size = PAGE_SIZE << gfporder;
 447
 448	/*
 449	 * The slab management structure can be either off the slab or
 450	 * on it. For the latter case, the memory allocated for a
 451	 * slab is used for:
 452	 *
 453	 * - @buffer_size bytes for each object
 454	 * - One freelist_idx_t for each object
 455	 *
 456	 * We don't need to consider alignment of freelist because
 457	 * freelist will be at the end of slab page. The objects will be
 458	 * at the correct alignment.
 459	 *
 460	 * If the slab management structure is off the slab, then the
 461	 * alignment will already be calculated into the size. Because
 462	 * the slabs are all pages aligned, the objects will be at the
 463	 * correct alignment when allocated.
 464	 */
 465	if (flags & (CFLGS_OBJFREELIST_SLAB | CFLGS_OFF_SLAB)) {
 466		num = slab_size / buffer_size;
 467		*left_over = slab_size % buffer_size;
 468	} else {
 469		num = slab_size / (buffer_size + sizeof(freelist_idx_t));
 470		*left_over = slab_size %
 471			(buffer_size + sizeof(freelist_idx_t));
 472	}
 473
 474	return num;
 475}
 476
 477#if DEBUG
 478#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
 479
 480static void __slab_error(const char *function, struct kmem_cache *cachep,
 481			char *msg)
 482{
 483	pr_err("slab error in %s(): cache `%s': %s\n",
 484	       function, cachep->name, msg);
 485	dump_stack();
 486	add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
 487}
 488#endif
 489
 490/*
 491 * By default on NUMA we use alien caches to stage the freeing of
 492 * objects allocated from other nodes. This causes massive memory
 493 * inefficiencies when using fake NUMA setup to split memory into a
 494 * large number of small nodes, so it can be disabled on the command
 495 * line
 496  */
 497
 498static int use_alien_caches __read_mostly = 1;
 499static int __init noaliencache_setup(char *s)
 500{
 501	use_alien_caches = 0;
 502	return 1;
 503}
 504__setup("noaliencache", noaliencache_setup);
 505
 506static int __init slab_max_order_setup(char *str)
 507{
 508	get_option(&str, &slab_max_order);
 509	slab_max_order = slab_max_order < 0 ? 0 :
 510				min(slab_max_order, MAX_ORDER - 1);
 511	slab_max_order_set = true;
 512
 513	return 1;
 514}
 515__setup("slab_max_order=", slab_max_order_setup);
 516
 517#ifdef CONFIG_NUMA
 518/*
 519 * Special reaping functions for NUMA systems called from cache_reap().
 520 * These take care of doing round robin flushing of alien caches (containing
 521 * objects freed on different nodes from which they were allocated) and the
 522 * flushing of remote pcps by calling drain_node_pages.
 523 */
 524static DEFINE_PER_CPU(unsigned long, slab_reap_node);
 525
 526static void init_reap_node(int cpu)
 527{
 528	per_cpu(slab_reap_node, cpu) = next_node_in(cpu_to_mem(cpu),
 529						    node_online_map);
 
 
 
 
 
 530}
 531
 532static void next_reap_node(void)
 533{
 534	int node = __this_cpu_read(slab_reap_node);
 535
 536	node = next_node_in(node, node_online_map);
 
 
 537	__this_cpu_write(slab_reap_node, node);
 538}
 539
 540#else
 541#define init_reap_node(cpu) do { } while (0)
 542#define next_reap_node(void) do { } while (0)
 543#endif
 544
 545/*
 546 * Initiate the reap timer running on the target CPU.  We run at around 1 to 2Hz
 547 * via the workqueue/eventd.
 548 * Add the CPU number into the expiration time to minimize the possibility of
 549 * the CPUs getting into lockstep and contending for the global cache chain
 550 * lock.
 551 */
 552static void start_cpu_timer(int cpu)
 553{
 554	struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
 555
 556	if (reap_work->work.func == NULL) {
 
 
 
 
 
 557		init_reap_node(cpu);
 558		INIT_DEFERRABLE_WORK(reap_work, cache_reap);
 559		schedule_delayed_work_on(cpu, reap_work,
 560					__round_jiffies_relative(HZ, cpu));
 561	}
 562}
 563
 564static void init_arraycache(struct array_cache *ac, int limit, int batch)
 565{
 566	/*
 567	 * The array_cache structures contain pointers to free object.
 568	 * However, when such objects are allocated or transferred to another
 569	 * cache the pointers are not cleared and they could be counted as
 570	 * valid references during a kmemleak scan. Therefore, kmemleak must
 571	 * not scan such objects.
 572	 */
 573	kmemleak_no_scan(ac);
 574	if (ac) {
 575		ac->avail = 0;
 576		ac->limit = limit;
 577		ac->batchcount = batch;
 578		ac->touched = 0;
 579	}
 580}
 581
 582static struct array_cache *alloc_arraycache(int node, int entries,
 583					    int batchcount, gfp_t gfp)
 584{
 585	size_t memsize = sizeof(void *) * entries + sizeof(struct array_cache);
 586	struct array_cache *ac = NULL;
 587
 588	ac = kmalloc_node(memsize, gfp, node);
 589	init_arraycache(ac, entries, batchcount);
 590	return ac;
 591}
 592
 593static noinline void cache_free_pfmemalloc(struct kmem_cache *cachep,
 594					struct page *page, void *objp)
 595{
 596	struct kmem_cache_node *n;
 597	int page_node;
 598	LIST_HEAD(list);
 599
 600	page_node = page_to_nid(page);
 601	n = get_node(cachep, page_node);
 602
 603	spin_lock(&n->list_lock);
 604	free_block(cachep, &objp, 1, page_node, &list);
 605	spin_unlock(&n->list_lock);
 606
 607	slabs_destroy(cachep, &list);
 608}
 609
 610/*
 611 * Transfer objects in one arraycache to another.
 612 * Locking must be handled by the caller.
 613 *
 614 * Return the number of entries transferred.
 615 */
 616static int transfer_objects(struct array_cache *to,
 617		struct array_cache *from, unsigned int max)
 618{
 619	/* Figure out how many entries to transfer */
 620	int nr = min3(from->avail, max, to->limit - to->avail);
 621
 622	if (!nr)
 623		return 0;
 624
 625	memcpy(to->entry + to->avail, from->entry + from->avail -nr,
 626			sizeof(void *) *nr);
 627
 628	from->avail -= nr;
 629	to->avail += nr;
 630	return nr;
 631}
 632
 633#ifndef CONFIG_NUMA
 634
 635#define drain_alien_cache(cachep, alien) do { } while (0)
 636#define reap_alien(cachep, n) do { } while (0)
 637
 638static inline struct alien_cache **alloc_alien_cache(int node,
 639						int limit, gfp_t gfp)
 640{
 641	return NULL;
 642}
 643
 644static inline void free_alien_cache(struct alien_cache **ac_ptr)
 645{
 646}
 647
 648static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
 649{
 650	return 0;
 651}
 652
 653static inline void *alternate_node_alloc(struct kmem_cache *cachep,
 654		gfp_t flags)
 655{
 656	return NULL;
 657}
 658
 659static inline void *____cache_alloc_node(struct kmem_cache *cachep,
 660		 gfp_t flags, int nodeid)
 661{
 662	return NULL;
 663}
 664
 665static inline gfp_t gfp_exact_node(gfp_t flags)
 666{
 667	return flags & ~__GFP_NOFAIL;
 668}
 669
 670#else	/* CONFIG_NUMA */
 671
 672static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
 673static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
 674
 675static struct alien_cache *__alloc_alien_cache(int node, int entries,
 676						int batch, gfp_t gfp)
 677{
 678	size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache);
 679	struct alien_cache *alc = NULL;
 680
 681	alc = kmalloc_node(memsize, gfp, node);
 682	init_arraycache(&alc->ac, entries, batch);
 683	spin_lock_init(&alc->lock);
 684	return alc;
 685}
 686
 687static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
 688{
 689	struct alien_cache **alc_ptr;
 690	size_t memsize = sizeof(void *) * nr_node_ids;
 691	int i;
 692
 693	if (limit > 1)
 694		limit = 12;
 695	alc_ptr = kzalloc_node(memsize, gfp, node);
 696	if (!alc_ptr)
 697		return NULL;
 698
 699	for_each_node(i) {
 700		if (i == node || !node_online(i))
 701			continue;
 702		alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp);
 703		if (!alc_ptr[i]) {
 704			for (i--; i >= 0; i--)
 705				kfree(alc_ptr[i]);
 706			kfree(alc_ptr);
 707			return NULL;
 708		}
 709	}
 710	return alc_ptr;
 711}
 712
 713static void free_alien_cache(struct alien_cache **alc_ptr)
 714{
 715	int i;
 716
 717	if (!alc_ptr)
 718		return;
 719	for_each_node(i)
 720	    kfree(alc_ptr[i]);
 721	kfree(alc_ptr);
 722}
 723
 724static void __drain_alien_cache(struct kmem_cache *cachep,
 725				struct array_cache *ac, int node,
 726				struct list_head *list)
 727{
 728	struct kmem_cache_node *n = get_node(cachep, node);
 729
 730	if (ac->avail) {
 731		spin_lock(&n->list_lock);
 732		/*
 733		 * Stuff objects into the remote nodes shared array first.
 734		 * That way we could avoid the overhead of putting the objects
 735		 * into the free lists and getting them back later.
 736		 */
 737		if (n->shared)
 738			transfer_objects(n->shared, ac, ac->limit);
 739
 740		free_block(cachep, ac->entry, ac->avail, node, list);
 741		ac->avail = 0;
 742		spin_unlock(&n->list_lock);
 743	}
 744}
 745
 746/*
 747 * Called from cache_reap() to regularly drain alien caches round robin.
 748 */
 749static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
 750{
 751	int node = __this_cpu_read(slab_reap_node);
 752
 753	if (n->alien) {
 754		struct alien_cache *alc = n->alien[node];
 755		struct array_cache *ac;
 756
 757		if (alc) {
 758			ac = &alc->ac;
 759			if (ac->avail && spin_trylock_irq(&alc->lock)) {
 760				LIST_HEAD(list);
 761
 762				__drain_alien_cache(cachep, ac, node, &list);
 763				spin_unlock_irq(&alc->lock);
 764				slabs_destroy(cachep, &list);
 765			}
 766		}
 767	}
 768}
 769
 770static void drain_alien_cache(struct kmem_cache *cachep,
 771				struct alien_cache **alien)
 772{
 773	int i = 0;
 774	struct alien_cache *alc;
 775	struct array_cache *ac;
 776	unsigned long flags;
 777
 778	for_each_online_node(i) {
 779		alc = alien[i];
 780		if (alc) {
 781			LIST_HEAD(list);
 782
 783			ac = &alc->ac;
 784			spin_lock_irqsave(&alc->lock, flags);
 785			__drain_alien_cache(cachep, ac, i, &list);
 786			spin_unlock_irqrestore(&alc->lock, flags);
 787			slabs_destroy(cachep, &list);
 788		}
 789	}
 790}
 791
 792static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
 793				int node, int page_node)
 794{
 795	struct kmem_cache_node *n;
 796	struct alien_cache *alien = NULL;
 797	struct array_cache *ac;
 798	LIST_HEAD(list);
 799
 800	n = get_node(cachep, node);
 801	STATS_INC_NODEFREES(cachep);
 802	if (n->alien && n->alien[page_node]) {
 803		alien = n->alien[page_node];
 804		ac = &alien->ac;
 805		spin_lock(&alien->lock);
 806		if (unlikely(ac->avail == ac->limit)) {
 807			STATS_INC_ACOVERFLOW(cachep);
 808			__drain_alien_cache(cachep, ac, page_node, &list);
 809		}
 810		ac->entry[ac->avail++] = objp;
 811		spin_unlock(&alien->lock);
 812		slabs_destroy(cachep, &list);
 813	} else {
 814		n = get_node(cachep, page_node);
 815		spin_lock(&n->list_lock);
 816		free_block(cachep, &objp, 1, page_node, &list);
 817		spin_unlock(&n->list_lock);
 818		slabs_destroy(cachep, &list);
 819	}
 820	return 1;
 821}
 822
 823static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
 824{
 825	int page_node = page_to_nid(virt_to_page(objp));
 826	int node = numa_mem_id();
 827	/*
 828	 * Make sure we are not freeing a object from another node to the array
 829	 * cache on this cpu.
 830	 */
 831	if (likely(node == page_node))
 832		return 0;
 833
 834	return __cache_free_alien(cachep, objp, node, page_node);
 835}
 836
 837/*
 838 * Construct gfp mask to allocate from a specific node but do not reclaim or
 839 * warn about failures.
 840 */
 841static inline gfp_t gfp_exact_node(gfp_t flags)
 842{
 843	return (flags | __GFP_THISNODE | __GFP_NOWARN) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
 844}
 845#endif
 846
 847static int init_cache_node(struct kmem_cache *cachep, int node, gfp_t gfp)
 848{
 849	struct kmem_cache_node *n;
 850
 851	/*
 852	 * Set up the kmem_cache_node for cpu before we can
 853	 * begin anything. Make sure some other cpu on this
 854	 * node has not already allocated this
 855	 */
 856	n = get_node(cachep, node);
 857	if (n) {
 858		spin_lock_irq(&n->list_lock);
 859		n->free_limit = (1 + nr_cpus_node(node)) * cachep->batchcount +
 860				cachep->num;
 861		spin_unlock_irq(&n->list_lock);
 862
 863		return 0;
 864	}
 865
 866	n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
 867	if (!n)
 868		return -ENOMEM;
 869
 870	kmem_cache_node_init(n);
 871	n->next_reap = jiffies + REAPTIMEOUT_NODE +
 872		    ((unsigned long)cachep) % REAPTIMEOUT_NODE;
 873
 874	n->free_limit =
 875		(1 + nr_cpus_node(node)) * cachep->batchcount + cachep->num;
 876
 877	/*
 878	 * The kmem_cache_nodes don't come and go as CPUs
 879	 * come and go.  slab_mutex is sufficient
 880	 * protection here.
 881	 */
 882	cachep->node[node] = n;
 883
 884	return 0;
 885}
 886
 887#if (defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)) || defined(CONFIG_SMP)
 888/*
 889 * Allocates and initializes node for a node on each slab cache, used for
 890 * either memory or cpu hotplug.  If memory is being hot-added, the kmem_cache_node
 891 * will be allocated off-node since memory is not yet online for the new node.
 892 * When hotplugging memory or a cpu, existing node are not replaced if
 893 * already in use.
 894 *
 895 * Must hold slab_mutex.
 896 */
 897static int init_cache_node_node(int node)
 898{
 899	int ret;
 900	struct kmem_cache *cachep;
 
 
 901
 902	list_for_each_entry(cachep, &slab_caches, list) {
 903		ret = init_cache_node(cachep, node, GFP_KERNEL);
 904		if (ret)
 905			return ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 906	}
 907
 908	return 0;
 909}
 910#endif
 911
 912static int setup_kmem_cache_node(struct kmem_cache *cachep,
 913				int node, gfp_t gfp, bool force_change)
 914{
 915	int ret = -ENOMEM;
 916	struct kmem_cache_node *n;
 917	struct array_cache *old_shared = NULL;
 918	struct array_cache *new_shared = NULL;
 919	struct alien_cache **new_alien = NULL;
 920	LIST_HEAD(list);
 921
 922	if (use_alien_caches) {
 923		new_alien = alloc_alien_cache(node, cachep->limit, gfp);
 924		if (!new_alien)
 925			goto fail;
 926	}
 927
 928	if (cachep->shared) {
 929		new_shared = alloc_arraycache(node,
 930			cachep->shared * cachep->batchcount, 0xbaadf00d, gfp);
 931		if (!new_shared)
 932			goto fail;
 933	}
 934
 935	ret = init_cache_node(cachep, node, gfp);
 936	if (ret)
 937		goto fail;
 938
 939	n = get_node(cachep, node);
 940	spin_lock_irq(&n->list_lock);
 941	if (n->shared && force_change) {
 942		free_block(cachep, n->shared->entry,
 943				n->shared->avail, node, &list);
 944		n->shared->avail = 0;
 945	}
 946
 947	if (!n->shared || force_change) {
 948		old_shared = n->shared;
 949		n->shared = new_shared;
 950		new_shared = NULL;
 951	}
 952
 953	if (!n->alien) {
 954		n->alien = new_alien;
 955		new_alien = NULL;
 956	}
 957
 958	spin_unlock_irq(&n->list_lock);
 959	slabs_destroy(cachep, &list);
 960
 961	/*
 962	 * To protect lockless access to n->shared during irq disabled context.
 963	 * If n->shared isn't NULL in irq disabled context, accessing to it is
 964	 * guaranteed to be valid until irq is re-enabled, because it will be
 965	 * freed after synchronize_sched().
 966	 */
 967	if (old_shared && force_change)
 968		synchronize_sched();
 969
 970fail:
 971	kfree(old_shared);
 972	kfree(new_shared);
 973	free_alien_cache(new_alien);
 974
 975	return ret;
 976}
 977
 978#ifdef CONFIG_SMP
 979
 980static void cpuup_canceled(long cpu)
 981{
 982	struct kmem_cache *cachep;
 983	struct kmem_cache_node *n = NULL;
 984	int node = cpu_to_mem(cpu);
 985	const struct cpumask *mask = cpumask_of_node(node);
 986
 987	list_for_each_entry(cachep, &slab_caches, list) {
 988		struct array_cache *nc;
 989		struct array_cache *shared;
 990		struct alien_cache **alien;
 991		LIST_HEAD(list);
 992
 993		n = get_node(cachep, node);
 994		if (!n)
 995			continue;
 996
 997		spin_lock_irq(&n->list_lock);
 998
 999		/* Free limit for this kmem_cache_node */
1000		n->free_limit -= cachep->batchcount;
1001
1002		/* cpu is dead; no one can alloc from it. */
1003		nc = per_cpu_ptr(cachep->cpu_cache, cpu);
1004		if (nc) {
1005			free_block(cachep, nc->entry, nc->avail, node, &list);
1006			nc->avail = 0;
1007		}
1008
1009		if (!cpumask_empty(mask)) {
1010			spin_unlock_irq(&n->list_lock);
1011			goto free_slab;
1012		}
1013
1014		shared = n->shared;
1015		if (shared) {
1016			free_block(cachep, shared->entry,
1017				   shared->avail, node, &list);
1018			n->shared = NULL;
1019		}
1020
1021		alien = n->alien;
1022		n->alien = NULL;
1023
1024		spin_unlock_irq(&n->list_lock);
1025
1026		kfree(shared);
1027		if (alien) {
1028			drain_alien_cache(cachep, alien);
1029			free_alien_cache(alien);
1030		}
1031
1032free_slab:
1033		slabs_destroy(cachep, &list);
1034	}
1035	/*
1036	 * In the previous loop, all the objects were freed to
1037	 * the respective cache's slabs,  now we can go ahead and
1038	 * shrink each nodelist to its limit.
1039	 */
1040	list_for_each_entry(cachep, &slab_caches, list) {
1041		n = get_node(cachep, node);
1042		if (!n)
1043			continue;
1044		drain_freelist(cachep, n, INT_MAX);
1045	}
1046}
1047
1048static int cpuup_prepare(long cpu)
1049{
1050	struct kmem_cache *cachep;
 
1051	int node = cpu_to_mem(cpu);
1052	int err;
1053
1054	/*
1055	 * We need to do this right in the beginning since
1056	 * alloc_arraycache's are going to use this list.
1057	 * kmalloc_node allows us to add the slab to the right
1058	 * kmem_cache_node and not this cpu's kmem_cache_node
1059	 */
1060	err = init_cache_node_node(node);
1061	if (err < 0)
1062		goto bad;
1063
1064	/*
1065	 * Now we can go ahead with allocating the shared arrays and
1066	 * array caches
1067	 */
1068	list_for_each_entry(cachep, &slab_caches, list) {
1069		err = setup_kmem_cache_node(cachep, node, GFP_KERNEL, false);
1070		if (err)
1071			goto bad;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1072	}
1073
1074	return 0;
1075bad:
1076	cpuup_canceled(cpu);
1077	return -ENOMEM;
1078}
1079
1080int slab_prepare_cpu(unsigned int cpu)
 
1081{
1082	int err;
 
1083
1084	mutex_lock(&slab_mutex);
1085	err = cpuup_prepare(cpu);
1086	mutex_unlock(&slab_mutex);
1087	return err;
1088}
1089
1090/*
1091 * This is called for a failed online attempt and for a successful
1092 * offline.
1093 *
1094 * Even if all the cpus of a node are down, we don't free the
1095 * kmem_list3 of any cache. This to avoid a race between cpu_down, and
1096 * a kmalloc allocation from another cpu for memory from the node of
1097 * the cpu going down.  The list3 structure is usually allocated from
1098 * kmem_cache_create() and gets destroyed at kmem_cache_destroy().
1099 */
1100int slab_dead_cpu(unsigned int cpu)
1101{
1102	mutex_lock(&slab_mutex);
1103	cpuup_canceled(cpu);
1104	mutex_unlock(&slab_mutex);
1105	return 0;
1106}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1107#endif
1108
1109static int slab_online_cpu(unsigned int cpu)
1110{
1111	start_cpu_timer(cpu);
1112	return 0;
 
 
 
1113}
1114
1115static int slab_offline_cpu(unsigned int cpu)
1116{
1117	/*
1118	 * Shutdown cache reaper. Note that the slab_mutex is held so
1119	 * that if cache_reap() is invoked it cannot do anything
1120	 * expensive but will only modify reap_work and reschedule the
1121	 * timer.
1122	 */
1123	cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1124	/* Now the cache_reaper is guaranteed to be not running. */
1125	per_cpu(slab_reap_work, cpu).work.func = NULL;
1126	return 0;
1127}
1128
1129#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1130/*
1131 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1132 * Returns -EBUSY if all objects cannot be drained so that the node is not
1133 * removed.
1134 *
1135 * Must hold slab_mutex.
1136 */
1137static int __meminit drain_cache_node_node(int node)
1138{
1139	struct kmem_cache *cachep;
1140	int ret = 0;
1141
1142	list_for_each_entry(cachep, &slab_caches, list) {
1143		struct kmem_cache_node *n;
1144
1145		n = get_node(cachep, node);
1146		if (!n)
1147			continue;
1148
1149		drain_freelist(cachep, n, INT_MAX);
1150
1151		if (!list_empty(&n->slabs_full) ||
1152		    !list_empty(&n->slabs_partial)) {
1153			ret = -EBUSY;
1154			break;
1155		}
1156	}
1157	return ret;
1158}
1159
1160static int __meminit slab_memory_callback(struct notifier_block *self,
1161					unsigned long action, void *arg)
1162{
1163	struct memory_notify *mnb = arg;
1164	int ret = 0;
1165	int nid;
1166
1167	nid = mnb->status_change_nid;
1168	if (nid < 0)
1169		goto out;
1170
1171	switch (action) {
1172	case MEM_GOING_ONLINE:
1173		mutex_lock(&slab_mutex);
1174		ret = init_cache_node_node(nid);
1175		mutex_unlock(&slab_mutex);
1176		break;
1177	case MEM_GOING_OFFLINE:
1178		mutex_lock(&slab_mutex);
1179		ret = drain_cache_node_node(nid);
1180		mutex_unlock(&slab_mutex);
1181		break;
1182	case MEM_ONLINE:
1183	case MEM_OFFLINE:
1184	case MEM_CANCEL_ONLINE:
1185	case MEM_CANCEL_OFFLINE:
1186		break;
1187	}
1188out:
1189	return notifier_from_errno(ret);
1190}
1191#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1192
1193/*
1194 * swap the static kmem_cache_node with kmalloced memory
1195 */
1196static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1197				int nodeid)
1198{
1199	struct kmem_cache_node *ptr;
1200
1201	ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1202	BUG_ON(!ptr);
1203
1204	memcpy(ptr, list, sizeof(struct kmem_cache_node));
1205	/*
1206	 * Do not assume that spinlocks can be initialized via memcpy:
1207	 */
1208	spin_lock_init(&ptr->list_lock);
1209
1210	MAKE_ALL_LISTS(cachep, ptr, nodeid);
1211	cachep->node[nodeid] = ptr;
1212}
1213
1214/*
1215 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1216 * size of kmem_cache_node.
1217 */
1218static void __init set_up_node(struct kmem_cache *cachep, int index)
1219{
1220	int node;
1221
1222	for_each_online_node(node) {
1223		cachep->node[node] = &init_kmem_cache_node[index + node];
1224		cachep->node[node]->next_reap = jiffies +
1225		    REAPTIMEOUT_NODE +
1226		    ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1227	}
1228}
1229
1230/*
1231 * Initialisation.  Called after the page allocator have been initialised and
1232 * before smp_init().
1233 */
1234void __init kmem_cache_init(void)
1235{
1236	int i;
1237
1238	BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1239					sizeof(struct rcu_head));
1240	kmem_cache = &kmem_cache_boot;
1241
1242	if (!IS_ENABLED(CONFIG_NUMA) || num_possible_nodes() == 1)
1243		use_alien_caches = 0;
1244
1245	for (i = 0; i < NUM_INIT_LISTS; i++)
1246		kmem_cache_node_init(&init_kmem_cache_node[i]);
1247
1248	/*
1249	 * Fragmentation resistance on low memory - only use bigger
1250	 * page orders on machines with more than 32MB of memory if
1251	 * not overridden on the command line.
1252	 */
1253	if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1254		slab_max_order = SLAB_MAX_ORDER_HI;
1255
1256	/* Bootstrap is tricky, because several objects are allocated
1257	 * from caches that do not exist yet:
1258	 * 1) initialize the kmem_cache cache: it contains the struct
1259	 *    kmem_cache structures of all caches, except kmem_cache itself:
1260	 *    kmem_cache is statically allocated.
1261	 *    Initially an __init data area is used for the head array and the
1262	 *    kmem_cache_node structures, it's replaced with a kmalloc allocated
1263	 *    array at the end of the bootstrap.
1264	 * 2) Create the first kmalloc cache.
1265	 *    The struct kmem_cache for the new cache is allocated normally.
1266	 *    An __init data area is used for the head array.
1267	 * 3) Create the remaining kmalloc caches, with minimally sized
1268	 *    head arrays.
1269	 * 4) Replace the __init data head arrays for kmem_cache and the first
1270	 *    kmalloc cache with kmalloc allocated arrays.
1271	 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1272	 *    the other cache's with kmalloc allocated memory.
1273	 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1274	 */
1275
1276	/* 1) create the kmem_cache */
1277
1278	/*
1279	 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1280	 */
1281	create_boot_cache(kmem_cache, "kmem_cache",
1282		offsetof(struct kmem_cache, node) +
1283				  nr_node_ids * sizeof(struct kmem_cache_node *),
1284				  SLAB_HWCACHE_ALIGN, 0, 0);
1285	list_add(&kmem_cache->list, &slab_caches);
1286	memcg_link_cache(kmem_cache);
1287	slab_state = PARTIAL;
1288
1289	/*
1290	 * Initialize the caches that provide memory for the  kmem_cache_node
1291	 * structures first.  Without this, further allocations will bug.
1292	 */
1293	kmalloc_caches[INDEX_NODE] = create_kmalloc_cache(
1294				kmalloc_info[INDEX_NODE].name,
1295				kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS,
1296				0, kmalloc_size(INDEX_NODE));
1297	slab_state = PARTIAL_NODE;
1298	setup_kmalloc_cache_index_table();
1299
1300	slab_early_init = 0;
1301
1302	/* 5) Replace the bootstrap kmem_cache_node */
1303	{
1304		int nid;
1305
1306		for_each_online_node(nid) {
1307			init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
1308
1309			init_list(kmalloc_caches[INDEX_NODE],
1310					  &init_kmem_cache_node[SIZE_NODE + nid], nid);
1311		}
1312	}
1313
1314	create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
1315}
1316
1317void __init kmem_cache_init_late(void)
1318{
1319	struct kmem_cache *cachep;
1320
 
 
1321	/* 6) resize the head arrays to their final sizes */
1322	mutex_lock(&slab_mutex);
1323	list_for_each_entry(cachep, &slab_caches, list)
1324		if (enable_cpucache(cachep, GFP_NOWAIT))
1325			BUG();
1326	mutex_unlock(&slab_mutex);
1327
1328	/* Done! */
1329	slab_state = FULL;
1330
 
 
 
 
 
 
1331#ifdef CONFIG_NUMA
1332	/*
1333	 * Register a memory hotplug callback that initializes and frees
1334	 * node.
1335	 */
1336	hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1337#endif
1338
1339	/*
1340	 * The reap timers are started later, with a module init call: That part
1341	 * of the kernel is not yet operational.
1342	 */
1343}
1344
1345static int __init cpucache_init(void)
1346{
1347	int ret;
1348
1349	/*
1350	 * Register the timers that return unneeded pages to the page allocator
1351	 */
1352	ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "SLAB online",
1353				slab_online_cpu, slab_offline_cpu);
1354	WARN_ON(ret < 0);
1355
 
 
1356	return 0;
1357}
1358__initcall(cpucache_init);
1359
1360static noinline void
1361slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1362{
1363#if DEBUG
1364	struct kmem_cache_node *n;
 
1365	unsigned long flags;
1366	int node;
1367	static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1368				      DEFAULT_RATELIMIT_BURST);
1369
1370	if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1371		return;
1372
1373	pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1374		nodeid, gfpflags, &gfpflags);
1375	pr_warn("  cache: %s, object size: %d, order: %d\n",
1376		cachep->name, cachep->size, cachep->gfporder);
1377
1378	for_each_kmem_cache_node(cachep, node, n) {
1379		unsigned long total_slabs, free_slabs, free_objs;
 
1380
1381		spin_lock_irqsave(&n->list_lock, flags);
1382		total_slabs = n->total_slabs;
1383		free_slabs = n->free_slabs;
1384		free_objs = n->free_objects;
 
 
 
 
 
 
 
 
 
1385		spin_unlock_irqrestore(&n->list_lock, flags);
1386
1387		pr_warn("  node %d: slabs: %ld/%ld, objs: %ld/%ld\n",
1388			node, total_slabs - free_slabs, total_slabs,
1389			(total_slabs * cachep->num) - free_objs,
1390			total_slabs * cachep->num);
 
1391	}
1392#endif
1393}
1394
1395/*
1396 * Interface to system's page allocator. No need to hold the
1397 * kmem_cache_node ->list_lock.
1398 *
1399 * If we requested dmaable memory, we will get it. Even if we
1400 * did not request dmaable memory, we might get it, but that
1401 * would be relatively rare and ignorable.
1402 */
1403static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1404								int nodeid)
1405{
1406	struct page *page;
1407	int nr_pages;
1408
1409	flags |= cachep->allocflags;
 
 
1410
1411	page = __alloc_pages_node(nodeid, flags, cachep->gfporder);
1412	if (!page) {
1413		slab_out_of_memory(cachep, flags, nodeid);
1414		return NULL;
1415	}
1416
1417	if (memcg_charge_slab(page, flags, cachep->gfporder, cachep)) {
1418		__free_pages(page, cachep->gfporder);
1419		return NULL;
1420	}
1421
1422	nr_pages = (1 << cachep->gfporder);
1423	if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1424		mod_lruvec_page_state(page, NR_SLAB_RECLAIMABLE, nr_pages);
 
1425	else
1426		mod_lruvec_page_state(page, NR_SLAB_UNRECLAIMABLE, nr_pages);
 
1427
1428	__SetPageSlab(page);
1429	/* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1430	if (sk_memalloc_socks() && page_is_pfmemalloc(page))
1431		SetPageSlabPfmemalloc(page);
1432
 
 
 
 
 
 
 
 
 
1433	return page;
1434}
1435
1436/*
1437 * Interface to system's page release.
1438 */
1439static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
1440{
1441	int order = cachep->gfporder;
1442	unsigned long nr_freed = (1 << order);
1443
 
 
1444	if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1445		mod_lruvec_page_state(page, NR_SLAB_RECLAIMABLE, -nr_freed);
 
1446	else
1447		mod_lruvec_page_state(page, NR_SLAB_UNRECLAIMABLE, -nr_freed);
 
1448
1449	BUG_ON(!PageSlab(page));
1450	__ClearPageSlabPfmemalloc(page);
1451	__ClearPageSlab(page);
1452	page_mapcount_reset(page);
1453	page->mapping = NULL;
1454
1455	if (current->reclaim_state)
1456		current->reclaim_state->reclaimed_slab += nr_freed;
1457	memcg_uncharge_slab(page, order, cachep);
1458	__free_pages(page, order);
1459}
1460
1461static void kmem_rcu_free(struct rcu_head *head)
1462{
1463	struct kmem_cache *cachep;
1464	struct page *page;
1465
1466	page = container_of(head, struct page, rcu_head);
1467	cachep = page->slab_cache;
1468
1469	kmem_freepages(cachep, page);
1470}
1471
1472#if DEBUG
1473static bool is_debug_pagealloc_cache(struct kmem_cache *cachep)
1474{
1475	if (debug_pagealloc_enabled() && OFF_SLAB(cachep) &&
1476		(cachep->size % PAGE_SIZE) == 0)
1477		return true;
1478
1479	return false;
1480}
1481
1482#ifdef CONFIG_DEBUG_PAGEALLOC
1483static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1484			    unsigned long caller)
1485{
1486	int size = cachep->object_size;
1487
1488	addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1489
1490	if (size < 5 * sizeof(unsigned long))
1491		return;
1492
1493	*addr++ = 0x12345678;
1494	*addr++ = caller;
1495	*addr++ = smp_processor_id();
1496	size -= 3 * sizeof(unsigned long);
1497	{
1498		unsigned long *sptr = &caller;
1499		unsigned long svalue;
1500
1501		while (!kstack_end(sptr)) {
1502			svalue = *sptr++;
1503			if (kernel_text_address(svalue)) {
1504				*addr++ = svalue;
1505				size -= sizeof(unsigned long);
1506				if (size <= sizeof(unsigned long))
1507					break;
1508			}
1509		}
1510
1511	}
1512	*addr++ = 0x87654321;
1513}
1514
1515static void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1516				int map, unsigned long caller)
1517{
1518	if (!is_debug_pagealloc_cache(cachep))
1519		return;
1520
1521	if (caller)
1522		store_stackinfo(cachep, objp, caller);
1523
1524	kernel_map_pages(virt_to_page(objp), cachep->size / PAGE_SIZE, map);
1525}
1526
1527#else
1528static inline void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1529				int map, unsigned long caller) {}
1530
1531#endif
1532
1533static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1534{
1535	int size = cachep->object_size;
1536	addr = &((char *)addr)[obj_offset(cachep)];
1537
1538	memset(addr, val, size);
1539	*(unsigned char *)(addr + size - 1) = POISON_END;
1540}
1541
1542static void dump_line(char *data, int offset, int limit)
1543{
1544	int i;
1545	unsigned char error = 0;
1546	int bad_count = 0;
1547
1548	pr_err("%03x: ", offset);
1549	for (i = 0; i < limit; i++) {
1550		if (data[offset + i] != POISON_FREE) {
1551			error = data[offset + i];
1552			bad_count++;
1553		}
1554	}
1555	print_hex_dump(KERN_CONT, "", 0, 16, 1,
1556			&data[offset], limit, 1);
1557
1558	if (bad_count == 1) {
1559		error ^= POISON_FREE;
1560		if (!(error & (error - 1))) {
1561			pr_err("Single bit error detected. Probably bad RAM.\n");
1562#ifdef CONFIG_X86
1563			pr_err("Run memtest86+ or a similar memory test tool.\n");
1564#else
1565			pr_err("Run a memory test tool.\n");
1566#endif
1567		}
1568	}
1569}
1570#endif
1571
1572#if DEBUG
1573
1574static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1575{
1576	int i, size;
1577	char *realobj;
1578
1579	if (cachep->flags & SLAB_RED_ZONE) {
1580		pr_err("Redzone: 0x%llx/0x%llx\n",
1581		       *dbg_redzone1(cachep, objp),
1582		       *dbg_redzone2(cachep, objp));
1583	}
1584
1585	if (cachep->flags & SLAB_STORE_USER)
1586		pr_err("Last user: (%pSR)\n", *dbg_userword(cachep, objp));
 
 
 
1587	realobj = (char *)objp + obj_offset(cachep);
1588	size = cachep->object_size;
1589	for (i = 0; i < size && lines; i += 16, lines--) {
1590		int limit;
1591		limit = 16;
1592		if (i + limit > size)
1593			limit = size - i;
1594		dump_line(realobj, i, limit);
1595	}
1596}
1597
1598static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1599{
1600	char *realobj;
1601	int size, i;
1602	int lines = 0;
1603
1604	if (is_debug_pagealloc_cache(cachep))
1605		return;
1606
1607	realobj = (char *)objp + obj_offset(cachep);
1608	size = cachep->object_size;
1609
1610	for (i = 0; i < size; i++) {
1611		char exp = POISON_FREE;
1612		if (i == size - 1)
1613			exp = POISON_END;
1614		if (realobj[i] != exp) {
1615			int limit;
1616			/* Mismatch ! */
1617			/* Print header */
1618			if (lines == 0) {
1619				pr_err("Slab corruption (%s): %s start=%px, len=%d\n",
1620				       print_tainted(), cachep->name,
1621				       realobj, size);
1622				print_objinfo(cachep, objp, 0);
1623			}
1624			/* Hexdump the affected line */
1625			i = (i / 16) * 16;
1626			limit = 16;
1627			if (i + limit > size)
1628				limit = size - i;
1629			dump_line(realobj, i, limit);
1630			i += 16;
1631			lines++;
1632			/* Limit to 5 lines */
1633			if (lines > 5)
1634				break;
1635		}
1636	}
1637	if (lines != 0) {
1638		/* Print some data about the neighboring objects, if they
1639		 * exist:
1640		 */
1641		struct page *page = virt_to_head_page(objp);
1642		unsigned int objnr;
1643
1644		objnr = obj_to_index(cachep, page, objp);
1645		if (objnr) {
1646			objp = index_to_obj(cachep, page, objnr - 1);
1647			realobj = (char *)objp + obj_offset(cachep);
1648			pr_err("Prev obj: start=%px, len=%d\n", realobj, size);
1649			print_objinfo(cachep, objp, 2);
1650		}
1651		if (objnr + 1 < cachep->num) {
1652			objp = index_to_obj(cachep, page, objnr + 1);
1653			realobj = (char *)objp + obj_offset(cachep);
1654			pr_err("Next obj: start=%px, len=%d\n", realobj, size);
1655			print_objinfo(cachep, objp, 2);
1656		}
1657	}
1658}
1659#endif
1660
1661#if DEBUG
1662static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1663						struct page *page)
1664{
1665	int i;
1666
1667	if (OBJFREELIST_SLAB(cachep) && cachep->flags & SLAB_POISON) {
1668		poison_obj(cachep, page->freelist - obj_offset(cachep),
1669			POISON_FREE);
1670	}
1671
1672	for (i = 0; i < cachep->num; i++) {
1673		void *objp = index_to_obj(cachep, page, i);
1674
1675		if (cachep->flags & SLAB_POISON) {
1676			check_poison_obj(cachep, objp);
1677			slab_kernel_map(cachep, objp, 1, 0);
1678		}
1679		if (cachep->flags & SLAB_RED_ZONE) {
1680			if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1681				slab_error(cachep, "start of a freed object was overwritten");
1682			if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1683				slab_error(cachep, "end of a freed object was overwritten");
1684		}
1685	}
1686}
1687#else
1688static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1689						struct page *page)
1690{
1691}
1692#endif
1693
1694/**
1695 * slab_destroy - destroy and release all objects in a slab
1696 * @cachep: cache pointer being destroyed
1697 * @page: page pointer being destroyed
1698 *
1699 * Destroy all the objs in a slab page, and release the mem back to the system.
1700 * Before calling the slab page must have been unlinked from the cache. The
1701 * kmem_cache_node ->list_lock is not held/needed.
1702 */
1703static void slab_destroy(struct kmem_cache *cachep, struct page *page)
1704{
1705	void *freelist;
1706
1707	freelist = page->freelist;
1708	slab_destroy_debugcheck(cachep, page);
1709	if (unlikely(cachep->flags & SLAB_TYPESAFE_BY_RCU))
1710		call_rcu(&page->rcu_head, kmem_rcu_free);
1711	else
1712		kmem_freepages(cachep, page);
1713
1714	/*
1715	 * From now on, we don't use freelist
1716	 * although actual page can be freed in rcu context
1717	 */
1718	if (OFF_SLAB(cachep))
1719		kmem_cache_free(cachep->freelist_cache, freelist);
1720}
1721
1722static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
1723{
1724	struct page *page, *n;
1725
1726	list_for_each_entry_safe(page, n, list, lru) {
1727		list_del(&page->lru);
1728		slab_destroy(cachep, page);
1729	}
1730}
1731
1732/**
1733 * calculate_slab_order - calculate size (page order) of slabs
1734 * @cachep: pointer to the cache that is being created
1735 * @size: size of objects to be created in this cache.
1736 * @flags: slab allocation flags
1737 *
1738 * Also calculates the number of objects per slab.
1739 *
1740 * This could be made much more intelligent.  For now, try to avoid using
1741 * high order pages for slabs.  When the gfp() functions are more friendly
1742 * towards high-order requests, this should be changed.
1743 */
1744static size_t calculate_slab_order(struct kmem_cache *cachep,
1745				size_t size, slab_flags_t flags)
1746{
1747	size_t left_over = 0;
1748	int gfporder;
1749
1750	for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
1751		unsigned int num;
1752		size_t remainder;
1753
1754		num = cache_estimate(gfporder, size, flags, &remainder);
1755		if (!num)
1756			continue;
1757
1758		/* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1759		if (num > SLAB_OBJ_MAX_NUM)
1760			break;
1761
1762		if (flags & CFLGS_OFF_SLAB) {
1763			struct kmem_cache *freelist_cache;
1764			size_t freelist_size;
1765
1766			freelist_size = num * sizeof(freelist_idx_t);
1767			freelist_cache = kmalloc_slab(freelist_size, 0u);
1768			if (!freelist_cache)
1769				continue;
1770
1771			/*
1772			 * Needed to avoid possible looping condition
1773			 * in cache_grow_begin()
1774			 */
1775			if (OFF_SLAB(freelist_cache))
1776				continue;
1777
1778			/* check if off slab has enough benefit */
1779			if (freelist_cache->size > cachep->size / 2)
1780				continue;
1781		}
1782
1783		/* Found something acceptable - save it away */
1784		cachep->num = num;
1785		cachep->gfporder = gfporder;
1786		left_over = remainder;
1787
1788		/*
1789		 * A VFS-reclaimable slab tends to have most allocations
1790		 * as GFP_NOFS and we really don't want to have to be allocating
1791		 * higher-order pages when we are unable to shrink dcache.
1792		 */
1793		if (flags & SLAB_RECLAIM_ACCOUNT)
1794			break;
1795
1796		/*
1797		 * Large number of objects is good, but very large slabs are
1798		 * currently bad for the gfp()s.
1799		 */
1800		if (gfporder >= slab_max_order)
1801			break;
1802
1803		/*
1804		 * Acceptable internal fragmentation?
1805		 */
1806		if (left_over * 8 <= (PAGE_SIZE << gfporder))
1807			break;
1808	}
1809	return left_over;
1810}
1811
1812static struct array_cache __percpu *alloc_kmem_cache_cpus(
1813		struct kmem_cache *cachep, int entries, int batchcount)
1814{
1815	int cpu;
1816	size_t size;
1817	struct array_cache __percpu *cpu_cache;
1818
1819	size = sizeof(void *) * entries + sizeof(struct array_cache);
1820	cpu_cache = __alloc_percpu(size, sizeof(void *));
1821
1822	if (!cpu_cache)
1823		return NULL;
1824
1825	for_each_possible_cpu(cpu) {
1826		init_arraycache(per_cpu_ptr(cpu_cache, cpu),
1827				entries, batchcount);
1828	}
1829
1830	return cpu_cache;
1831}
1832
1833static int __ref setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
1834{
1835	if (slab_state >= FULL)
1836		return enable_cpucache(cachep, gfp);
1837
1838	cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1);
1839	if (!cachep->cpu_cache)
1840		return 1;
1841
1842	if (slab_state == DOWN) {
1843		/* Creation of first cache (kmem_cache). */
1844		set_up_node(kmem_cache, CACHE_CACHE);
1845	} else if (slab_state == PARTIAL) {
1846		/* For kmem_cache_node */
1847		set_up_node(cachep, SIZE_NODE);
1848	} else {
1849		int node;
1850
1851		for_each_online_node(node) {
1852			cachep->node[node] = kmalloc_node(
1853				sizeof(struct kmem_cache_node), gfp, node);
1854			BUG_ON(!cachep->node[node]);
1855			kmem_cache_node_init(cachep->node[node]);
1856		}
1857	}
1858
1859	cachep->node[numa_mem_id()]->next_reap =
1860			jiffies + REAPTIMEOUT_NODE +
1861			((unsigned long)cachep) % REAPTIMEOUT_NODE;
1862
1863	cpu_cache_get(cachep)->avail = 0;
1864	cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1865	cpu_cache_get(cachep)->batchcount = 1;
1866	cpu_cache_get(cachep)->touched = 0;
1867	cachep->batchcount = 1;
1868	cachep->limit = BOOT_CPUCACHE_ENTRIES;
1869	return 0;
1870}
1871
1872slab_flags_t kmem_cache_flags(unsigned int object_size,
1873	slab_flags_t flags, const char *name,
1874	void (*ctor)(void *))
1875{
1876	return flags;
1877}
1878
1879struct kmem_cache *
1880__kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
1881		   slab_flags_t flags, void (*ctor)(void *))
1882{
1883	struct kmem_cache *cachep;
1884
1885	cachep = find_mergeable(size, align, flags, name, ctor);
1886	if (cachep) {
1887		cachep->refcount++;
1888
1889		/*
1890		 * Adjust the object sizes so that we clear
1891		 * the complete object on kzalloc.
1892		 */
1893		cachep->object_size = max_t(int, cachep->object_size, size);
1894	}
1895	return cachep;
1896}
1897
1898static bool set_objfreelist_slab_cache(struct kmem_cache *cachep,
1899			size_t size, slab_flags_t flags)
1900{
1901	size_t left;
1902
1903	cachep->num = 0;
1904
1905	if (cachep->ctor || flags & SLAB_TYPESAFE_BY_RCU)
1906		return false;
1907
1908	left = calculate_slab_order(cachep, size,
1909			flags | CFLGS_OBJFREELIST_SLAB);
1910	if (!cachep->num)
1911		return false;
1912
1913	if (cachep->num * sizeof(freelist_idx_t) > cachep->object_size)
1914		return false;
1915
1916	cachep->colour = left / cachep->colour_off;
1917
1918	return true;
1919}
1920
1921static bool set_off_slab_cache(struct kmem_cache *cachep,
1922			size_t size, slab_flags_t flags)
1923{
1924	size_t left;
1925
1926	cachep->num = 0;
1927
1928	/*
1929	 * Always use on-slab management when SLAB_NOLEAKTRACE
1930	 * to avoid recursive calls into kmemleak.
1931	 */
1932	if (flags & SLAB_NOLEAKTRACE)
1933		return false;
1934
1935	/*
1936	 * Size is large, assume best to place the slab management obj
1937	 * off-slab (should allow better packing of objs).
1938	 */
1939	left = calculate_slab_order(cachep, size, flags | CFLGS_OFF_SLAB);
1940	if (!cachep->num)
1941		return false;
1942
1943	/*
1944	 * If the slab has been placed off-slab, and we have enough space then
1945	 * move it on-slab. This is at the expense of any extra colouring.
1946	 */
1947	if (left >= cachep->num * sizeof(freelist_idx_t))
1948		return false;
1949
1950	cachep->colour = left / cachep->colour_off;
1951
1952	return true;
1953}
1954
1955static bool set_on_slab_cache(struct kmem_cache *cachep,
1956			size_t size, slab_flags_t flags)
1957{
1958	size_t left;
1959
1960	cachep->num = 0;
1961
1962	left = calculate_slab_order(cachep, size, flags);
1963	if (!cachep->num)
1964		return false;
1965
1966	cachep->colour = left / cachep->colour_off;
1967
1968	return true;
1969}
1970
1971/**
1972 * __kmem_cache_create - Create a cache.
1973 * @cachep: cache management descriptor
1974 * @flags: SLAB flags
1975 *
1976 * Returns a ptr to the cache on success, NULL on failure.
1977 * Cannot be called within a int, but can be interrupted.
1978 * The @ctor is run when new pages are allocated by the cache.
1979 *
1980 * The flags are
1981 *
1982 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1983 * to catch references to uninitialised memory.
1984 *
1985 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1986 * for buffer overruns.
1987 *
1988 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1989 * cacheline.  This can be beneficial if you're counting cycles as closely
1990 * as davem.
1991 */
1992int __kmem_cache_create(struct kmem_cache *cachep, slab_flags_t flags)
 
1993{
1994	size_t ralign = BYTES_PER_WORD;
1995	gfp_t gfp;
1996	int err;
1997	unsigned int size = cachep->size;
1998
1999#if DEBUG
2000#if FORCED_DEBUG
2001	/*
2002	 * Enable redzoning and last user accounting, except for caches with
2003	 * large objects, if the increased size would increase the object size
2004	 * above the next power of two: caches with object sizes just above a
2005	 * power of two have a significant amount of internal fragmentation.
2006	 */
2007	if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2008						2 * sizeof(unsigned long long)))
2009		flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2010	if (!(flags & SLAB_TYPESAFE_BY_RCU))
2011		flags |= SLAB_POISON;
2012#endif
2013#endif
2014
2015	/*
2016	 * Check that size is in terms of words.  This is needed to avoid
2017	 * unaligned accesses for some archs when redzoning is used, and makes
2018	 * sure any on-slab bufctl's are also correctly aligned.
2019	 */
2020	size = ALIGN(size, BYTES_PER_WORD);
 
 
 
2021
2022	if (flags & SLAB_RED_ZONE) {
2023		ralign = REDZONE_ALIGN;
2024		/* If redzoning, ensure that the second redzone is suitably
2025		 * aligned, by adjusting the object size accordingly. */
2026		size = ALIGN(size, REDZONE_ALIGN);
 
2027	}
2028
2029	/* 3) caller mandated alignment */
2030	if (ralign < cachep->align) {
2031		ralign = cachep->align;
2032	}
2033	/* disable debug if necessary */
2034	if (ralign > __alignof__(unsigned long long))
2035		flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2036	/*
2037	 * 4) Store it.
2038	 */
2039	cachep->align = ralign;
2040	cachep->colour_off = cache_line_size();
2041	/* Offset must be a multiple of the alignment. */
2042	if (cachep->colour_off < cachep->align)
2043		cachep->colour_off = cachep->align;
2044
2045	if (slab_is_available())
2046		gfp = GFP_KERNEL;
2047	else
2048		gfp = GFP_NOWAIT;
2049
2050#if DEBUG
2051
2052	/*
2053	 * Both debugging options require word-alignment which is calculated
2054	 * into align above.
2055	 */
2056	if (flags & SLAB_RED_ZONE) {
2057		/* add space for red zone words */
2058		cachep->obj_offset += sizeof(unsigned long long);
2059		size += 2 * sizeof(unsigned long long);
2060	}
2061	if (flags & SLAB_STORE_USER) {
2062		/* user store requires one word storage behind the end of
2063		 * the real object. But if the second red zone needs to be
2064		 * aligned to 64 bits, we must allow that much space.
2065		 */
2066		if (flags & SLAB_RED_ZONE)
2067			size += REDZONE_ALIGN;
2068		else
2069			size += BYTES_PER_WORD;
2070	}
2071#endif
2072
2073	kasan_cache_create(cachep, &size, &flags);
2074
2075	size = ALIGN(size, cachep->align);
2076	/*
2077	 * We should restrict the number of objects in a slab to implement
2078	 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2079	 */
2080	if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2081		size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
2082
2083#if DEBUG
2084	/*
2085	 * To activate debug pagealloc, off-slab management is necessary
2086	 * requirement. In early phase of initialization, small sized slab
2087	 * doesn't get initialized so it would not be possible. So, we need
2088	 * to check size >= 256. It guarantees that all necessary small
2089	 * sized slab is initialized in current slab initialization sequence.
2090	 */
2091	if (debug_pagealloc_enabled() && (flags & SLAB_POISON) &&
2092		size >= 256 && cachep->object_size > cache_line_size()) {
2093		if (size < PAGE_SIZE || size % PAGE_SIZE == 0) {
2094			size_t tmp_size = ALIGN(size, PAGE_SIZE);
2095
2096			if (set_off_slab_cache(cachep, tmp_size, flags)) {
2097				flags |= CFLGS_OFF_SLAB;
2098				cachep->obj_offset += tmp_size - size;
2099				size = tmp_size;
2100				goto done;
2101			}
2102		}
2103	}
2104#endif
2105
2106	if (set_objfreelist_slab_cache(cachep, size, flags)) {
2107		flags |= CFLGS_OBJFREELIST_SLAB;
2108		goto done;
2109	}
2110
2111	if (set_off_slab_cache(cachep, size, flags)) {
2112		flags |= CFLGS_OFF_SLAB;
2113		goto done;
2114	}
2115
2116	if (set_on_slab_cache(cachep, size, flags))
2117		goto done;
2118
2119	return -E2BIG;
2120
2121done:
2122	cachep->freelist_size = cachep->num * sizeof(freelist_idx_t);
2123	cachep->flags = flags;
2124	cachep->allocflags = __GFP_COMP;
2125	if (flags & SLAB_CACHE_DMA)
2126		cachep->allocflags |= GFP_DMA;
2127	if (flags & SLAB_RECLAIM_ACCOUNT)
2128		cachep->allocflags |= __GFP_RECLAIMABLE;
2129	cachep->size = size;
2130	cachep->reciprocal_buffer_size = reciprocal_value(size);
2131
2132#if DEBUG
2133	/*
2134	 * If we're going to use the generic kernel_map_pages()
2135	 * poisoning, then it's going to smash the contents of
2136	 * the redzone and userword anyhow, so switch them off.
2137	 */
2138	if (IS_ENABLED(CONFIG_PAGE_POISONING) &&
2139		(cachep->flags & SLAB_POISON) &&
2140		is_debug_pagealloc_cache(cachep))
2141		cachep->flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2142#endif
2143
2144	if (OFF_SLAB(cachep)) {
2145		cachep->freelist_cache =
2146			kmalloc_slab(cachep->freelist_size, 0u);
2147	}
2148
2149	err = setup_cpu_cache(cachep, gfp);
2150	if (err) {
2151		__kmem_cache_release(cachep);
2152		return err;
2153	}
2154
2155	return 0;
2156}
2157
2158#if DEBUG
2159static void check_irq_off(void)
2160{
2161	BUG_ON(!irqs_disabled());
2162}
2163
2164static void check_irq_on(void)
2165{
2166	BUG_ON(irqs_disabled());
2167}
2168
2169static void check_mutex_acquired(void)
2170{
2171	BUG_ON(!mutex_is_locked(&slab_mutex));
2172}
2173
2174static void check_spinlock_acquired(struct kmem_cache *cachep)
2175{
2176#ifdef CONFIG_SMP
2177	check_irq_off();
2178	assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
2179#endif
2180}
2181
2182static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2183{
2184#ifdef CONFIG_SMP
2185	check_irq_off();
2186	assert_spin_locked(&get_node(cachep, node)->list_lock);
2187#endif
2188}
2189
2190#else
2191#define check_irq_off()	do { } while(0)
2192#define check_irq_on()	do { } while(0)
2193#define check_mutex_acquired()	do { } while(0)
2194#define check_spinlock_acquired(x) do { } while(0)
2195#define check_spinlock_acquired_node(x, y) do { } while(0)
2196#endif
2197
2198static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac,
2199				int node, bool free_all, struct list_head *list)
2200{
2201	int tofree;
2202
2203	if (!ac || !ac->avail)
2204		return;
2205
2206	tofree = free_all ? ac->avail : (ac->limit + 4) / 5;
2207	if (tofree > ac->avail)
2208		tofree = (ac->avail + 1) / 2;
2209
2210	free_block(cachep, ac->entry, tofree, node, list);
2211	ac->avail -= tofree;
2212	memmove(ac->entry, &(ac->entry[tofree]), sizeof(void *) * ac->avail);
2213}
2214
2215static void do_drain(void *arg)
2216{
2217	struct kmem_cache *cachep = arg;
2218	struct array_cache *ac;
2219	int node = numa_mem_id();
2220	struct kmem_cache_node *n;
2221	LIST_HEAD(list);
2222
2223	check_irq_off();
2224	ac = cpu_cache_get(cachep);
2225	n = get_node(cachep, node);
2226	spin_lock(&n->list_lock);
2227	free_block(cachep, ac->entry, ac->avail, node, &list);
2228	spin_unlock(&n->list_lock);
2229	slabs_destroy(cachep, &list);
2230	ac->avail = 0;
2231}
2232
2233static void drain_cpu_caches(struct kmem_cache *cachep)
2234{
2235	struct kmem_cache_node *n;
2236	int node;
2237	LIST_HEAD(list);
2238
2239	on_each_cpu(do_drain, cachep, 1);
2240	check_irq_on();
2241	for_each_kmem_cache_node(cachep, node, n)
2242		if (n->alien)
2243			drain_alien_cache(cachep, n->alien);
2244
2245	for_each_kmem_cache_node(cachep, node, n) {
2246		spin_lock_irq(&n->list_lock);
2247		drain_array_locked(cachep, n->shared, node, true, &list);
2248		spin_unlock_irq(&n->list_lock);
2249
2250		slabs_destroy(cachep, &list);
2251	}
2252}
2253
2254/*
2255 * Remove slabs from the list of free slabs.
2256 * Specify the number of slabs to drain in tofree.
2257 *
2258 * Returns the actual number of slabs released.
2259 */
2260static int drain_freelist(struct kmem_cache *cache,
2261			struct kmem_cache_node *n, int tofree)
2262{
2263	struct list_head *p;
2264	int nr_freed;
2265	struct page *page;
2266
2267	nr_freed = 0;
2268	while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
2269
2270		spin_lock_irq(&n->list_lock);
2271		p = n->slabs_free.prev;
2272		if (p == &n->slabs_free) {
2273			spin_unlock_irq(&n->list_lock);
2274			goto out;
2275		}
2276
2277		page = list_entry(p, struct page, lru);
2278		list_del(&page->lru);
2279		n->free_slabs--;
2280		n->total_slabs--;
2281		/*
2282		 * Safe to drop the lock. The slab is no longer linked
2283		 * to the cache.
2284		 */
2285		n->free_objects -= cache->num;
2286		spin_unlock_irq(&n->list_lock);
2287		slab_destroy(cache, page);
2288		nr_freed++;
2289	}
2290out:
2291	return nr_freed;
2292}
2293
2294bool __kmem_cache_empty(struct kmem_cache *s)
2295{
2296	int node;
2297	struct kmem_cache_node *n;
2298
2299	for_each_kmem_cache_node(s, node, n)
2300		if (!list_empty(&n->slabs_full) ||
2301		    !list_empty(&n->slabs_partial))
2302			return false;
2303	return true;
2304}
2305
2306int __kmem_cache_shrink(struct kmem_cache *cachep)
2307{
2308	int ret = 0;
2309	int node;
2310	struct kmem_cache_node *n;
2311
2312	drain_cpu_caches(cachep);
2313
2314	check_irq_on();
2315	for_each_kmem_cache_node(cachep, node, n) {
2316		drain_freelist(cachep, n, INT_MAX);
2317
2318		ret += !list_empty(&n->slabs_full) ||
2319			!list_empty(&n->slabs_partial);
2320	}
2321	return (ret ? 1 : 0);
2322}
2323
2324#ifdef CONFIG_MEMCG
2325void __kmemcg_cache_deactivate(struct kmem_cache *cachep)
2326{
2327	__kmem_cache_shrink(cachep);
2328}
2329#endif
2330
2331int __kmem_cache_shutdown(struct kmem_cache *cachep)
2332{
2333	return __kmem_cache_shrink(cachep);
2334}
2335
2336void __kmem_cache_release(struct kmem_cache *cachep)
2337{
2338	int i;
2339	struct kmem_cache_node *n;
2340
2341	cache_random_seq_destroy(cachep);
2342
2343	free_percpu(cachep->cpu_cache);
2344
2345	/* NUMA: free the node structures */
2346	for_each_kmem_cache_node(cachep, i, n) {
2347		kfree(n->shared);
2348		free_alien_cache(n->alien);
2349		kfree(n);
2350		cachep->node[i] = NULL;
2351	}
2352}
2353
2354/*
2355 * Get the memory for a slab management obj.
2356 *
2357 * For a slab cache when the slab descriptor is off-slab, the
2358 * slab descriptor can't come from the same cache which is being created,
2359 * Because if it is the case, that means we defer the creation of
2360 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2361 * And we eventually call down to __kmem_cache_create(), which
2362 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2363 * This is a "chicken-and-egg" problem.
2364 *
2365 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2366 * which are all initialized during kmem_cache_init().
2367 */
2368static void *alloc_slabmgmt(struct kmem_cache *cachep,
2369				   struct page *page, int colour_off,
2370				   gfp_t local_flags, int nodeid)
2371{
2372	void *freelist;
2373	void *addr = page_address(page);
2374
2375	page->s_mem = addr + colour_off;
2376	page->active = 0;
2377
2378	if (OBJFREELIST_SLAB(cachep))
2379		freelist = NULL;
2380	else if (OFF_SLAB(cachep)) {
2381		/* Slab management obj is off-slab. */
2382		freelist = kmem_cache_alloc_node(cachep->freelist_cache,
2383					      local_flags, nodeid);
2384		if (!freelist)
2385			return NULL;
2386	} else {
2387		/* We will use last bytes at the slab for freelist */
2388		freelist = addr + (PAGE_SIZE << cachep->gfporder) -
2389				cachep->freelist_size;
2390	}
2391
2392	return freelist;
2393}
2394
2395static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
2396{
2397	return ((freelist_idx_t *)page->freelist)[idx];
2398}
2399
2400static inline void set_free_obj(struct page *page,
2401					unsigned int idx, freelist_idx_t val)
2402{
2403	((freelist_idx_t *)(page->freelist))[idx] = val;
2404}
2405
2406static void cache_init_objs_debug(struct kmem_cache *cachep, struct page *page)
2407{
2408#if DEBUG
2409	int i;
2410
2411	for (i = 0; i < cachep->num; i++) {
2412		void *objp = index_to_obj(cachep, page, i);
2413
2414		if (cachep->flags & SLAB_STORE_USER)
2415			*dbg_userword(cachep, objp) = NULL;
2416
2417		if (cachep->flags & SLAB_RED_ZONE) {
2418			*dbg_redzone1(cachep, objp) = RED_INACTIVE;
2419			*dbg_redzone2(cachep, objp) = RED_INACTIVE;
2420		}
2421		/*
2422		 * Constructors are not allowed to allocate memory from the same
2423		 * cache which they are a constructor for.  Otherwise, deadlock.
2424		 * They must also be threaded.
2425		 */
2426		if (cachep->ctor && !(cachep->flags & SLAB_POISON)) {
2427			kasan_unpoison_object_data(cachep,
2428						   objp + obj_offset(cachep));
2429			cachep->ctor(objp + obj_offset(cachep));
2430			kasan_poison_object_data(
2431				cachep, objp + obj_offset(cachep));
2432		}
2433
2434		if (cachep->flags & SLAB_RED_ZONE) {
2435			if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2436				slab_error(cachep, "constructor overwrote the end of an object");
2437			if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2438				slab_error(cachep, "constructor overwrote the start of an object");
2439		}
2440		/* need to poison the objs? */
2441		if (cachep->flags & SLAB_POISON) {
2442			poison_obj(cachep, objp, POISON_FREE);
2443			slab_kernel_map(cachep, objp, 0, 0);
2444		}
2445	}
2446#endif
2447}
2448
2449#ifdef CONFIG_SLAB_FREELIST_RANDOM
2450/* Hold information during a freelist initialization */
2451union freelist_init_state {
2452	struct {
2453		unsigned int pos;
2454		unsigned int *list;
2455		unsigned int count;
2456	};
2457	struct rnd_state rnd_state;
2458};
2459
2460/*
2461 * Initialize the state based on the randomization methode available.
2462 * return true if the pre-computed list is available, false otherwize.
2463 */
2464static bool freelist_state_initialize(union freelist_init_state *state,
2465				struct kmem_cache *cachep,
2466				unsigned int count)
2467{
2468	bool ret;
2469	unsigned int rand;
2470
2471	/* Use best entropy available to define a random shift */
2472	rand = get_random_int();
2473
2474	/* Use a random state if the pre-computed list is not available */
2475	if (!cachep->random_seq) {
2476		prandom_seed_state(&state->rnd_state, rand);
2477		ret = false;
2478	} else {
2479		state->list = cachep->random_seq;
2480		state->count = count;
2481		state->pos = rand % count;
2482		ret = true;
2483	}
2484	return ret;
2485}
2486
2487/* Get the next entry on the list and randomize it using a random shift */
2488static freelist_idx_t next_random_slot(union freelist_init_state *state)
2489{
2490	if (state->pos >= state->count)
2491		state->pos = 0;
2492	return state->list[state->pos++];
2493}
2494
2495/* Swap two freelist entries */
2496static void swap_free_obj(struct page *page, unsigned int a, unsigned int b)
2497{
2498	swap(((freelist_idx_t *)page->freelist)[a],
2499		((freelist_idx_t *)page->freelist)[b]);
2500}
2501
2502/*
2503 * Shuffle the freelist initialization state based on pre-computed lists.
2504 * return true if the list was successfully shuffled, false otherwise.
2505 */
2506static bool shuffle_freelist(struct kmem_cache *cachep, struct page *page)
2507{
2508	unsigned int objfreelist = 0, i, rand, count = cachep->num;
2509	union freelist_init_state state;
2510	bool precomputed;
2511
2512	if (count < 2)
2513		return false;
2514
2515	precomputed = freelist_state_initialize(&state, cachep, count);
2516
2517	/* Take a random entry as the objfreelist */
2518	if (OBJFREELIST_SLAB(cachep)) {
2519		if (!precomputed)
2520			objfreelist = count - 1;
2521		else
2522			objfreelist = next_random_slot(&state);
2523		page->freelist = index_to_obj(cachep, page, objfreelist) +
2524						obj_offset(cachep);
2525		count--;
2526	}
2527
2528	/*
2529	 * On early boot, generate the list dynamically.
2530	 * Later use a pre-computed list for speed.
2531	 */
2532	if (!precomputed) {
2533		for (i = 0; i < count; i++)
2534			set_free_obj(page, i, i);
2535
2536		/* Fisher-Yates shuffle */
2537		for (i = count - 1; i > 0; i--) {
2538			rand = prandom_u32_state(&state.rnd_state);
2539			rand %= (i + 1);
2540			swap_free_obj(page, i, rand);
2541		}
2542	} else {
2543		for (i = 0; i < count; i++)
2544			set_free_obj(page, i, next_random_slot(&state));
2545	}
2546
2547	if (OBJFREELIST_SLAB(cachep))
2548		set_free_obj(page, cachep->num - 1, objfreelist);
2549
2550	return true;
2551}
2552#else
2553static inline bool shuffle_freelist(struct kmem_cache *cachep,
2554				struct page *page)
2555{
2556	return false;
2557}
2558#endif /* CONFIG_SLAB_FREELIST_RANDOM */
2559
2560static void cache_init_objs(struct kmem_cache *cachep,
2561			    struct page *page)
2562{
2563	int i;
2564	void *objp;
2565	bool shuffled;
2566
2567	cache_init_objs_debug(cachep, page);
2568
2569	/* Try to randomize the freelist if enabled */
2570	shuffled = shuffle_freelist(cachep, page);
2571
2572	if (!shuffled && OBJFREELIST_SLAB(cachep)) {
2573		page->freelist = index_to_obj(cachep, page, cachep->num - 1) +
2574						obj_offset(cachep);
2575	}
2576
2577	for (i = 0; i < cachep->num; i++) {
2578		objp = index_to_obj(cachep, page, i);
2579		kasan_init_slab_obj(cachep, objp);
2580
2581		/* constructor could break poison info */
2582		if (DEBUG == 0 && cachep->ctor) {
 
2583			kasan_unpoison_object_data(cachep, objp);
2584			cachep->ctor(objp);
2585			kasan_poison_object_data(cachep, objp);
2586		}
2587
2588		if (!shuffled)
2589			set_free_obj(page, i, i);
 
 
 
 
 
 
 
 
 
2590	}
2591}
2592
2593static void *slab_get_obj(struct kmem_cache *cachep, struct page *page)
2594{
2595	void *objp;
2596
2597	objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
2598	page->active++;
2599
2600#if DEBUG
2601	if (cachep->flags & SLAB_STORE_USER)
2602		set_store_user_dirty(cachep);
2603#endif
2604
2605	return objp;
2606}
2607
2608static void slab_put_obj(struct kmem_cache *cachep,
2609			struct page *page, void *objp)
2610{
2611	unsigned int objnr = obj_to_index(cachep, page, objp);
2612#if DEBUG
2613	unsigned int i;
2614
2615	/* Verify double free bug */
2616	for (i = page->active; i < cachep->num; i++) {
2617		if (get_free_obj(page, i) == objnr) {
2618			pr_err("slab: double free detected in cache '%s', objp %px\n",
2619			       cachep->name, objp);
2620			BUG();
2621		}
2622	}
2623#endif
2624	page->active--;
2625	if (!page->freelist)
2626		page->freelist = objp + obj_offset(cachep);
2627
2628	set_free_obj(page, page->active, objnr);
2629}
2630
2631/*
2632 * Map pages beginning at addr to the given cache and slab. This is required
2633 * for the slab allocator to be able to lookup the cache and slab of a
2634 * virtual address for kfree, ksize, and slab debugging.
2635 */
2636static void slab_map_pages(struct kmem_cache *cache, struct page *page,
2637			   void *freelist)
2638{
2639	page->slab_cache = cache;
2640	page->freelist = freelist;
2641}
2642
2643/*
2644 * Grow (by 1) the number of slabs within a cache.  This is called by
2645 * kmem_cache_alloc() when there are no active objs left in a cache.
2646 */
2647static struct page *cache_grow_begin(struct kmem_cache *cachep,
2648				gfp_t flags, int nodeid)
2649{
2650	void *freelist;
2651	size_t offset;
2652	gfp_t local_flags;
2653	int page_node;
2654	struct kmem_cache_node *n;
2655	struct page *page;
2656
2657	/*
2658	 * Be lazy and only check for valid flags here,  keeping it out of the
2659	 * critical path in kmem_cache_alloc().
2660	 */
2661	if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
2662		gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK;
2663		flags &= ~GFP_SLAB_BUG_MASK;
2664		pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
2665				invalid_mask, &invalid_mask, flags, &flags);
2666		dump_stack();
2667	}
2668	local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2669
 
2670	check_irq_off();
 
 
 
 
 
 
 
 
 
 
 
 
2671	if (gfpflags_allow_blocking(local_flags))
2672		local_irq_enable();
2673
2674	/*
 
 
 
 
 
 
 
 
2675	 * Get mem for the objs.  Attempt to allocate a physical page from
2676	 * 'nodeid'.
2677	 */
2678	page = kmem_getpages(cachep, local_flags, nodeid);
 
2679	if (!page)
2680		goto failed;
2681
2682	page_node = page_to_nid(page);
2683	n = get_node(cachep, page_node);
2684
2685	/* Get colour for the slab, and cal the next value. */
2686	n->colour_next++;
2687	if (n->colour_next >= cachep->colour)
2688		n->colour_next = 0;
2689
2690	offset = n->colour_next;
2691	if (offset >= cachep->colour)
2692		offset = 0;
2693
2694	offset *= cachep->colour_off;
2695
2696	/* Get slab management. */
2697	freelist = alloc_slabmgmt(cachep, page, offset,
2698			local_flags & ~GFP_CONSTRAINT_MASK, page_node);
2699	if (OFF_SLAB(cachep) && !freelist)
2700		goto opps1;
2701
2702	slab_map_pages(cachep, page, freelist);
2703
2704	kasan_poison_slab(page);
2705	cache_init_objs(cachep, page);
2706
2707	if (gfpflags_allow_blocking(local_flags))
2708		local_irq_disable();
 
 
2709
2710	return page;
2711
 
 
 
 
2712opps1:
2713	kmem_freepages(cachep, page);
2714failed:
2715	if (gfpflags_allow_blocking(local_flags))
2716		local_irq_disable();
2717	return NULL;
2718}
2719
2720static void cache_grow_end(struct kmem_cache *cachep, struct page *page)
2721{
2722	struct kmem_cache_node *n;
2723	void *list = NULL;
2724
2725	check_irq_off();
2726
2727	if (!page)
2728		return;
2729
2730	INIT_LIST_HEAD(&page->lru);
2731	n = get_node(cachep, page_to_nid(page));
2732
2733	spin_lock(&n->list_lock);
2734	n->total_slabs++;
2735	if (!page->active) {
2736		list_add_tail(&page->lru, &(n->slabs_free));
2737		n->free_slabs++;
2738	} else
2739		fixup_slab_list(cachep, n, page, &list);
2740
2741	STATS_INC_GROWN(cachep);
2742	n->free_objects += cachep->num - page->active;
2743	spin_unlock(&n->list_lock);
2744
2745	fixup_objfreelist_debug(cachep, &list);
2746}
2747
2748#if DEBUG
2749
2750/*
2751 * Perform extra freeing checks:
2752 * - detect bad pointers.
2753 * - POISON/RED_ZONE checking
2754 */
2755static void kfree_debugcheck(const void *objp)
2756{
2757	if (!virt_addr_valid(objp)) {
2758		pr_err("kfree_debugcheck: out of range ptr %lxh\n",
2759		       (unsigned long)objp);
2760		BUG();
2761	}
2762}
2763
2764static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2765{
2766	unsigned long long redzone1, redzone2;
2767
2768	redzone1 = *dbg_redzone1(cache, obj);
2769	redzone2 = *dbg_redzone2(cache, obj);
2770
2771	/*
2772	 * Redzone is ok.
2773	 */
2774	if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2775		return;
2776
2777	if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2778		slab_error(cache, "double free detected");
2779	else
2780		slab_error(cache, "memory outside object was overwritten");
2781
2782	pr_err("%px: redzone 1:0x%llx, redzone 2:0x%llx\n",
2783	       obj, redzone1, redzone2);
2784}
2785
2786static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2787				   unsigned long caller)
2788{
2789	unsigned int objnr;
2790	struct page *page;
2791
2792	BUG_ON(virt_to_cache(objp) != cachep);
2793
2794	objp -= obj_offset(cachep);
2795	kfree_debugcheck(objp);
2796	page = virt_to_head_page(objp);
2797
2798	if (cachep->flags & SLAB_RED_ZONE) {
2799		verify_redzone_free(cachep, objp);
2800		*dbg_redzone1(cachep, objp) = RED_INACTIVE;
2801		*dbg_redzone2(cachep, objp) = RED_INACTIVE;
2802	}
2803	if (cachep->flags & SLAB_STORE_USER) {
2804		set_store_user_dirty(cachep);
2805		*dbg_userword(cachep, objp) = (void *)caller;
2806	}
2807
2808	objnr = obj_to_index(cachep, page, objp);
2809
2810	BUG_ON(objnr >= cachep->num);
2811	BUG_ON(objp != index_to_obj(cachep, page, objnr));
2812
2813	if (cachep->flags & SLAB_POISON) {
2814		poison_obj(cachep, objp, POISON_FREE);
2815		slab_kernel_map(cachep, objp, 0, caller);
2816	}
2817	return objp;
2818}
2819
2820#else
2821#define kfree_debugcheck(x) do { } while(0)
2822#define cache_free_debugcheck(x,objp,z) (objp)
2823#endif
2824
2825static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
2826						void **list)
2827{
2828#if DEBUG
2829	void *next = *list;
2830	void *objp;
2831
2832	while (next) {
2833		objp = next - obj_offset(cachep);
2834		next = *(void **)next;
2835		poison_obj(cachep, objp, POISON_FREE);
2836	}
2837#endif
2838}
2839
2840static inline void fixup_slab_list(struct kmem_cache *cachep,
2841				struct kmem_cache_node *n, struct page *page,
2842				void **list)
2843{
2844	/* move slabp to correct slabp list: */
2845	list_del(&page->lru);
2846	if (page->active == cachep->num) {
2847		list_add(&page->lru, &n->slabs_full);
2848		if (OBJFREELIST_SLAB(cachep)) {
2849#if DEBUG
2850			/* Poisoning will be done without holding the lock */
2851			if (cachep->flags & SLAB_POISON) {
2852				void **objp = page->freelist;
2853
2854				*objp = *list;
2855				*list = objp;
2856			}
2857#endif
2858			page->freelist = NULL;
2859		}
2860	} else
2861		list_add(&page->lru, &n->slabs_partial);
2862}
2863
2864/* Try to find non-pfmemalloc slab if needed */
2865static noinline struct page *get_valid_first_slab(struct kmem_cache_node *n,
2866					struct page *page, bool pfmemalloc)
2867{
2868	if (!page)
2869		return NULL;
2870
2871	if (pfmemalloc)
2872		return page;
2873
2874	if (!PageSlabPfmemalloc(page))
2875		return page;
2876
2877	/* No need to keep pfmemalloc slab if we have enough free objects */
2878	if (n->free_objects > n->free_limit) {
2879		ClearPageSlabPfmemalloc(page);
2880		return page;
2881	}
2882
2883	/* Move pfmemalloc slab to the end of list to speed up next search */
2884	list_del(&page->lru);
2885	if (!page->active) {
2886		list_add_tail(&page->lru, &n->slabs_free);
2887		n->free_slabs++;
2888	} else
2889		list_add_tail(&page->lru, &n->slabs_partial);
2890
2891	list_for_each_entry(page, &n->slabs_partial, lru) {
2892		if (!PageSlabPfmemalloc(page))
2893			return page;
2894	}
2895
2896	n->free_touched = 1;
2897	list_for_each_entry(page, &n->slabs_free, lru) {
2898		if (!PageSlabPfmemalloc(page)) {
2899			n->free_slabs--;
2900			return page;
2901		}
2902	}
2903
2904	return NULL;
2905}
2906
2907static struct page *get_first_slab(struct kmem_cache_node *n, bool pfmemalloc)
2908{
2909	struct page *page;
2910
2911	assert_spin_locked(&n->list_lock);
2912	page = list_first_entry_or_null(&n->slabs_partial, struct page, lru);
2913	if (!page) {
2914		n->free_touched = 1;
2915		page = list_first_entry_or_null(&n->slabs_free, struct page,
2916						lru);
2917		if (page)
2918			n->free_slabs--;
2919	}
2920
2921	if (sk_memalloc_socks())
2922		page = get_valid_first_slab(n, page, pfmemalloc);
2923
2924	return page;
2925}
2926
2927static noinline void *cache_alloc_pfmemalloc(struct kmem_cache *cachep,
2928				struct kmem_cache_node *n, gfp_t flags)
2929{
2930	struct page *page;
2931	void *obj;
2932	void *list = NULL;
2933
2934	if (!gfp_pfmemalloc_allowed(flags))
2935		return NULL;
2936
2937	spin_lock(&n->list_lock);
2938	page = get_first_slab(n, true);
2939	if (!page) {
2940		spin_unlock(&n->list_lock);
2941		return NULL;
2942	}
2943
2944	obj = slab_get_obj(cachep, page);
2945	n->free_objects--;
2946
2947	fixup_slab_list(cachep, n, page, &list);
2948
2949	spin_unlock(&n->list_lock);
2950	fixup_objfreelist_debug(cachep, &list);
2951
2952	return obj;
2953}
2954
2955/*
2956 * Slab list should be fixed up by fixup_slab_list() for existing slab
2957 * or cache_grow_end() for new slab
2958 */
2959static __always_inline int alloc_block(struct kmem_cache *cachep,
2960		struct array_cache *ac, struct page *page, int batchcount)
2961{
2962	/*
2963	 * There must be at least one object available for
2964	 * allocation.
2965	 */
2966	BUG_ON(page->active >= cachep->num);
2967
2968	while (page->active < cachep->num && batchcount--) {
2969		STATS_INC_ALLOCED(cachep);
2970		STATS_INC_ACTIVE(cachep);
2971		STATS_SET_HIGH(cachep);
2972
2973		ac->entry[ac->avail++] = slab_get_obj(cachep, page);
2974	}
2975
2976	return batchcount;
2977}
2978
2979static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2980{
2981	int batchcount;
2982	struct kmem_cache_node *n;
2983	struct array_cache *ac, *shared;
2984	int node;
2985	void *list = NULL;
2986	struct page *page;
2987
2988	check_irq_off();
2989	node = numa_mem_id();
2990
 
2991	ac = cpu_cache_get(cachep);
2992	batchcount = ac->batchcount;
2993	if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2994		/*
2995		 * If there was little recent activity on this cache, then
2996		 * perform only a partial refill.  Otherwise we could generate
2997		 * refill bouncing.
2998		 */
2999		batchcount = BATCHREFILL_LIMIT;
3000	}
3001	n = get_node(cachep, node);
3002
3003	BUG_ON(ac->avail > 0 || !n);
3004	shared = READ_ONCE(n->shared);
3005	if (!n->free_objects && (!shared || !shared->avail))
3006		goto direct_grow;
3007
3008	spin_lock(&n->list_lock);
3009	shared = READ_ONCE(n->shared);
3010
3011	/* See if we can refill from the shared array */
3012	if (shared && transfer_objects(ac, shared, batchcount)) {
3013		shared->touched = 1;
3014		goto alloc_done;
3015	}
3016
3017	while (batchcount > 0) {
 
3018		/* Get slab alloc is to come from. */
3019		page = get_first_slab(n, false);
3020		if (!page)
3021			goto must_grow;
3022
3023		check_spinlock_acquired(cachep);
3024
3025		batchcount = alloc_block(cachep, ac, page, batchcount);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3026		fixup_slab_list(cachep, n, page, &list);
3027	}
3028
3029must_grow:
3030	n->free_objects -= ac->avail;
3031alloc_done:
3032	spin_unlock(&n->list_lock);
3033	fixup_objfreelist_debug(cachep, &list);
3034
3035direct_grow:
3036	if (unlikely(!ac->avail)) {
 
 
3037		/* Check if we can use obj in pfmemalloc slab */
3038		if (sk_memalloc_socks()) {
3039			void *obj = cache_alloc_pfmemalloc(cachep, n, flags);
3040
3041			if (obj)
3042				return obj;
3043		}
3044
3045		page = cache_grow_begin(cachep, gfp_exact_node(flags), node);
3046
3047		/*
3048		 * cache_grow_begin() can reenable interrupts,
3049		 * then ac could change.
3050		 */
3051		ac = cpu_cache_get(cachep);
3052		if (!ac->avail && page)
3053			alloc_block(cachep, ac, page, batchcount);
3054		cache_grow_end(cachep, page);
3055
3056		if (!ac->avail)
 
3057			return NULL;
 
 
 
3058	}
3059	ac->touched = 1;
3060
3061	return ac->entry[--ac->avail];
3062}
3063
3064static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3065						gfp_t flags)
3066{
3067	might_sleep_if(gfpflags_allow_blocking(flags));
 
 
 
3068}
3069
3070#if DEBUG
3071static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3072				gfp_t flags, void *objp, unsigned long caller)
3073{
3074	if (!objp)
3075		return objp;
3076	if (cachep->flags & SLAB_POISON) {
3077		check_poison_obj(cachep, objp);
3078		slab_kernel_map(cachep, objp, 1, 0);
3079		poison_obj(cachep, objp, POISON_INUSE);
3080	}
3081	if (cachep->flags & SLAB_STORE_USER)
3082		*dbg_userword(cachep, objp) = (void *)caller;
3083
3084	if (cachep->flags & SLAB_RED_ZONE) {
3085		if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3086				*dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3087			slab_error(cachep, "double free, or memory outside object was overwritten");
3088			pr_err("%px: redzone 1:0x%llx, redzone 2:0x%llx\n",
3089			       objp, *dbg_redzone1(cachep, objp),
3090			       *dbg_redzone2(cachep, objp));
3091		}
3092		*dbg_redzone1(cachep, objp) = RED_ACTIVE;
3093		*dbg_redzone2(cachep, objp) = RED_ACTIVE;
3094	}
3095
3096	objp += obj_offset(cachep);
3097	if (cachep->ctor && cachep->flags & SLAB_POISON)
3098		cachep->ctor(objp);
3099	if (ARCH_SLAB_MINALIGN &&
3100	    ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3101		pr_err("0x%px: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3102		       objp, (int)ARCH_SLAB_MINALIGN);
3103	}
3104	return objp;
3105}
3106#else
3107#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3108#endif
3109
3110static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3111{
3112	void *objp;
3113	struct array_cache *ac;
3114
3115	check_irq_off();
3116
3117	ac = cpu_cache_get(cachep);
3118	if (likely(ac->avail)) {
3119		ac->touched = 1;
3120		objp = ac->entry[--ac->avail];
3121
3122		STATS_INC_ALLOCHIT(cachep);
3123		goto out;
3124	}
3125
3126	STATS_INC_ALLOCMISS(cachep);
3127	objp = cache_alloc_refill(cachep, flags);
3128	/*
3129	 * the 'ac' may be updated by cache_alloc_refill(),
3130	 * and kmemleak_erase() requires its correct value.
3131	 */
3132	ac = cpu_cache_get(cachep);
3133
3134out:
3135	/*
3136	 * To avoid a false negative, if an object that is in one of the
3137	 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3138	 * treat the array pointers as a reference to the object.
3139	 */
3140	if (objp)
3141		kmemleak_erase(&ac->entry[ac->avail]);
3142	return objp;
3143}
3144
3145#ifdef CONFIG_NUMA
3146/*
3147 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
3148 *
3149 * If we are in_interrupt, then process context, including cpusets and
3150 * mempolicy, may not apply and should not be used for allocation policy.
3151 */
3152static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3153{
3154	int nid_alloc, nid_here;
3155
3156	if (in_interrupt() || (flags & __GFP_THISNODE))
3157		return NULL;
3158	nid_alloc = nid_here = numa_mem_id();
3159	if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3160		nid_alloc = cpuset_slab_spread_node();
3161	else if (current->mempolicy)
3162		nid_alloc = mempolicy_slab_node();
3163	if (nid_alloc != nid_here)
3164		return ____cache_alloc_node(cachep, flags, nid_alloc);
3165	return NULL;
3166}
3167
3168/*
3169 * Fallback function if there was no memory available and no objects on a
3170 * certain node and fall back is permitted. First we scan all the
3171 * available node for available objects. If that fails then we
3172 * perform an allocation without specifying a node. This allows the page
3173 * allocator to do its reclaim / fallback magic. We then insert the
3174 * slab into the proper nodelist and then allocate from it.
3175 */
3176static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3177{
3178	struct zonelist *zonelist;
 
3179	struct zoneref *z;
3180	struct zone *zone;
3181	enum zone_type high_zoneidx = gfp_zone(flags);
3182	void *obj = NULL;
3183	struct page *page;
3184	int nid;
3185	unsigned int cpuset_mems_cookie;
3186
3187	if (flags & __GFP_THISNODE)
3188		return NULL;
3189
 
 
3190retry_cpuset:
3191	cpuset_mems_cookie = read_mems_allowed_begin();
3192	zonelist = node_zonelist(mempolicy_slab_node(), flags);
3193
3194retry:
3195	/*
3196	 * Look through allowed nodes for objects available
3197	 * from existing per node queues.
3198	 */
3199	for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3200		nid = zone_to_nid(zone);
3201
3202		if (cpuset_zone_allowed(zone, flags) &&
3203			get_node(cache, nid) &&
3204			get_node(cache, nid)->free_objects) {
3205				obj = ____cache_alloc_node(cache,
3206					gfp_exact_node(flags), nid);
3207				if (obj)
3208					break;
3209		}
3210	}
3211
3212	if (!obj) {
3213		/*
3214		 * This allocation will be performed within the constraints
3215		 * of the current cpuset / memory policy requirements.
3216		 * We may trigger various forms of reclaim on the allowed
3217		 * set and go into memory reserves if necessary.
3218		 */
3219		page = cache_grow_begin(cache, flags, numa_mem_id());
3220		cache_grow_end(cache, page);
 
 
 
 
 
 
3221		if (page) {
3222			nid = page_to_nid(page);
3223			obj = ____cache_alloc_node(cache,
3224				gfp_exact_node(flags), nid);
3225
3226			/*
3227			 * Another processor may allocate the objects in
3228			 * the slab since we are not holding any locks.
3229			 */
3230			if (!obj)
3231				goto retry;
 
 
 
 
 
 
 
 
 
 
 
 
 
3232		}
3233	}
3234
3235	if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
3236		goto retry_cpuset;
3237	return obj;
3238}
3239
3240/*
3241 * A interface to enable slab creation on nodeid
3242 */
3243static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3244				int nodeid)
3245{
3246	struct page *page;
3247	struct kmem_cache_node *n;
3248	void *obj = NULL;
3249	void *list = NULL;
 
3250
3251	VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES);
3252	n = get_node(cachep, nodeid);
3253	BUG_ON(!n);
3254
 
3255	check_irq_off();
3256	spin_lock(&n->list_lock);
3257	page = get_first_slab(n, false);
3258	if (!page)
3259		goto must_grow;
3260
3261	check_spinlock_acquired_node(cachep, nodeid);
3262
3263	STATS_INC_NODEALLOCS(cachep);
3264	STATS_INC_ACTIVE(cachep);
3265	STATS_SET_HIGH(cachep);
3266
3267	BUG_ON(page->active == cachep->num);
3268
3269	obj = slab_get_obj(cachep, page);
3270	n->free_objects--;
3271
3272	fixup_slab_list(cachep, n, page, &list);
3273
3274	spin_unlock(&n->list_lock);
3275	fixup_objfreelist_debug(cachep, &list);
3276	return obj;
3277
3278must_grow:
3279	spin_unlock(&n->list_lock);
3280	page = cache_grow_begin(cachep, gfp_exact_node(flags), nodeid);
3281	if (page) {
3282		/* This slab isn't counted yet so don't update free_objects */
3283		obj = slab_get_obj(cachep, page);
3284	}
3285	cache_grow_end(cachep, page);
3286
3287	return obj ? obj : fallback_alloc(cachep, flags);
 
3288}
3289
3290static __always_inline void *
3291slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3292		   unsigned long caller)
3293{
3294	unsigned long save_flags;
3295	void *ptr;
3296	int slab_node = numa_mem_id();
3297
3298	flags &= gfp_allowed_mask;
3299	cachep = slab_pre_alloc_hook(cachep, flags);
3300	if (unlikely(!cachep))
3301		return NULL;
3302
3303	cache_alloc_debugcheck_before(cachep, flags);
3304	local_irq_save(save_flags);
3305
3306	if (nodeid == NUMA_NO_NODE)
3307		nodeid = slab_node;
3308
3309	if (unlikely(!get_node(cachep, nodeid))) {
3310		/* Node not bootstrapped yet */
3311		ptr = fallback_alloc(cachep, flags);
3312		goto out;
3313	}
3314
3315	if (nodeid == slab_node) {
3316		/*
3317		 * Use the locally cached objects if possible.
3318		 * However ____cache_alloc does not allow fallback
3319		 * to other nodes. It may fail while we still have
3320		 * objects on other nodes available.
3321		 */
3322		ptr = ____cache_alloc(cachep, flags);
3323		if (ptr)
3324			goto out;
3325	}
3326	/* ___cache_alloc_node can fall back to other nodes */
3327	ptr = ____cache_alloc_node(cachep, flags, nodeid);
3328  out:
3329	local_irq_restore(save_flags);
3330	ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3331
3332	if (unlikely(flags & __GFP_ZERO) && ptr)
3333		memset(ptr, 0, cachep->object_size);
3334
3335	slab_post_alloc_hook(cachep, flags, 1, &ptr);
3336	return ptr;
3337}
3338
3339static __always_inline void *
3340__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3341{
3342	void *objp;
3343
3344	if (current->mempolicy || cpuset_do_slab_mem_spread()) {
3345		objp = alternate_node_alloc(cache, flags);
3346		if (objp)
3347			goto out;
3348	}
3349	objp = ____cache_alloc(cache, flags);
3350
3351	/*
3352	 * We may just have run out of memory on the local node.
3353	 * ____cache_alloc_node() knows how to locate memory on other nodes
3354	 */
3355	if (!objp)
3356		objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3357
3358  out:
3359	return objp;
3360}
3361#else
3362
3363static __always_inline void *
3364__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3365{
3366	return ____cache_alloc(cachep, flags);
3367}
3368
3369#endif /* CONFIG_NUMA */
3370
3371static __always_inline void *
3372slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3373{
3374	unsigned long save_flags;
3375	void *objp;
3376
3377	flags &= gfp_allowed_mask;
3378	cachep = slab_pre_alloc_hook(cachep, flags);
3379	if (unlikely(!cachep))
3380		return NULL;
3381
3382	cache_alloc_debugcheck_before(cachep, flags);
3383	local_irq_save(save_flags);
3384	objp = __do_cache_alloc(cachep, flags);
3385	local_irq_restore(save_flags);
3386	objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3387	prefetchw(objp);
3388
3389	if (unlikely(flags & __GFP_ZERO) && objp)
3390		memset(objp, 0, cachep->object_size);
3391
3392	slab_post_alloc_hook(cachep, flags, 1, &objp);
3393	return objp;
3394}
3395
3396/*
3397 * Caller needs to acquire correct kmem_cache_node's list_lock
3398 * @list: List of detached free slabs should be freed by caller
3399 */
3400static void free_block(struct kmem_cache *cachep, void **objpp,
3401			int nr_objects, int node, struct list_head *list)
3402{
3403	int i;
3404	struct kmem_cache_node *n = get_node(cachep, node);
3405	struct page *page;
3406
3407	n->free_objects += nr_objects;
3408
3409	for (i = 0; i < nr_objects; i++) {
3410		void *objp;
3411		struct page *page;
3412
3413		objp = objpp[i];
3414
3415		page = virt_to_head_page(objp);
3416		list_del(&page->lru);
3417		check_spinlock_acquired_node(cachep, node);
3418		slab_put_obj(cachep, page, objp);
3419		STATS_DEC_ACTIVE(cachep);
 
3420
3421		/* fixup slab chains */
3422		if (page->active == 0) {
3423			list_add(&page->lru, &n->slabs_free);
3424			n->free_slabs++;
 
 
 
 
3425		} else {
3426			/* Unconditionally move a slab to the end of the
3427			 * partial list on free - maximum time for the
3428			 * other objects to be freed, too.
3429			 */
3430			list_add_tail(&page->lru, &n->slabs_partial);
3431		}
3432	}
3433
3434	while (n->free_objects > n->free_limit && !list_empty(&n->slabs_free)) {
3435		n->free_objects -= cachep->num;
3436
3437		page = list_last_entry(&n->slabs_free, struct page, lru);
3438		list_move(&page->lru, list);
3439		n->free_slabs--;
3440		n->total_slabs--;
3441	}
3442}
3443
3444static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3445{
3446	int batchcount;
3447	struct kmem_cache_node *n;
3448	int node = numa_mem_id();
3449	LIST_HEAD(list);
3450
3451	batchcount = ac->batchcount;
3452
3453	check_irq_off();
3454	n = get_node(cachep, node);
3455	spin_lock(&n->list_lock);
3456	if (n->shared) {
3457		struct array_cache *shared_array = n->shared;
3458		int max = shared_array->limit - shared_array->avail;
3459		if (max) {
3460			if (batchcount > max)
3461				batchcount = max;
3462			memcpy(&(shared_array->entry[shared_array->avail]),
3463			       ac->entry, sizeof(void *) * batchcount);
3464			shared_array->avail += batchcount;
3465			goto free_done;
3466		}
3467	}
3468
3469	free_block(cachep, ac->entry, batchcount, node, &list);
3470free_done:
3471#if STATS
3472	{
3473		int i = 0;
3474		struct page *page;
3475
3476		list_for_each_entry(page, &n->slabs_free, lru) {
3477			BUG_ON(page->active);
3478
3479			i++;
3480		}
3481		STATS_SET_FREEABLE(cachep, i);
3482	}
3483#endif
3484	spin_unlock(&n->list_lock);
3485	slabs_destroy(cachep, &list);
3486	ac->avail -= batchcount;
3487	memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3488}
3489
3490/*
3491 * Release an obj back to its cache. If the obj has a constructed state, it must
3492 * be in this state _before_ it is released.  Called with disabled ints.
3493 */
3494static __always_inline void __cache_free(struct kmem_cache *cachep, void *objp,
3495					 unsigned long caller)
3496{
3497	/* Put the object into the quarantine, don't touch it for now. */
3498	if (kasan_slab_free(cachep, objp, _RET_IP_))
3499		return;
3500
3501	___cache_free(cachep, objp, caller);
3502}
3503
3504void ___cache_free(struct kmem_cache *cachep, void *objp,
3505		unsigned long caller)
3506{
3507	struct array_cache *ac = cpu_cache_get(cachep);
3508
3509	check_irq_off();
3510	kmemleak_free_recursive(objp, cachep->flags);
3511	objp = cache_free_debugcheck(cachep, objp, caller);
3512
 
 
3513	/*
3514	 * Skip calling cache_free_alien() when the platform is not numa.
3515	 * This will avoid cache misses that happen while accessing slabp (which
3516	 * is per page memory  reference) to get nodeid. Instead use a global
3517	 * variable to skip the call, which is mostly likely to be present in
3518	 * the cache.
3519	 */
3520	if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3521		return;
3522
3523	if (ac->avail < ac->limit) {
3524		STATS_INC_FREEHIT(cachep);
3525	} else {
3526		STATS_INC_FREEMISS(cachep);
3527		cache_flusharray(cachep, ac);
3528	}
3529
3530	if (sk_memalloc_socks()) {
3531		struct page *page = virt_to_head_page(objp);
3532
3533		if (unlikely(PageSlabPfmemalloc(page))) {
3534			cache_free_pfmemalloc(cachep, page, objp);
3535			return;
3536		}
3537	}
3538
3539	ac->entry[ac->avail++] = objp;
3540}
3541
3542/**
3543 * kmem_cache_alloc - Allocate an object
3544 * @cachep: The cache to allocate from.
3545 * @flags: See kmalloc().
3546 *
3547 * Allocate an object from this cache.  The flags are only relevant
3548 * if the cache has no available objects.
3549 */
3550void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3551{
3552	void *ret = slab_alloc(cachep, flags, _RET_IP_);
3553
3554	kasan_slab_alloc(cachep, ret, flags);
3555	trace_kmem_cache_alloc(_RET_IP_, ret,
3556			       cachep->object_size, cachep->size, flags);
3557
3558	return ret;
3559}
3560EXPORT_SYMBOL(kmem_cache_alloc);
3561
3562static __always_inline void
3563cache_alloc_debugcheck_after_bulk(struct kmem_cache *s, gfp_t flags,
3564				  size_t size, void **p, unsigned long caller)
3565{
3566	size_t i;
3567
3568	for (i = 0; i < size; i++)
3569		p[i] = cache_alloc_debugcheck_after(s, flags, p[i], caller);
3570}
3571
3572int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3573			  void **p)
3574{
3575	size_t i;
3576
3577	s = slab_pre_alloc_hook(s, flags);
3578	if (!s)
3579		return 0;
3580
3581	cache_alloc_debugcheck_before(s, flags);
3582
3583	local_irq_disable();
3584	for (i = 0; i < size; i++) {
3585		void *objp = __do_cache_alloc(s, flags);
3586
3587		if (unlikely(!objp))
3588			goto error;
3589		p[i] = objp;
3590	}
3591	local_irq_enable();
3592
3593	cache_alloc_debugcheck_after_bulk(s, flags, size, p, _RET_IP_);
3594
3595	/* Clear memory outside IRQ disabled section */
3596	if (unlikely(flags & __GFP_ZERO))
3597		for (i = 0; i < size; i++)
3598			memset(p[i], 0, s->object_size);
3599
3600	slab_post_alloc_hook(s, flags, size, p);
3601	/* FIXME: Trace call missing. Christoph would like a bulk variant */
3602	return size;
3603error:
3604	local_irq_enable();
3605	cache_alloc_debugcheck_after_bulk(s, flags, i, p, _RET_IP_);
3606	slab_post_alloc_hook(s, flags, i, p);
3607	__kmem_cache_free_bulk(s, i, p);
3608	return 0;
3609}
3610EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3611
3612#ifdef CONFIG_TRACING
3613void *
3614kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3615{
3616	void *ret;
3617
3618	ret = slab_alloc(cachep, flags, _RET_IP_);
3619
3620	kasan_kmalloc(cachep, ret, size, flags);
3621	trace_kmalloc(_RET_IP_, ret,
3622		      size, cachep->size, flags);
3623	return ret;
3624}
3625EXPORT_SYMBOL(kmem_cache_alloc_trace);
3626#endif
3627
3628#ifdef CONFIG_NUMA
3629/**
3630 * kmem_cache_alloc_node - Allocate an object on the specified node
3631 * @cachep: The cache to allocate from.
3632 * @flags: See kmalloc().
3633 * @nodeid: node number of the target node.
3634 *
3635 * Identical to kmem_cache_alloc but it will allocate memory on the given
3636 * node, which can improve the performance for cpu bound structures.
3637 *
3638 * Fallback to other node is possible if __GFP_THISNODE is not set.
3639 */
3640void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3641{
3642	void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3643
3644	kasan_slab_alloc(cachep, ret, flags);
3645	trace_kmem_cache_alloc_node(_RET_IP_, ret,
3646				    cachep->object_size, cachep->size,
3647				    flags, nodeid);
3648
3649	return ret;
3650}
3651EXPORT_SYMBOL(kmem_cache_alloc_node);
3652
3653#ifdef CONFIG_TRACING
3654void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3655				  gfp_t flags,
3656				  int nodeid,
3657				  size_t size)
3658{
3659	void *ret;
3660
3661	ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3662
3663	kasan_kmalloc(cachep, ret, size, flags);
3664	trace_kmalloc_node(_RET_IP_, ret,
3665			   size, cachep->size,
3666			   flags, nodeid);
3667	return ret;
3668}
3669EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3670#endif
3671
3672static __always_inline void *
3673__do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3674{
3675	struct kmem_cache *cachep;
3676	void *ret;
3677
3678	cachep = kmalloc_slab(size, flags);
3679	if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3680		return cachep;
3681	ret = kmem_cache_alloc_node_trace(cachep, flags, node, size);
3682	kasan_kmalloc(cachep, ret, size, flags);
3683
3684	return ret;
3685}
3686
3687void *__kmalloc_node(size_t size, gfp_t flags, int node)
3688{
3689	return __do_kmalloc_node(size, flags, node, _RET_IP_);
3690}
3691EXPORT_SYMBOL(__kmalloc_node);
3692
3693void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3694		int node, unsigned long caller)
3695{
3696	return __do_kmalloc_node(size, flags, node, caller);
3697}
3698EXPORT_SYMBOL(__kmalloc_node_track_caller);
3699#endif /* CONFIG_NUMA */
3700
3701/**
3702 * __do_kmalloc - allocate memory
3703 * @size: how many bytes of memory are required.
3704 * @flags: the type of memory to allocate (see kmalloc).
3705 * @caller: function caller for debug tracking of the caller
3706 */
3707static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3708					  unsigned long caller)
3709{
3710	struct kmem_cache *cachep;
3711	void *ret;
3712
3713	cachep = kmalloc_slab(size, flags);
3714	if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3715		return cachep;
3716	ret = slab_alloc(cachep, flags, caller);
3717
3718	kasan_kmalloc(cachep, ret, size, flags);
3719	trace_kmalloc(caller, ret,
3720		      size, cachep->size, flags);
3721
3722	return ret;
3723}
3724
3725void *__kmalloc(size_t size, gfp_t flags)
3726{
3727	return __do_kmalloc(size, flags, _RET_IP_);
3728}
3729EXPORT_SYMBOL(__kmalloc);
3730
3731void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3732{
3733	return __do_kmalloc(size, flags, caller);
3734}
3735EXPORT_SYMBOL(__kmalloc_track_caller);
3736
3737/**
3738 * kmem_cache_free - Deallocate an object
3739 * @cachep: The cache the allocation was from.
3740 * @objp: The previously allocated object.
3741 *
3742 * Free an object which was previously allocated from this
3743 * cache.
3744 */
3745void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3746{
3747	unsigned long flags;
3748	cachep = cache_from_obj(cachep, objp);
3749	if (!cachep)
3750		return;
3751
3752	local_irq_save(flags);
3753	debug_check_no_locks_freed(objp, cachep->object_size);
3754	if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3755		debug_check_no_obj_freed(objp, cachep->object_size);
3756	__cache_free(cachep, objp, _RET_IP_);
3757	local_irq_restore(flags);
3758
3759	trace_kmem_cache_free(_RET_IP_, objp);
3760}
3761EXPORT_SYMBOL(kmem_cache_free);
3762
3763void kmem_cache_free_bulk(struct kmem_cache *orig_s, size_t size, void **p)
3764{
3765	struct kmem_cache *s;
3766	size_t i;
3767
3768	local_irq_disable();
3769	for (i = 0; i < size; i++) {
3770		void *objp = p[i];
3771
3772		if (!orig_s) /* called via kfree_bulk */
3773			s = virt_to_cache(objp);
3774		else
3775			s = cache_from_obj(orig_s, objp);
3776
3777		debug_check_no_locks_freed(objp, s->object_size);
3778		if (!(s->flags & SLAB_DEBUG_OBJECTS))
3779			debug_check_no_obj_freed(objp, s->object_size);
3780
3781		__cache_free(s, objp, _RET_IP_);
3782	}
3783	local_irq_enable();
3784
3785	/* FIXME: add tracing */
3786}
3787EXPORT_SYMBOL(kmem_cache_free_bulk);
3788
3789/**
3790 * kfree - free previously allocated memory
3791 * @objp: pointer returned by kmalloc.
3792 *
3793 * If @objp is NULL, no operation is performed.
3794 *
3795 * Don't free memory not originally allocated by kmalloc()
3796 * or you will run into trouble.
3797 */
3798void kfree(const void *objp)
3799{
3800	struct kmem_cache *c;
3801	unsigned long flags;
3802
3803	trace_kfree(_RET_IP_, objp);
3804
3805	if (unlikely(ZERO_OR_NULL_PTR(objp)))
3806		return;
3807	local_irq_save(flags);
3808	kfree_debugcheck(objp);
3809	c = virt_to_cache(objp);
3810	debug_check_no_locks_freed(objp, c->object_size);
3811
3812	debug_check_no_obj_freed(objp, c->object_size);
3813	__cache_free(c, (void *)objp, _RET_IP_);
3814	local_irq_restore(flags);
3815}
3816EXPORT_SYMBOL(kfree);
3817
3818/*
3819 * This initializes kmem_cache_node or resizes various caches for all nodes.
3820 */
3821static int setup_kmem_cache_nodes(struct kmem_cache *cachep, gfp_t gfp)
3822{
3823	int ret;
3824	int node;
3825	struct kmem_cache_node *n;
 
 
3826
3827	for_each_online_node(node) {
3828		ret = setup_kmem_cache_node(cachep, node, gfp, true);
3829		if (ret)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3830			goto fail;
 
3831
 
 
 
 
 
 
 
 
3832	}
3833
3834	return 0;
3835
3836fail:
3837	if (!cachep->list.next) {
3838		/* Cache is not active yet. Roll back what we did */
3839		node--;
3840		while (node >= 0) {
3841			n = get_node(cachep, node);
3842			if (n) {
3843				kfree(n->shared);
3844				free_alien_cache(n->alien);
3845				kfree(n);
3846				cachep->node[node] = NULL;
3847			}
3848			node--;
3849		}
3850	}
3851	return -ENOMEM;
3852}
3853
3854/* Always called with the slab_mutex held */
3855static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
3856				int batchcount, int shared, gfp_t gfp)
3857{
3858	struct array_cache __percpu *cpu_cache, *prev;
3859	int cpu;
3860
3861	cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount);
3862	if (!cpu_cache)
3863		return -ENOMEM;
3864
3865	prev = cachep->cpu_cache;
3866	cachep->cpu_cache = cpu_cache;
3867	/*
3868	 * Without a previous cpu_cache there's no need to synchronize remote
3869	 * cpus, so skip the IPIs.
3870	 */
3871	if (prev)
3872		kick_all_cpus_sync();
3873
3874	check_irq_on();
3875	cachep->batchcount = batchcount;
3876	cachep->limit = limit;
3877	cachep->shared = shared;
3878
3879	if (!prev)
3880		goto setup_node;
3881
3882	for_each_online_cpu(cpu) {
3883		LIST_HEAD(list);
3884		int node;
3885		struct kmem_cache_node *n;
3886		struct array_cache *ac = per_cpu_ptr(prev, cpu);
3887
3888		node = cpu_to_mem(cpu);
3889		n = get_node(cachep, node);
3890		spin_lock_irq(&n->list_lock);
3891		free_block(cachep, ac->entry, ac->avail, node, &list);
3892		spin_unlock_irq(&n->list_lock);
3893		slabs_destroy(cachep, &list);
3894	}
3895	free_percpu(prev);
3896
3897setup_node:
3898	return setup_kmem_cache_nodes(cachep, gfp);
3899}
3900
3901static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3902				int batchcount, int shared, gfp_t gfp)
3903{
3904	int ret;
3905	struct kmem_cache *c;
3906
3907	ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3908
3909	if (slab_state < FULL)
3910		return ret;
3911
3912	if ((ret < 0) || !is_root_cache(cachep))
3913		return ret;
3914
3915	lockdep_assert_held(&slab_mutex);
3916	for_each_memcg_cache(c, cachep) {
3917		/* return value determined by the root cache only */
3918		__do_tune_cpucache(c, limit, batchcount, shared, gfp);
3919	}
3920
3921	return ret;
3922}
3923
3924/* Called with slab_mutex held always */
3925static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3926{
3927	int err;
3928	int limit = 0;
3929	int shared = 0;
3930	int batchcount = 0;
3931
3932	err = cache_random_seq_create(cachep, cachep->num, gfp);
3933	if (err)
3934		goto end;
3935
3936	if (!is_root_cache(cachep)) {
3937		struct kmem_cache *root = memcg_root_cache(cachep);
3938		limit = root->limit;
3939		shared = root->shared;
3940		batchcount = root->batchcount;
3941	}
3942
3943	if (limit && shared && batchcount)
3944		goto skip_setup;
3945	/*
3946	 * The head array serves three purposes:
3947	 * - create a LIFO ordering, i.e. return objects that are cache-warm
3948	 * - reduce the number of spinlock operations.
3949	 * - reduce the number of linked list operations on the slab and
3950	 *   bufctl chains: array operations are cheaper.
3951	 * The numbers are guessed, we should auto-tune as described by
3952	 * Bonwick.
3953	 */
3954	if (cachep->size > 131072)
3955		limit = 1;
3956	else if (cachep->size > PAGE_SIZE)
3957		limit = 8;
3958	else if (cachep->size > 1024)
3959		limit = 24;
3960	else if (cachep->size > 256)
3961		limit = 54;
3962	else
3963		limit = 120;
3964
3965	/*
3966	 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3967	 * allocation behaviour: Most allocs on one cpu, most free operations
3968	 * on another cpu. For these cases, an efficient object passing between
3969	 * cpus is necessary. This is provided by a shared array. The array
3970	 * replaces Bonwick's magazine layer.
3971	 * On uniprocessor, it's functionally equivalent (but less efficient)
3972	 * to a larger limit. Thus disabled by default.
3973	 */
3974	shared = 0;
3975	if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
3976		shared = 8;
3977
3978#if DEBUG
3979	/*
3980	 * With debugging enabled, large batchcount lead to excessively long
3981	 * periods with disabled local interrupts. Limit the batchcount
3982	 */
3983	if (limit > 32)
3984		limit = 32;
3985#endif
3986	batchcount = (limit + 1) / 2;
3987skip_setup:
3988	err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3989end:
3990	if (err)
3991		pr_err("enable_cpucache failed for %s, error %d\n",
3992		       cachep->name, -err);
3993	return err;
3994}
3995
3996/*
3997 * Drain an array if it contains any elements taking the node lock only if
3998 * necessary. Note that the node listlock also protects the array_cache
3999 * if drain_array() is used on the shared array.
4000 */
4001static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
4002			 struct array_cache *ac, int node)
4003{
4004	LIST_HEAD(list);
4005
4006	/* ac from n->shared can be freed if we don't hold the slab_mutex. */
4007	check_mutex_acquired();
4008
4009	if (!ac || !ac->avail)
4010		return;
4011
4012	if (ac->touched) {
4013		ac->touched = 0;
4014		return;
 
 
 
 
 
 
 
 
 
 
 
 
4015	}
4016
4017	spin_lock_irq(&n->list_lock);
4018	drain_array_locked(cachep, ac, node, false, &list);
4019	spin_unlock_irq(&n->list_lock);
4020
4021	slabs_destroy(cachep, &list);
4022}
4023
4024/**
4025 * cache_reap - Reclaim memory from caches.
4026 * @w: work descriptor
4027 *
4028 * Called from workqueue/eventd every few seconds.
4029 * Purpose:
4030 * - clear the per-cpu caches for this CPU.
4031 * - return freeable pages to the main free memory pool.
4032 *
4033 * If we cannot acquire the cache chain mutex then just give up - we'll try
4034 * again on the next iteration.
4035 */
4036static void cache_reap(struct work_struct *w)
4037{
4038	struct kmem_cache *searchp;
4039	struct kmem_cache_node *n;
4040	int node = numa_mem_id();
4041	struct delayed_work *work = to_delayed_work(w);
4042
4043	if (!mutex_trylock(&slab_mutex))
4044		/* Give up. Setup the next iteration. */
4045		goto out;
4046
4047	list_for_each_entry(searchp, &slab_caches, list) {
4048		check_irq_on();
4049
4050		/*
4051		 * We only take the node lock if absolutely necessary and we
4052		 * have established with reasonable certainty that
4053		 * we can do some work if the lock was obtained.
4054		 */
4055		n = get_node(searchp, node);
4056
4057		reap_alien(searchp, n);
4058
4059		drain_array(searchp, n, cpu_cache_get(searchp), node);
4060
4061		/*
4062		 * These are racy checks but it does not matter
4063		 * if we skip one check or scan twice.
4064		 */
4065		if (time_after(n->next_reap, jiffies))
4066			goto next;
4067
4068		n->next_reap = jiffies + REAPTIMEOUT_NODE;
4069
4070		drain_array(searchp, n, n->shared, node);
4071
4072		if (n->free_touched)
4073			n->free_touched = 0;
4074		else {
4075			int freed;
4076
4077			freed = drain_freelist(searchp, n, (n->free_limit +
4078				5 * searchp->num - 1) / (5 * searchp->num));
4079			STATS_ADD_REAPED(searchp, freed);
4080		}
4081next:
4082		cond_resched();
4083	}
4084	check_irq_on();
4085	mutex_unlock(&slab_mutex);
4086	next_reap_node();
4087out:
4088	/* Set up the next iteration */
4089	schedule_delayed_work_on(smp_processor_id(), work,
4090				round_jiffies_relative(REAPTIMEOUT_AC));
4091}
4092
 
4093void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
4094{
4095	unsigned long active_objs, num_objs, active_slabs;
4096	unsigned long total_slabs = 0, free_objs = 0, shared_avail = 0;
4097	unsigned long free_slabs = 0;
 
 
 
 
4098	int node;
4099	struct kmem_cache_node *n;
4100
 
 
4101	for_each_kmem_cache_node(cachep, node, n) {
 
4102		check_irq_on();
4103		spin_lock_irq(&n->list_lock);
4104
4105		total_slabs += n->total_slabs;
4106		free_slabs += n->free_slabs;
4107		free_objs += n->free_objects;
4108
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4109		if (n->shared)
4110			shared_avail += n->shared->avail;
4111
4112		spin_unlock_irq(&n->list_lock);
4113	}
4114	num_objs = total_slabs * cachep->num;
4115	active_slabs = total_slabs - free_slabs;
4116	active_objs = num_objs - free_objs;
 
 
 
 
 
4117
4118	sinfo->active_objs = active_objs;
4119	sinfo->num_objs = num_objs;
4120	sinfo->active_slabs = active_slabs;
4121	sinfo->num_slabs = total_slabs;
4122	sinfo->shared_avail = shared_avail;
4123	sinfo->limit = cachep->limit;
4124	sinfo->batchcount = cachep->batchcount;
4125	sinfo->shared = cachep->shared;
4126	sinfo->objects_per_slab = cachep->num;
4127	sinfo->cache_order = cachep->gfporder;
4128}
4129
4130void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4131{
4132#if STATS
4133	{			/* node stats */
4134		unsigned long high = cachep->high_mark;
4135		unsigned long allocs = cachep->num_allocations;
4136		unsigned long grown = cachep->grown;
4137		unsigned long reaped = cachep->reaped;
4138		unsigned long errors = cachep->errors;
4139		unsigned long max_freeable = cachep->max_freeable;
4140		unsigned long node_allocs = cachep->node_allocs;
4141		unsigned long node_frees = cachep->node_frees;
4142		unsigned long overflows = cachep->node_overflow;
4143
4144		seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
4145			   allocs, high, grown,
4146			   reaped, errors, max_freeable, node_allocs,
4147			   node_frees, overflows);
4148	}
4149	/* cpu stats */
4150	{
4151		unsigned long allochit = atomic_read(&cachep->allochit);
4152		unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4153		unsigned long freehit = atomic_read(&cachep->freehit);
4154		unsigned long freemiss = atomic_read(&cachep->freemiss);
4155
4156		seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4157			   allochit, allocmiss, freehit, freemiss);
4158	}
4159#endif
4160}
4161
4162#define MAX_SLABINFO_WRITE 128
4163/**
4164 * slabinfo_write - Tuning for the slab allocator
4165 * @file: unused
4166 * @buffer: user buffer
4167 * @count: data length
4168 * @ppos: unused
4169 */
4170ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4171		       size_t count, loff_t *ppos)
4172{
4173	char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4174	int limit, batchcount, shared, res;
4175	struct kmem_cache *cachep;
4176
4177	if (count > MAX_SLABINFO_WRITE)
4178		return -EINVAL;
4179	if (copy_from_user(&kbuf, buffer, count))
4180		return -EFAULT;
4181	kbuf[MAX_SLABINFO_WRITE] = '\0';
4182
4183	tmp = strchr(kbuf, ' ');
4184	if (!tmp)
4185		return -EINVAL;
4186	*tmp = '\0';
4187	tmp++;
4188	if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4189		return -EINVAL;
4190
4191	/* Find the cache in the chain of caches. */
4192	mutex_lock(&slab_mutex);
4193	res = -EINVAL;
4194	list_for_each_entry(cachep, &slab_caches, list) {
4195		if (!strcmp(cachep->name, kbuf)) {
4196			if (limit < 1 || batchcount < 1 ||
4197					batchcount > limit || shared < 0) {
4198				res = 0;
4199			} else {
4200				res = do_tune_cpucache(cachep, limit,
4201						       batchcount, shared,
4202						       GFP_KERNEL);
4203			}
4204			break;
4205		}
4206	}
4207	mutex_unlock(&slab_mutex);
4208	if (res >= 0)
4209		res = count;
4210	return res;
4211}
4212
4213#ifdef CONFIG_DEBUG_SLAB_LEAK
4214
4215static inline int add_caller(unsigned long *n, unsigned long v)
4216{
4217	unsigned long *p;
4218	int l;
4219	if (!v)
4220		return 1;
4221	l = n[1];
4222	p = n + 2;
4223	while (l) {
4224		int i = l/2;
4225		unsigned long *q = p + 2 * i;
4226		if (*q == v) {
4227			q[1]++;
4228			return 1;
4229		}
4230		if (*q > v) {
4231			l = i;
4232		} else {
4233			p = q + 2;
4234			l -= i + 1;
4235		}
4236	}
4237	if (++n[1] == n[0])
4238		return 0;
4239	memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4240	p[0] = v;
4241	p[1] = 1;
4242	return 1;
4243}
4244
4245static void handle_slab(unsigned long *n, struct kmem_cache *c,
4246						struct page *page)
4247{
4248	void *p;
4249	int i, j;
4250	unsigned long v;
4251
4252	if (n[0] == n[1])
4253		return;
4254	for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
4255		bool active = true;
4256
4257		for (j = page->active; j < c->num; j++) {
4258			if (get_free_obj(page, j) == i) {
4259				active = false;
4260				break;
4261			}
4262		}
4263
4264		if (!active)
4265			continue;
4266
4267		/*
4268		 * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table
4269		 * mapping is established when actual object allocation and
4270		 * we could mistakenly access the unmapped object in the cpu
4271		 * cache.
4272		 */
4273		if (probe_kernel_read(&v, dbg_userword(c, p), sizeof(v)))
4274			continue;
4275
4276		if (!add_caller(n, v))
4277			return;
4278	}
4279}
4280
4281static void show_symbol(struct seq_file *m, unsigned long address)
4282{
4283#ifdef CONFIG_KALLSYMS
4284	unsigned long offset, size;
4285	char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4286
4287	if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4288		seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4289		if (modname[0])
4290			seq_printf(m, " [%s]", modname);
4291		return;
4292	}
4293#endif
4294	seq_printf(m, "%px", (void *)address);
4295}
4296
4297static int leaks_show(struct seq_file *m, void *p)
4298{
4299	struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4300	struct page *page;
4301	struct kmem_cache_node *n;
4302	const char *name;
4303	unsigned long *x = m->private;
4304	int node;
4305	int i;
4306
4307	if (!(cachep->flags & SLAB_STORE_USER))
4308		return 0;
4309	if (!(cachep->flags & SLAB_RED_ZONE))
4310		return 0;
4311
4312	/*
4313	 * Set store_user_clean and start to grab stored user information
4314	 * for all objects on this cache. If some alloc/free requests comes
4315	 * during the processing, information would be wrong so restart
4316	 * whole processing.
4317	 */
4318	do {
4319		set_store_user_clean(cachep);
4320		drain_cpu_caches(cachep);
4321
4322		x[1] = 0;
4323
4324		for_each_kmem_cache_node(cachep, node, n) {
4325
4326			check_irq_on();
4327			spin_lock_irq(&n->list_lock);
4328
4329			list_for_each_entry(page, &n->slabs_full, lru)
4330				handle_slab(x, cachep, page);
4331			list_for_each_entry(page, &n->slabs_partial, lru)
4332				handle_slab(x, cachep, page);
4333			spin_unlock_irq(&n->list_lock);
4334		}
4335	} while (!is_store_user_clean(cachep));
4336
4337	name = cachep->name;
4338	if (x[0] == x[1]) {
4339		/* Increase the buffer size */
4340		mutex_unlock(&slab_mutex);
4341		m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4342		if (!m->private) {
4343			/* Too bad, we are really out */
4344			m->private = x;
4345			mutex_lock(&slab_mutex);
4346			return -ENOMEM;
4347		}
4348		*(unsigned long *)m->private = x[0] * 2;
4349		kfree(x);
4350		mutex_lock(&slab_mutex);
4351		/* Now make sure this entry will be retried */
4352		m->count = m->size;
4353		return 0;
4354	}
4355	for (i = 0; i < x[1]; i++) {
4356		seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4357		show_symbol(m, x[2*i+2]);
4358		seq_putc(m, '\n');
4359	}
4360
4361	return 0;
4362}
4363
4364static const struct seq_operations slabstats_op = {
4365	.start = slab_start,
4366	.next = slab_next,
4367	.stop = slab_stop,
4368	.show = leaks_show,
4369};
4370
4371static int slabstats_open(struct inode *inode, struct file *file)
4372{
4373	unsigned long *n;
4374
4375	n = __seq_open_private(file, &slabstats_op, PAGE_SIZE);
4376	if (!n)
4377		return -ENOMEM;
4378
4379	*n = PAGE_SIZE / (2 * sizeof(unsigned long));
4380
4381	return 0;
4382}
4383
4384static const struct file_operations proc_slabstats_operations = {
4385	.open		= slabstats_open,
4386	.read		= seq_read,
4387	.llseek		= seq_lseek,
4388	.release	= seq_release_private,
4389};
4390#endif
4391
4392static int __init slab_proc_init(void)
4393{
4394#ifdef CONFIG_DEBUG_SLAB_LEAK
4395	proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4396#endif
4397	return 0;
4398}
4399module_init(slab_proc_init);
4400
4401#ifdef CONFIG_HARDENED_USERCOPY
4402/*
4403 * Rejects incorrectly sized objects and objects that are to be copied
4404 * to/from userspace but do not fall entirely within the containing slab
4405 * cache's usercopy region.
4406 *
4407 * Returns NULL if check passes, otherwise const char * to name of cache
4408 * to indicate an error.
4409 */
4410void __check_heap_object(const void *ptr, unsigned long n, struct page *page,
4411			 bool to_user)
4412{
4413	struct kmem_cache *cachep;
4414	unsigned int objnr;
4415	unsigned long offset;
4416
4417	/* Find and validate object. */
4418	cachep = page->slab_cache;
4419	objnr = obj_to_index(cachep, page, (void *)ptr);
4420	BUG_ON(objnr >= cachep->num);
4421
4422	/* Find offset within object. */
4423	offset = ptr - index_to_obj(cachep, page, objnr) - obj_offset(cachep);
4424
4425	/* Allow address range falling entirely within usercopy region. */
4426	if (offset >= cachep->useroffset &&
4427	    offset - cachep->useroffset <= cachep->usersize &&
4428	    n <= cachep->useroffset - offset + cachep->usersize)
4429		return;
4430
4431	/*
4432	 * If the copy is still within the allocated object, produce
4433	 * a warning instead of rejecting the copy. This is intended
4434	 * to be a temporary method to find any missing usercopy
4435	 * whitelists.
4436	 */
4437	if (usercopy_fallback &&
4438	    offset <= cachep->object_size &&
4439	    n <= cachep->object_size - offset) {
4440		usercopy_warn("SLAB object", cachep->name, to_user, offset, n);
4441		return;
4442	}
4443
4444	usercopy_abort("SLAB object", cachep->name, to_user, offset, n);
4445}
4446#endif /* CONFIG_HARDENED_USERCOPY */
4447
4448/**
4449 * ksize - get the actual amount of memory allocated for a given object
4450 * @objp: Pointer to the object
4451 *
4452 * kmalloc may internally round up allocations and return more memory
4453 * than requested. ksize() can be used to determine the actual amount of
4454 * memory allocated. The caller may use this additional memory, even though
4455 * a smaller amount of memory was initially specified with the kmalloc call.
4456 * The caller must guarantee that objp points to a valid object previously
4457 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4458 * must not be freed during the duration of the call.
4459 */
4460size_t ksize(const void *objp)
4461{
4462	size_t size;
4463
4464	BUG_ON(!objp);
4465	if (unlikely(objp == ZERO_SIZE_PTR))
4466		return 0;
4467
4468	size = virt_to_cache(objp)->object_size;
4469	/* We assume that ksize callers could use the whole allocated area,
4470	 * so we need to unpoison this area.
4471	 */
4472	kasan_unpoison_shadow(objp, size);
4473
4474	return size;
4475}
4476EXPORT_SYMBOL(ksize);