Linux Audio

Check our new training course

Loading...
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);
v5.14.15
   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/kfence.h>
 104#include	<linux/cpu.h>
 105#include	<linux/__KEEPIDENTS__BD.h>
 106#include	<linux/__KEEPIDENTS__BE.h>
 107#include	<linux/rcupdate.h>
 108#include	<linux/__KEEPIDENTS__BF.h>
 109#include	<linux/__KEEPIDENTS__BG.h>
 110#include	<linux/__KEEPIDENTS__BH.h>
 111#include	<linux/kmemleak.h>
 112#include	<linux/__KEEPIDENTS__BI.h>
 113#include	<linux/__KEEPIDENTS__BJ.h>
 114#include	<linux/__KEEPIDENTS__CA-__KEEPIDENTS__CB.h>
 115#include	<linux/__KEEPIDENTS__CC.h>
 116#include	<linux/reciprocal_div.h>
 117#include	<linux/debugobjects.h>
 118#include	<linux/__KEEPIDENTS__CD.h>
 119#include	<linux/__KEEPIDENTS__CE.h>
 120#include	<linux/__KEEPIDENTS__CF/task_stack.h>
 121
 122#include	<net/__KEEPIDENTS__CG.h>
 123
 124#include	<asm/cacheflush.h>
 125#include	<asm/tlbflush.h>
 126#include	<asm/page.h>
 127
 128#include <trace/events/kmem.h>
 129
 130#include	"internal.h"
 131
 132#include	"slab.h"
 133
 134/*
 135 * DEBUG	- 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
 136 *		  0 for faster, smaller code (especially in the critical paths).
 137 *
 138 * STATS	- 1 to collect stats for /proc/slabinfo.
 139 *		  0 for faster, smaller code (especially in the critical paths).
 140 *
 141 * FORCED_DEBUG	- 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
 142 */
 143
 144#ifdef CONFIG_DEBUG_SLAB
 145#define	DEBUG		1
 146#define	STATS		1
 147#define	FORCED_DEBUG	1
 148#else
 149#define	DEBUG		0
 150#define	STATS		0
 151#define	FORCED_DEBUG	0
 152#endif
 153
 154/* Shouldn't this be in a header file somewhere? */
 155#define	BYTES_PER_WORD		sizeof(void *)
 156#define	REDZONE_ALIGN		max(BYTES_PER_WORD, __alignof__(unsigned long long))
 157
 158#ifndef ARCH_KMALLOC_FLAGS
 159#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
 160#endif
 161
 162#define FREELIST_BYTE_INDEX (((PAGE_SIZE >> BITS_PER_BYTE) \
 163				<= SLAB_OBJ_MIN_SIZE) ? 1 : 0)
 164
 165#if FREELIST_BYTE_INDEX
 166typedef unsigned char freelist_idx_t;
 167#else
 168typedef unsigned short freelist_idx_t;
 169#endif
 170
 171#define SLAB_OBJ_MAX_NUM ((1 << sizeof(freelist_idx_t) * BITS_PER_BYTE) - 1)
 172
 173/*
 174 * struct array_cache
 175 *
 176 * Purpose:
 177 * - LIFO ordering, to hand out cache-warm objects from _alloc
 178 * - reduce the number of linked list operations
 179 * - reduce spinlock operations
 180 *
 181 * The limit is stored in the per-cpu structure to reduce the data cache
 182 * footprint.
 183 *
 184 */
 185struct array_cache {
 186	unsigned int avail;
 187	unsigned int limit;
 188	unsigned int batchcount;
 189	unsigned int touched;
 190	void *entry[];	/*
 191			 * Must have this definition in here for the proper
 192			 * alignment of array_cache. Also simplifies accessing
 193			 * the entries.
 194			 */
 195};
 196
 197struct alien_cache {
 198	spinlock_t lock;
 199	struct array_cache ac;
 200};
 201
 202/*
 203 * Need this for bootstrapping a per node allocator.
 204 */
 205#define NUM_INIT_LISTS (2 * MAX_NUMNODES)
 206static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
 207#define	CACHE_CACHE 0
 208#define	SIZE_NODE (MAX_NUMNODES)
 209
 210static int drain_freelist(struct kmem_cache *cache,
 211			struct kmem_cache_node *n, int tofree);
 212static void free_block(struct kmem_cache *cachep, void **objpp, int len,
 213			int node, struct list_head *list);
 214static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list);
 215static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
 216static void cache_reap(struct work_struct *unused);
 217
 218static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
 219						void **list);
 220static inline void fixup_slab_list(struct kmem_cache *cachep,
 221				struct kmem_cache_node *n, struct page *page,
 222				void **list);
 223static int slab_early_init = 1;
 224
 225#define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
 226
 227static void kmem_cache_node_init(struct kmem_cache_node *parent)
 228{
 229	INIT_LIST_HEAD(&parent->slabs_full);
 230	INIT_LIST_HEAD(&parent->slabs_partial);
 231	INIT_LIST_HEAD(&parent->slabs_free);
 232	parent->total_slabs = 0;
 233	parent->free_slabs = 0;
 234	parent->shared = NULL;
 235	parent->alien = NULL;
 236	parent->colour_next = 0;
 237	spin_lock_init(&parent->list_lock);
 238	parent->free_objects = 0;
 239	parent->free_touched = 0;
 240}
 241
 242#define MAKE_LIST(cachep, listp, slab, nodeid)				\
 243	do {								\
 244		INIT_LIST_HEAD(listp);					\
 245		list_splice(&get_node(cachep, nodeid)->slab, listp);	\
 246	} while (0)
 247
 248#define	MAKE_ALL_LISTS(cachep, ptr, nodeid)				\
 249	do {								\
 250	MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid);	\
 251	MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
 252	MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid);	\
 253	} while (0)
 254
 255#define CFLGS_OBJFREELIST_SLAB	((slab_flags_t __force)0x40000000U)
 256#define CFLGS_OFF_SLAB		((slab_flags_t __force)0x80000000U)
 257#define	OBJFREELIST_SLAB(x)	((x)->flags & CFLGS_OBJFREELIST_SLAB)
 258#define	OFF_SLAB(x)	((x)->flags & CFLGS_OFF_SLAB)
 259
 260#define BATCHREFILL_LIMIT	16
 261/*
 262 * Optimization question: fewer reaps means less probability for unnecessary
 263 * cpucache drain/refill cycles.
 264 *
 265 * OTOH the cpuarrays can contain lots of objects,
 266 * which could lock up otherwise freeable slabs.
 267 */
 268#define REAPTIMEOUT_AC		(2*HZ)
 269#define REAPTIMEOUT_NODE	(4*HZ)
 270
 271#if STATS
 272#define	STATS_INC_ACTIVE(x)	((x)->num_active++)
 273#define	STATS_DEC_ACTIVE(x)	((x)->num_active--)
 274#define	STATS_INC_ALLOCED(x)	((x)->num_allocations++)
 275#define	STATS_INC_GROWN(x)	((x)->grown++)
 276#define	STATS_ADD_REAPED(x, y)	((x)->reaped += (y))
 277#define	STATS_SET_HIGH(x)						\
 278	do {								\
 279		if ((x)->num_active > (x)->high_mark)			\
 280			(x)->high_mark = (x)->num_active;		\
 281	} while (0)
 282#define	STATS_INC_ERR(x)	((x)->errors++)
 283#define	STATS_INC_NODEALLOCS(x)	((x)->node_allocs++)
 284#define	STATS_INC_NODEFREES(x)	((x)->node_frees++)
 285#define STATS_INC_ACOVERFLOW(x)   ((x)->node_overflow++)
 286#define	STATS_SET_FREEABLE(x, i)					\
 287	do {								\
 288		if ((x)->max_freeable < i)				\
 289			(x)->max_freeable = i;				\
 290	} while (0)
 291#define STATS_INC_ALLOCHIT(x)	atomic_inc(&(x)->allochit)
 292#define STATS_INC_ALLOCMISS(x)	atomic_inc(&(x)->allocmiss)
 293#define STATS_INC_FREEHIT(x)	atomic_inc(&(x)->freehit)
 294#define STATS_INC_FREEMISS(x)	atomic_inc(&(x)->freemiss)
 295#else
 296#define	STATS_INC_ACTIVE(x)	do { } while (0)
 297#define	STATS_DEC_ACTIVE(x)	do { } while (0)
 298#define	STATS_INC_ALLOCED(x)	do { } while (0)
 299#define	STATS_INC_GROWN(x)	do { } while (0)
 300#define	STATS_ADD_REAPED(x, y)	do { (void)(y); } while (0)
 301#define	STATS_SET_HIGH(x)	do { } while (0)
 302#define	STATS_INC_ERR(x)	do { } while (0)
 303#define	STATS_INC_NODEALLOCS(x)	do { } while (0)
 304#define	STATS_INC_NODEFREES(x)	do { } while (0)
 305#define STATS_INC_ACOVERFLOW(x)   do { } while (0)
 306#define	STATS_SET_FREEABLE(x, i) do { } while (0)
 307#define STATS_INC_ALLOCHIT(x)	do { } while (0)
 308#define STATS_INC_ALLOCMISS(x)	do { } while (0)
 309#define STATS_INC_FREEHIT(x)	do { } while (0)
 310#define STATS_INC_FREEMISS(x)	do { } while (0)
 311#endif
 312
 313#if DEBUG
 314
 315/*
 316 * memory layout of objects:
 317 * 0		: objp
 318 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
 319 * 		the end of an object is aligned with the end of the real
 320 * 		allocation. Catches writes behind the end of the allocation.
 321 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
 322 * 		redzone word.
 323 * cachep->obj_offset: The real object.
 324 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
 325 * cachep->size - 1* BYTES_PER_WORD: last caller address
 326 *					[BYTES_PER_WORD long]
 327 */
 328static int obj_offset(struct kmem_cache *cachep)
 329{
 330	return cachep->obj_offset;
 331}
 332
 333static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
 334{
 335	BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
 336	return (unsigned long long *) (objp + obj_offset(cachep) -
 337				      sizeof(unsigned long long));
 338}
 339
 340static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
 341{
 342	BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
 343	if (cachep->flags & SLAB_STORE_USER)
 344		return (unsigned long long *)(objp + cachep->size -
 345					      sizeof(unsigned long long) -
 346					      REDZONE_ALIGN);
 347	return (unsigned long long *) (objp + cachep->size -
 348				       sizeof(unsigned long long));
 349}
 350
 351static void **dbg_userword(struct kmem_cache *cachep, void *objp)
 352{
 353	BUG_ON(!(cachep->flags & SLAB_STORE_USER));
 354	return (void **)(objp + cachep->size - BYTES_PER_WORD);
 355}
 356
 357#else
 358
 359#define obj_offset(x)			0
 360#define dbg_redzone1(cachep, objp)	({BUG(); (unsigned long long *)NULL;})
 361#define dbg_redzone2(cachep, objp)	({BUG(); (unsigned long long *)NULL;})
 362#define dbg_userword(cachep, objp)	({BUG(); (void **)NULL;})
 363
 364#endif
 365
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 366/*
 367 * Do not go above this order unless 0 objects fit into the slab or
 368 * overridden on the command line.
 369 */
 370#define	SLAB_MAX_ORDER_HI	1
 371#define	SLAB_MAX_ORDER_LO	0
 372static int slab_max_order = SLAB_MAX_ORDER_LO;
 373static bool slab_max_order_set __initdata;
 374
 
 
 
 
 
 
 375static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
 376				 unsigned int idx)
 377{
 378	return page->s_mem + cache->size * idx;
 379}
 380
 
 
 
 
 
 
 
 
 
 
 
 
 
 381#define BOOT_CPUCACHE_ENTRIES	1
 382/* internal cache of cache description objs */
 383static struct kmem_cache kmem_cache_boot = {
 384	.batchcount = 1,
 385	.limit = BOOT_CPUCACHE_ENTRIES,
 386	.shared = 1,
 387	.size = sizeof(struct kmem_cache),
 388	.name = "kmem_cache",
 389};
 390
 391static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
 392
 393static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
 394{
 395	return this_cpu_ptr(cachep->cpu_cache);
 396}
 397
 398/*
 399 * Calculate the number of objects and left-over bytes for a given buffer size.
 400 */
 401static unsigned int cache_estimate(unsigned long gfporder, size_t buffer_size,
 402		slab_flags_t flags, size_t *left_over)
 403{
 404	unsigned int num;
 405	size_t slab_size = PAGE_SIZE << gfporder;
 406
 407	/*
 408	 * The slab management structure can be either off the slab or
 409	 * on it. For the latter case, the memory allocated for a
 410	 * slab is used for:
 411	 *
 412	 * - @buffer_size bytes for each object
 413	 * - One freelist_idx_t for each object
 414	 *
 415	 * We don't need to consider alignment of freelist because
 416	 * freelist will be at the end of slab page. The objects will be
 417	 * at the correct alignment.
 418	 *
 419	 * If the slab management structure is off the slab, then the
 420	 * alignment will already be calculated into the size. Because
 421	 * the slabs are all pages aligned, the objects will be at the
 422	 * correct alignment when allocated.
 423	 */
 424	if (flags & (CFLGS_OBJFREELIST_SLAB | CFLGS_OFF_SLAB)) {
 425		num = slab_size / buffer_size;
 426		*left_over = slab_size % buffer_size;
 427	} else {
 428		num = slab_size / (buffer_size + sizeof(freelist_idx_t));
 429		*left_over = slab_size %
 430			(buffer_size + sizeof(freelist_idx_t));
 431	}
 432
 433	return num;
 434}
 435
 436#if DEBUG
 437#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
 438
 439static void __slab_error(const char *function, struct kmem_cache *cachep,
 440			char *msg)
 441{
 442	pr_err("slab error in %s(): cache `%s': %s\n",
 443	       function, cachep->name, msg);
 444	dump_stack();
 445	add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
 446}
 447#endif
 448
 449/*
 450 * By default on NUMA we use alien caches to stage the freeing of
 451 * objects allocated from other nodes. This causes massive memory
 452 * inefficiencies when using fake NUMA setup to split memory into a
 453 * large number of small nodes, so it can be disabled on the command
 454 * line
 455  */
 456
 457static int use_alien_caches __read_mostly = 1;
 458static int __init noaliencache_setup(char *s)
 459{
 460	use_alien_caches = 0;
 461	return 1;
 462}
 463__setup("noaliencache", noaliencache_setup);
 464
 465static int __init slab_max_order_setup(char *str)
 466{
 467	get_option(&str, &slab_max_order);
 468	slab_max_order = slab_max_order < 0 ? 0 :
 469				min(slab_max_order, MAX_ORDER - 1);
 470	slab_max_order_set = true;
 471
 472	return 1;
 473}
 474__setup("slab_max_order=", slab_max_order_setup);
 475
 476#ifdef CONFIG_NUMA
 477/*
 478 * Special reaping functions for NUMA systems called from cache_reap().
 479 * These take care of doing round robin flushing of alien caches (containing
 480 * objects freed on different nodes from which they were allocated) and the
 481 * flushing of remote pcps by calling drain_node_pages.
 482 */
 483static DEFINE_PER_CPU(unsigned long, slab_reap_node);
 484
 485static void init_reap_node(int cpu)
 486{
 487	per_cpu(slab_reap_node, cpu) = next_node_in(cpu_to_mem(cpu),
 488						    node_online_map);
 489}
 490
 491static void next_reap_node(void)
 492{
 493	int node = __this_cpu_read(slab_reap_node);
 494
 495	node = next_node_in(node, node_online_map);
 496	__this_cpu_write(slab_reap_node, node);
 497}
 498
 499#else
 500#define init_reap_node(cpu) do { } while (0)
 501#define next_reap_node(void) do { } while (0)
 502#endif
 503
 504/*
 505 * Initiate the reap timer running on the target CPU.  We run at around 1 to 2Hz
 506 * via the workqueue/eventd.
 507 * Add the CPU number into the expiration time to minimize the possibility of
 508 * the CPUs getting into lockstep and contending for the global cache chain
 509 * lock.
 510 */
 511static void start_cpu_timer(int cpu)
 512{
 513	struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
 514
 515	if (reap_work->work.func == NULL) {
 516		init_reap_node(cpu);
 517		INIT_DEFERRABLE_WORK(reap_work, cache_reap);
 518		schedule_delayed_work_on(cpu, reap_work,
 519					__round_jiffies_relative(HZ, cpu));
 520	}
 521}
 522
 523static void init_arraycache(struct array_cache *ac, int limit, int batch)
 524{
 
 
 
 
 
 
 
 
 525	if (ac) {
 526		ac->avail = 0;
 527		ac->limit = limit;
 528		ac->batchcount = batch;
 529		ac->touched = 0;
 530	}
 531}
 532
 533static struct array_cache *alloc_arraycache(int node, int entries,
 534					    int batchcount, gfp_t gfp)
 535{
 536	size_t memsize = sizeof(void *) * entries + sizeof(struct array_cache);
 537	struct array_cache *ac = NULL;
 538
 539	ac = kmalloc_node(memsize, gfp, node);
 540	/*
 541	 * The array_cache structures contain pointers to free object.
 542	 * However, when such objects are allocated or transferred to another
 543	 * cache the pointers are not cleared and they could be counted as
 544	 * valid references during a kmemleak scan. Therefore, kmemleak must
 545	 * not scan such objects.
 546	 */
 547	kmemleak_no_scan(ac);
 548	init_arraycache(ac, entries, batchcount);
 549	return ac;
 550}
 551
 552static noinline void cache_free_pfmemalloc(struct kmem_cache *cachep,
 553					struct page *page, void *objp)
 554{
 555	struct kmem_cache_node *n;
 556	int page_node;
 557	LIST_HEAD(list);
 558
 559	page_node = page_to_nid(page);
 560	n = get_node(cachep, page_node);
 561
 562	spin_lock(&n->list_lock);
 563	free_block(cachep, &objp, 1, page_node, &list);
 564	spin_unlock(&n->list_lock);
 565
 566	slabs_destroy(cachep, &list);
 567}
 568
 569/*
 570 * Transfer objects in one arraycache to another.
 571 * Locking must be handled by the caller.
 572 *
 573 * Return the number of entries transferred.
 574 */
 575static int transfer_objects(struct array_cache *to,
 576		struct array_cache *from, unsigned int max)
 577{
 578	/* Figure out how many entries to transfer */
 579	int nr = min3(from->avail, max, to->limit - to->avail);
 580
 581	if (!nr)
 582		return 0;
 583
 584	memcpy(to->entry + to->avail, from->entry + from->avail - nr,
 585			sizeof(void *) *nr);
 586
 587	from->avail -= nr;
 588	to->avail += nr;
 589	return nr;
 590}
 591
 592/* &alien->lock must be held by alien callers. */
 593static __always_inline void __free_one(struct array_cache *ac, void *objp)
 594{
 595	/* Avoid trivial double-free. */
 596	if (IS_ENABLED(CONFIG_SLAB_FREELIST_HARDENED) &&
 597	    WARN_ON_ONCE(ac->avail > 0 && ac->entry[ac->avail - 1] == objp))
 598		return;
 599	ac->entry[ac->avail++] = objp;
 600}
 601
 602#ifndef CONFIG_NUMA
 603
 604#define drain_alien_cache(cachep, alien) do { } while (0)
 605#define reap_alien(cachep, n) do { } while (0)
 606
 607static inline struct alien_cache **alloc_alien_cache(int node,
 608						int limit, gfp_t gfp)
 609{
 610	return NULL;
 611}
 612
 613static inline void free_alien_cache(struct alien_cache **ac_ptr)
 614{
 615}
 616
 617static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
 618{
 619	return 0;
 620}
 621
 622static inline void *alternate_node_alloc(struct kmem_cache *cachep,
 623		gfp_t flags)
 624{
 625	return NULL;
 626}
 627
 628static inline void *____cache_alloc_node(struct kmem_cache *cachep,
 629		 gfp_t flags, int nodeid)
 630{
 631	return NULL;
 632}
 633
 634static inline gfp_t gfp_exact_node(gfp_t flags)
 635{
 636	return flags & ~__GFP_NOFAIL;
 637}
 638
 639#else	/* CONFIG_NUMA */
 640
 641static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
 642static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
 643
 644static struct alien_cache *__alloc_alien_cache(int node, int entries,
 645						int batch, gfp_t gfp)
 646{
 647	size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache);
 648	struct alien_cache *alc = NULL;
 649
 650	alc = kmalloc_node(memsize, gfp, node);
 651	if (alc) {
 652		kmemleak_no_scan(alc);
 653		init_arraycache(&alc->ac, entries, batch);
 654		spin_lock_init(&alc->lock);
 655	}
 656	return alc;
 657}
 658
 659static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
 660{
 661	struct alien_cache **alc_ptr;
 
 662	int i;
 663
 664	if (limit > 1)
 665		limit = 12;
 666	alc_ptr = kcalloc_node(nr_node_ids, sizeof(void *), gfp, node);
 667	if (!alc_ptr)
 668		return NULL;
 669
 670	for_each_node(i) {
 671		if (i == node || !node_online(i))
 672			continue;
 673		alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp);
 674		if (!alc_ptr[i]) {
 675			for (i--; i >= 0; i--)
 676				kfree(alc_ptr[i]);
 677			kfree(alc_ptr);
 678			return NULL;
 679		}
 680	}
 681	return alc_ptr;
 682}
 683
 684static void free_alien_cache(struct alien_cache **alc_ptr)
 685{
 686	int i;
 687
 688	if (!alc_ptr)
 689		return;
 690	for_each_node(i)
 691	    kfree(alc_ptr[i]);
 692	kfree(alc_ptr);
 693}
 694
 695static void __drain_alien_cache(struct kmem_cache *cachep,
 696				struct array_cache *ac, int node,
 697				struct list_head *list)
 698{
 699	struct kmem_cache_node *n = get_node(cachep, node);
 700
 701	if (ac->avail) {
 702		spin_lock(&n->list_lock);
 703		/*
 704		 * Stuff objects into the remote nodes shared array first.
 705		 * That way we could avoid the overhead of putting the objects
 706		 * into the free lists and getting them back later.
 707		 */
 708		if (n->shared)
 709			transfer_objects(n->shared, ac, ac->limit);
 710
 711		free_block(cachep, ac->entry, ac->avail, node, list);
 712		ac->avail = 0;
 713		spin_unlock(&n->list_lock);
 714	}
 715}
 716
 717/*
 718 * Called from cache_reap() to regularly drain alien caches round robin.
 719 */
 720static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
 721{
 722	int node = __this_cpu_read(slab_reap_node);
 723
 724	if (n->alien) {
 725		struct alien_cache *alc = n->alien[node];
 726		struct array_cache *ac;
 727
 728		if (alc) {
 729			ac = &alc->ac;
 730			if (ac->avail && spin_trylock_irq(&alc->lock)) {
 731				LIST_HEAD(list);
 732
 733				__drain_alien_cache(cachep, ac, node, &list);
 734				spin_unlock_irq(&alc->lock);
 735				slabs_destroy(cachep, &list);
 736			}
 737		}
 738	}
 739}
 740
 741static void drain_alien_cache(struct kmem_cache *cachep,
 742				struct alien_cache **alien)
 743{
 744	int i = 0;
 745	struct alien_cache *alc;
 746	struct array_cache *ac;
 747	unsigned long flags;
 748
 749	for_each_online_node(i) {
 750		alc = alien[i];
 751		if (alc) {
 752			LIST_HEAD(list);
 753
 754			ac = &alc->ac;
 755			spin_lock_irqsave(&alc->lock, flags);
 756			__drain_alien_cache(cachep, ac, i, &list);
 757			spin_unlock_irqrestore(&alc->lock, flags);
 758			slabs_destroy(cachep, &list);
 759		}
 760	}
 761}
 762
 763static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
 764				int node, int page_node)
 765{
 766	struct kmem_cache_node *n;
 767	struct alien_cache *alien = NULL;
 768	struct array_cache *ac;
 769	LIST_HEAD(list);
 770
 771	n = get_node(cachep, node);
 772	STATS_INC_NODEFREES(cachep);
 773	if (n->alien && n->alien[page_node]) {
 774		alien = n->alien[page_node];
 775		ac = &alien->ac;
 776		spin_lock(&alien->lock);
 777		if (unlikely(ac->avail == ac->limit)) {
 778			STATS_INC_ACOVERFLOW(cachep);
 779			__drain_alien_cache(cachep, ac, page_node, &list);
 780		}
 781		__free_one(ac, objp);
 782		spin_unlock(&alien->lock);
 783		slabs_destroy(cachep, &list);
 784	} else {
 785		n = get_node(cachep, page_node);
 786		spin_lock(&n->list_lock);
 787		free_block(cachep, &objp, 1, page_node, &list);
 788		spin_unlock(&n->list_lock);
 789		slabs_destroy(cachep, &list);
 790	}
 791	return 1;
 792}
 793
 794static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
 795{
 796	int page_node = page_to_nid(virt_to_page(objp));
 797	int node = numa_mem_id();
 798	/*
 799	 * Make sure we are not freeing a object from another node to the array
 800	 * cache on this cpu.
 801	 */
 802	if (likely(node == page_node))
 803		return 0;
 804
 805	return __cache_free_alien(cachep, objp, node, page_node);
 806}
 807
 808/*
 809 * Construct gfp mask to allocate from a specific node but do not reclaim or
 810 * warn about failures.
 811 */
 812static inline gfp_t gfp_exact_node(gfp_t flags)
 813{
 814	return (flags | __GFP_THISNODE | __GFP_NOWARN) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
 815}
 816#endif
 817
 818static int init_cache_node(struct kmem_cache *cachep, int node, gfp_t gfp)
 819{
 820	struct kmem_cache_node *n;
 821
 822	/*
 823	 * Set up the kmem_cache_node for cpu before we can
 824	 * begin anything. Make sure some other cpu on this
 825	 * node has not already allocated this
 826	 */
 827	n = get_node(cachep, node);
 828	if (n) {
 829		spin_lock_irq(&n->list_lock);
 830		n->free_limit = (1 + nr_cpus_node(node)) * cachep->batchcount +
 831				cachep->num;
 832		spin_unlock_irq(&n->list_lock);
 833
 834		return 0;
 835	}
 836
 837	n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
 838	if (!n)
 839		return -ENOMEM;
 840
 841	kmem_cache_node_init(n);
 842	n->next_reap = jiffies + REAPTIMEOUT_NODE +
 843		    ((unsigned long)cachep) % REAPTIMEOUT_NODE;
 844
 845	n->free_limit =
 846		(1 + nr_cpus_node(node)) * cachep->batchcount + cachep->num;
 847
 848	/*
 849	 * The kmem_cache_nodes don't come and go as CPUs
 850	 * come and go.  slab_mutex is sufficient
 851	 * protection here.
 852	 */
 853	cachep->node[node] = n;
 854
 855	return 0;
 856}
 857
 858#if (defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)) || defined(CONFIG_SMP)
 859/*
 860 * Allocates and initializes node for a node on each slab cache, used for
 861 * either memory or cpu hotplug.  If memory is being hot-added, the kmem_cache_node
 862 * will be allocated off-node since memory is not yet online for the new node.
 863 * When hotplugging memory or a cpu, existing node are not replaced if
 864 * already in use.
 865 *
 866 * Must hold slab_mutex.
 867 */
 868static int init_cache_node_node(int node)
 869{
 870	int ret;
 871	struct kmem_cache *cachep;
 872
 873	list_for_each_entry(cachep, &slab_caches, list) {
 874		ret = init_cache_node(cachep, node, GFP_KERNEL);
 875		if (ret)
 876			return ret;
 877	}
 878
 879	return 0;
 880}
 881#endif
 882
 883static int setup_kmem_cache_node(struct kmem_cache *cachep,
 884				int node, gfp_t gfp, bool force_change)
 885{
 886	int ret = -ENOMEM;
 887	struct kmem_cache_node *n;
 888	struct array_cache *old_shared = NULL;
 889	struct array_cache *new_shared = NULL;
 890	struct alien_cache **new_alien = NULL;
 891	LIST_HEAD(list);
 892
 893	if (use_alien_caches) {
 894		new_alien = alloc_alien_cache(node, cachep->limit, gfp);
 895		if (!new_alien)
 896			goto fail;
 897	}
 898
 899	if (cachep->shared) {
 900		new_shared = alloc_arraycache(node,
 901			cachep->shared * cachep->batchcount, 0xbaadf00d, gfp);
 902		if (!new_shared)
 903			goto fail;
 904	}
 905
 906	ret = init_cache_node(cachep, node, gfp);
 907	if (ret)
 908		goto fail;
 909
 910	n = get_node(cachep, node);
 911	spin_lock_irq(&n->list_lock);
 912	if (n->shared && force_change) {
 913		free_block(cachep, n->shared->entry,
 914				n->shared->avail, node, &list);
 915		n->shared->avail = 0;
 916	}
 917
 918	if (!n->shared || force_change) {
 919		old_shared = n->shared;
 920		n->shared = new_shared;
 921		new_shared = NULL;
 922	}
 923
 924	if (!n->alien) {
 925		n->alien = new_alien;
 926		new_alien = NULL;
 927	}
 928
 929	spin_unlock_irq(&n->list_lock);
 930	slabs_destroy(cachep, &list);
 931
 932	/*
 933	 * To protect lockless access to n->shared during irq disabled context.
 934	 * If n->shared isn't NULL in irq disabled context, accessing to it is
 935	 * guaranteed to be valid until irq is re-enabled, because it will be
 936	 * freed after synchronize_rcu().
 937	 */
 938	if (old_shared && force_change)
 939		synchronize_rcu();
 940
 941fail:
 942	kfree(old_shared);
 943	kfree(new_shared);
 944	free_alien_cache(new_alien);
 945
 946	return ret;
 947}
 948
 949#ifdef CONFIG_SMP
 950
 951static void cpuup_canceled(long cpu)
 952{
 953	struct kmem_cache *cachep;
 954	struct kmem_cache_node *n = NULL;
 955	int node = cpu_to_mem(cpu);
 956	const struct cpumask *mask = cpumask_of_node(node);
 957
 958	list_for_each_entry(cachep, &slab_caches, list) {
 959		struct array_cache *nc;
 960		struct array_cache *shared;
 961		struct alien_cache **alien;
 962		LIST_HEAD(list);
 963
 964		n = get_node(cachep, node);
 965		if (!n)
 966			continue;
 967
 968		spin_lock_irq(&n->list_lock);
 969
 970		/* Free limit for this kmem_cache_node */
 971		n->free_limit -= cachep->batchcount;
 972
 973		/* cpu is dead; no one can alloc from it. */
 974		nc = per_cpu_ptr(cachep->cpu_cache, cpu);
 975		free_block(cachep, nc->entry, nc->avail, node, &list);
 976		nc->avail = 0;
 
 
 977
 978		if (!cpumask_empty(mask)) {
 979			spin_unlock_irq(&n->list_lock);
 980			goto free_slab;
 981		}
 982
 983		shared = n->shared;
 984		if (shared) {
 985			free_block(cachep, shared->entry,
 986				   shared->avail, node, &list);
 987			n->shared = NULL;
 988		}
 989
 990		alien = n->alien;
 991		n->alien = NULL;
 992
 993		spin_unlock_irq(&n->list_lock);
 994
 995		kfree(shared);
 996		if (alien) {
 997			drain_alien_cache(cachep, alien);
 998			free_alien_cache(alien);
 999		}
1000
1001free_slab:
1002		slabs_destroy(cachep, &list);
1003	}
1004	/*
1005	 * In the previous loop, all the objects were freed to
1006	 * the respective cache's slabs,  now we can go ahead and
1007	 * shrink each nodelist to its limit.
1008	 */
1009	list_for_each_entry(cachep, &slab_caches, list) {
1010		n = get_node(cachep, node);
1011		if (!n)
1012			continue;
1013		drain_freelist(cachep, n, INT_MAX);
1014	}
1015}
1016
1017static int cpuup_prepare(long cpu)
1018{
1019	struct kmem_cache *cachep;
1020	int node = cpu_to_mem(cpu);
1021	int err;
1022
1023	/*
1024	 * We need to do this right in the beginning since
1025	 * alloc_arraycache's are going to use this list.
1026	 * kmalloc_node allows us to add the slab to the right
1027	 * kmem_cache_node and not this cpu's kmem_cache_node
1028	 */
1029	err = init_cache_node_node(node);
1030	if (err < 0)
1031		goto bad;
1032
1033	/*
1034	 * Now we can go ahead with allocating the shared arrays and
1035	 * array caches
1036	 */
1037	list_for_each_entry(cachep, &slab_caches, list) {
1038		err = setup_kmem_cache_node(cachep, node, GFP_KERNEL, false);
1039		if (err)
1040			goto bad;
1041	}
1042
1043	return 0;
1044bad:
1045	cpuup_canceled(cpu);
1046	return -ENOMEM;
1047}
1048
1049int slab_prepare_cpu(unsigned int cpu)
1050{
1051	int err;
1052
1053	mutex_lock(&slab_mutex);
1054	err = cpuup_prepare(cpu);
1055	mutex_unlock(&slab_mutex);
1056	return err;
1057}
1058
1059/*
1060 * This is called for a failed online attempt and for a successful
1061 * offline.
1062 *
1063 * Even if all the cpus of a node are down, we don't free the
1064 * kmem_cache_node of any cache. This to avoid a race between cpu_down, and
1065 * a kmalloc allocation from another cpu for memory from the node of
1066 * the cpu going down.  The kmem_cache_node structure is usually allocated from
1067 * kmem_cache_create() and gets destroyed at kmem_cache_destroy().
1068 */
1069int slab_dead_cpu(unsigned int cpu)
1070{
1071	mutex_lock(&slab_mutex);
1072	cpuup_canceled(cpu);
1073	mutex_unlock(&slab_mutex);
1074	return 0;
1075}
1076#endif
1077
1078static int slab_online_cpu(unsigned int cpu)
1079{
1080	start_cpu_timer(cpu);
1081	return 0;
1082}
1083
1084static int slab_offline_cpu(unsigned int cpu)
1085{
1086	/*
1087	 * Shutdown cache reaper. Note that the slab_mutex is held so
1088	 * that if cache_reap() is invoked it cannot do anything
1089	 * expensive but will only modify reap_work and reschedule the
1090	 * timer.
1091	 */
1092	cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1093	/* Now the cache_reaper is guaranteed to be not running. */
1094	per_cpu(slab_reap_work, cpu).work.func = NULL;
1095	return 0;
1096}
1097
1098#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1099/*
1100 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1101 * Returns -EBUSY if all objects cannot be drained so that the node is not
1102 * removed.
1103 *
1104 * Must hold slab_mutex.
1105 */
1106static int __meminit drain_cache_node_node(int node)
1107{
1108	struct kmem_cache *cachep;
1109	int ret = 0;
1110
1111	list_for_each_entry(cachep, &slab_caches, list) {
1112		struct kmem_cache_node *n;
1113
1114		n = get_node(cachep, node);
1115		if (!n)
1116			continue;
1117
1118		drain_freelist(cachep, n, INT_MAX);
1119
1120		if (!list_empty(&n->slabs_full) ||
1121		    !list_empty(&n->slabs_partial)) {
1122			ret = -EBUSY;
1123			break;
1124		}
1125	}
1126	return ret;
1127}
1128
1129static int __meminit slab_memory_callback(struct notifier_block *self,
1130					unsigned long action, void *arg)
1131{
1132	struct memory_notify *mnb = arg;
1133	int ret = 0;
1134	int nid;
1135
1136	nid = mnb->status_change_nid;
1137	if (nid < 0)
1138		goto out;
1139
1140	switch (action) {
1141	case MEM_GOING_ONLINE:
1142		mutex_lock(&slab_mutex);
1143		ret = init_cache_node_node(nid);
1144		mutex_unlock(&slab_mutex);
1145		break;
1146	case MEM_GOING_OFFLINE:
1147		mutex_lock(&slab_mutex);
1148		ret = drain_cache_node_node(nid);
1149		mutex_unlock(&slab_mutex);
1150		break;
1151	case MEM_ONLINE:
1152	case MEM_OFFLINE:
1153	case MEM_CANCEL_ONLINE:
1154	case MEM_CANCEL_OFFLINE:
1155		break;
1156	}
1157out:
1158	return notifier_from_errno(ret);
1159}
1160#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1161
1162/*
1163 * swap the static kmem_cache_node with kmalloced memory
1164 */
1165static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1166				int nodeid)
1167{
1168	struct kmem_cache_node *ptr;
1169
1170	ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1171	BUG_ON(!ptr);
1172
1173	memcpy(ptr, list, sizeof(struct kmem_cache_node));
1174	/*
1175	 * Do not assume that spinlocks can be initialized via memcpy:
1176	 */
1177	spin_lock_init(&ptr->list_lock);
1178
1179	MAKE_ALL_LISTS(cachep, ptr, nodeid);
1180	cachep->node[nodeid] = ptr;
1181}
1182
1183/*
1184 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1185 * size of kmem_cache_node.
1186 */
1187static void __init set_up_node(struct kmem_cache *cachep, int index)
1188{
1189	int node;
1190
1191	for_each_online_node(node) {
1192		cachep->node[node] = &init_kmem_cache_node[index + node];
1193		cachep->node[node]->next_reap = jiffies +
1194		    REAPTIMEOUT_NODE +
1195		    ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1196	}
1197}
1198
1199/*
1200 * Initialisation.  Called after the page allocator have been initialised and
1201 * before smp_init().
1202 */
1203void __init kmem_cache_init(void)
1204{
1205	int i;
1206
 
 
1207	kmem_cache = &kmem_cache_boot;
1208
1209	if (!IS_ENABLED(CONFIG_NUMA) || num_possible_nodes() == 1)
1210		use_alien_caches = 0;
1211
1212	for (i = 0; i < NUM_INIT_LISTS; i++)
1213		kmem_cache_node_init(&init_kmem_cache_node[i]);
1214
1215	/*
1216	 * Fragmentation resistance on low memory - only use bigger
1217	 * page orders on machines with more than 32MB of memory if
1218	 * not overridden on the command line.
1219	 */
1220	if (!slab_max_order_set && totalram_pages() > (32 << 20) >> PAGE_SHIFT)
1221		slab_max_order = SLAB_MAX_ORDER_HI;
1222
1223	/* Bootstrap is tricky, because several objects are allocated
1224	 * from caches that do not exist yet:
1225	 * 1) initialize the kmem_cache cache: it contains the struct
1226	 *    kmem_cache structures of all caches, except kmem_cache itself:
1227	 *    kmem_cache is statically allocated.
1228	 *    Initially an __init data area is used for the head array and the
1229	 *    kmem_cache_node structures, it's replaced with a kmalloc allocated
1230	 *    array at the end of the bootstrap.
1231	 * 2) Create the first kmalloc cache.
1232	 *    The struct kmem_cache for the new cache is allocated normally.
1233	 *    An __init data area is used for the head array.
1234	 * 3) Create the remaining kmalloc caches, with minimally sized
1235	 *    head arrays.
1236	 * 4) Replace the __init data head arrays for kmem_cache and the first
1237	 *    kmalloc cache with kmalloc allocated arrays.
1238	 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1239	 *    the other cache's with kmalloc allocated memory.
1240	 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1241	 */
1242
1243	/* 1) create the kmem_cache */
1244
1245	/*
1246	 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1247	 */
1248	create_boot_cache(kmem_cache, "kmem_cache",
1249		offsetof(struct kmem_cache, node) +
1250				  nr_node_ids * sizeof(struct kmem_cache_node *),
1251				  SLAB_HWCACHE_ALIGN, 0, 0);
1252	list_add(&kmem_cache->list, &slab_caches);
 
1253	slab_state = PARTIAL;
1254
1255	/*
1256	 * Initialize the caches that provide memory for the  kmem_cache_node
1257	 * structures first.  Without this, further allocations will bug.
1258	 */
1259	kmalloc_caches[KMALLOC_NORMAL][INDEX_NODE] = create_kmalloc_cache(
1260				kmalloc_info[INDEX_NODE].name[KMALLOC_NORMAL],
1261				kmalloc_info[INDEX_NODE].size,
1262				ARCH_KMALLOC_FLAGS, 0,
1263				kmalloc_info[INDEX_NODE].size);
1264	slab_state = PARTIAL_NODE;
1265	setup_kmalloc_cache_index_table();
1266
1267	slab_early_init = 0;
1268
1269	/* 5) Replace the bootstrap kmem_cache_node */
1270	{
1271		int nid;
1272
1273		for_each_online_node(nid) {
1274			init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
1275
1276			init_list(kmalloc_caches[KMALLOC_NORMAL][INDEX_NODE],
1277					  &init_kmem_cache_node[SIZE_NODE + nid], nid);
1278		}
1279	}
1280
1281	create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
1282}
1283
1284void __init kmem_cache_init_late(void)
1285{
1286	struct kmem_cache *cachep;
1287
1288	/* 6) resize the head arrays to their final sizes */
1289	mutex_lock(&slab_mutex);
1290	list_for_each_entry(cachep, &slab_caches, list)
1291		if (enable_cpucache(cachep, GFP_NOWAIT))
1292			BUG();
1293	mutex_unlock(&slab_mutex);
1294
1295	/* Done! */
1296	slab_state = FULL;
1297
1298#ifdef CONFIG_NUMA
1299	/*
1300	 * Register a memory hotplug callback that initializes and frees
1301	 * node.
1302	 */
1303	hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1304#endif
1305
1306	/*
1307	 * The reap timers are started later, with a module init call: That part
1308	 * of the kernel is not yet operational.
1309	 */
1310}
1311
1312static int __init cpucache_init(void)
1313{
1314	int ret;
1315
1316	/*
1317	 * Register the timers that return unneeded pages to the page allocator
1318	 */
1319	ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "SLAB online",
1320				slab_online_cpu, slab_offline_cpu);
1321	WARN_ON(ret < 0);
1322
1323	return 0;
1324}
1325__initcall(cpucache_init);
1326
1327static noinline void
1328slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1329{
1330#if DEBUG
1331	struct kmem_cache_node *n;
1332	unsigned long flags;
1333	int node;
1334	static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1335				      DEFAULT_RATELIMIT_BURST);
1336
1337	if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1338		return;
1339
1340	pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1341		nodeid, gfpflags, &gfpflags);
1342	pr_warn("  cache: %s, object size: %d, order: %d\n",
1343		cachep->name, cachep->size, cachep->gfporder);
1344
1345	for_each_kmem_cache_node(cachep, node, n) {
1346		unsigned long total_slabs, free_slabs, free_objs;
1347
1348		spin_lock_irqsave(&n->list_lock, flags);
1349		total_slabs = n->total_slabs;
1350		free_slabs = n->free_slabs;
1351		free_objs = n->free_objects;
1352		spin_unlock_irqrestore(&n->list_lock, flags);
1353
1354		pr_warn("  node %d: slabs: %ld/%ld, objs: %ld/%ld\n",
1355			node, total_slabs - free_slabs, total_slabs,
1356			(total_slabs * cachep->num) - free_objs,
1357			total_slabs * cachep->num);
1358	}
1359#endif
1360}
1361
1362/*
1363 * Interface to system's page allocator. No need to hold the
1364 * kmem_cache_node ->list_lock.
1365 *
1366 * If we requested dmaable memory, we will get it. Even if we
1367 * did not request dmaable memory, we might get it, but that
1368 * would be relatively rare and ignorable.
1369 */
1370static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1371								int nodeid)
1372{
1373	struct page *page;
 
1374
1375	flags |= cachep->allocflags;
1376
1377	page = __alloc_pages_node(nodeid, flags, cachep->gfporder);
1378	if (!page) {
1379		slab_out_of_memory(cachep, flags, nodeid);
1380		return NULL;
1381	}
1382
1383	account_slab_page(page, cachep->gfporder, cachep, flags);
 
 
 
 
 
 
 
 
 
 
1384	__SetPageSlab(page);
1385	/* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1386	if (sk_memalloc_socks() && page_is_pfmemalloc(page))
1387		SetPageSlabPfmemalloc(page);
1388
1389	return page;
1390}
1391
1392/*
1393 * Interface to system's page release.
1394 */
1395static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
1396{
1397	int order = cachep->gfporder;
 
 
 
 
 
 
1398
1399	BUG_ON(!PageSlab(page));
1400	__ClearPageSlabPfmemalloc(page);
1401	__ClearPageSlab(page);
1402	page_mapcount_reset(page);
1403	/* In union with page->mapping where page allocator expects NULL */
1404	page->slab_cache = NULL;
1405
1406	if (current->reclaim_state)
1407		current->reclaim_state->reclaimed_slab += 1 << order;
1408	unaccount_slab_page(page, order, cachep);
1409	__free_pages(page, order);
1410}
1411
1412static void kmem_rcu_free(struct rcu_head *head)
1413{
1414	struct kmem_cache *cachep;
1415	struct page *page;
1416
1417	page = container_of(head, struct page, rcu_head);
1418	cachep = page->slab_cache;
1419
1420	kmem_freepages(cachep, page);
1421}
1422
1423#if DEBUG
1424static bool is_debug_pagealloc_cache(struct kmem_cache *cachep)
1425{
1426	if (debug_pagealloc_enabled_static() && OFF_SLAB(cachep) &&
1427		(cachep->size % PAGE_SIZE) == 0)
1428		return true;
1429
1430	return false;
1431}
1432
1433#ifdef CONFIG_DEBUG_PAGEALLOC
1434static void slab_kernel_map(struct kmem_cache *cachep, void *objp, int map)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1435{
1436	if (!is_debug_pagealloc_cache(cachep))
1437		return;
1438
1439	__kernel_map_pages(virt_to_page(objp), cachep->size / PAGE_SIZE, map);
 
 
 
1440}
1441
1442#else
1443static inline void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1444				int map) {}
1445
1446#endif
1447
1448static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1449{
1450	int size = cachep->object_size;
1451	addr = &((char *)addr)[obj_offset(cachep)];
1452
1453	memset(addr, val, size);
1454	*(unsigned char *)(addr + size - 1) = POISON_END;
1455}
1456
1457static void dump_line(char *data, int offset, int limit)
1458{
1459	int i;
1460	unsigned char error = 0;
1461	int bad_count = 0;
1462
1463	pr_err("%03x: ", offset);
1464	for (i = 0; i < limit; i++) {
1465		if (data[offset + i] != POISON_FREE) {
1466			error = data[offset + i];
1467			bad_count++;
1468		}
1469	}
1470	print_hex_dump(KERN_CONT, "", 0, 16, 1,
1471			&data[offset], limit, 1);
1472
1473	if (bad_count == 1) {
1474		error ^= POISON_FREE;
1475		if (!(error & (error - 1))) {
1476			pr_err("Single bit error detected. Probably bad RAM.\n");
1477#ifdef CONFIG_X86
1478			pr_err("Run memtest86+ or a similar memory test tool.\n");
1479#else
1480			pr_err("Run a memory test tool.\n");
1481#endif
1482		}
1483	}
1484}
1485#endif
1486
1487#if DEBUG
1488
1489static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1490{
1491	int i, size;
1492	char *realobj;
1493
1494	if (cachep->flags & SLAB_RED_ZONE) {
1495		pr_err("Redzone: 0x%llx/0x%llx\n",
1496		       *dbg_redzone1(cachep, objp),
1497		       *dbg_redzone2(cachep, objp));
1498	}
1499
1500	if (cachep->flags & SLAB_STORE_USER)
1501		pr_err("Last user: (%pSR)\n", *dbg_userword(cachep, objp));
1502	realobj = (char *)objp + obj_offset(cachep);
1503	size = cachep->object_size;
1504	for (i = 0; i < size && lines; i += 16, lines--) {
1505		int limit;
1506		limit = 16;
1507		if (i + limit > size)
1508			limit = size - i;
1509		dump_line(realobj, i, limit);
1510	}
1511}
1512
1513static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1514{
1515	char *realobj;
1516	int size, i;
1517	int lines = 0;
1518
1519	if (is_debug_pagealloc_cache(cachep))
1520		return;
1521
1522	realobj = (char *)objp + obj_offset(cachep);
1523	size = cachep->object_size;
1524
1525	for (i = 0; i < size; i++) {
1526		char exp = POISON_FREE;
1527		if (i == size - 1)
1528			exp = POISON_END;
1529		if (realobj[i] != exp) {
1530			int limit;
1531			/* Mismatch ! */
1532			/* Print header */
1533			if (lines == 0) {
1534				pr_err("Slab corruption (%s): %s start=%px, len=%d\n",
1535				       print_tainted(), cachep->name,
1536				       realobj, size);
1537				print_objinfo(cachep, objp, 0);
1538			}
1539			/* Hexdump the affected line */
1540			i = (i / 16) * 16;
1541			limit = 16;
1542			if (i + limit > size)
1543				limit = size - i;
1544			dump_line(realobj, i, limit);
1545			i += 16;
1546			lines++;
1547			/* Limit to 5 lines */
1548			if (lines > 5)
1549				break;
1550		}
1551	}
1552	if (lines != 0) {
1553		/* Print some data about the neighboring objects, if they
1554		 * exist:
1555		 */
1556		struct page *page = virt_to_head_page(objp);
1557		unsigned int objnr;
1558
1559		objnr = obj_to_index(cachep, page, objp);
1560		if (objnr) {
1561			objp = index_to_obj(cachep, page, objnr - 1);
1562			realobj = (char *)objp + obj_offset(cachep);
1563			pr_err("Prev obj: start=%px, len=%d\n", realobj, size);
1564			print_objinfo(cachep, objp, 2);
1565		}
1566		if (objnr + 1 < cachep->num) {
1567			objp = index_to_obj(cachep, page, objnr + 1);
1568			realobj = (char *)objp + obj_offset(cachep);
1569			pr_err("Next obj: start=%px, len=%d\n", realobj, size);
1570			print_objinfo(cachep, objp, 2);
1571		}
1572	}
1573}
1574#endif
1575
1576#if DEBUG
1577static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1578						struct page *page)
1579{
1580	int i;
1581
1582	if (OBJFREELIST_SLAB(cachep) && cachep->flags & SLAB_POISON) {
1583		poison_obj(cachep, page->freelist - obj_offset(cachep),
1584			POISON_FREE);
1585	}
1586
1587	for (i = 0; i < cachep->num; i++) {
1588		void *objp = index_to_obj(cachep, page, i);
1589
1590		if (cachep->flags & SLAB_POISON) {
1591			check_poison_obj(cachep, objp);
1592			slab_kernel_map(cachep, objp, 1);
1593		}
1594		if (cachep->flags & SLAB_RED_ZONE) {
1595			if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1596				slab_error(cachep, "start of a freed object was overwritten");
1597			if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1598				slab_error(cachep, "end of a freed object was overwritten");
1599		}
1600	}
1601}
1602#else
1603static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1604						struct page *page)
1605{
1606}
1607#endif
1608
1609/**
1610 * slab_destroy - destroy and release all objects in a slab
1611 * @cachep: cache pointer being destroyed
1612 * @page: page pointer being destroyed
1613 *
1614 * Destroy all the objs in a slab page, and release the mem back to the system.
1615 * Before calling the slab page must have been unlinked from the cache. The
1616 * kmem_cache_node ->list_lock is not held/needed.
1617 */
1618static void slab_destroy(struct kmem_cache *cachep, struct page *page)
1619{
1620	void *freelist;
1621
1622	freelist = page->freelist;
1623	slab_destroy_debugcheck(cachep, page);
1624	if (unlikely(cachep->flags & SLAB_TYPESAFE_BY_RCU))
1625		call_rcu(&page->rcu_head, kmem_rcu_free);
1626	else
1627		kmem_freepages(cachep, page);
1628
1629	/*
1630	 * From now on, we don't use freelist
1631	 * although actual page can be freed in rcu context
1632	 */
1633	if (OFF_SLAB(cachep))
1634		kmem_cache_free(cachep->freelist_cache, freelist);
1635}
1636
1637/*
1638 * Update the size of the caches before calling slabs_destroy as it may
1639 * recursively call kfree.
1640 */
1641static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
1642{
1643	struct page *page, *n;
1644
1645	list_for_each_entry_safe(page, n, list, slab_list) {
1646		list_del(&page->slab_list);
1647		slab_destroy(cachep, page);
1648	}
1649}
1650
1651/**
1652 * calculate_slab_order - calculate size (page order) of slabs
1653 * @cachep: pointer to the cache that is being created
1654 * @size: size of objects to be created in this cache.
1655 * @flags: slab allocation flags
1656 *
1657 * Also calculates the number of objects per slab.
1658 *
1659 * This could be made much more intelligent.  For now, try to avoid using
1660 * high order pages for slabs.  When the gfp() functions are more friendly
1661 * towards high-order requests, this should be changed.
1662 *
1663 * Return: number of left-over bytes in a slab
1664 */
1665static size_t calculate_slab_order(struct kmem_cache *cachep,
1666				size_t size, slab_flags_t flags)
1667{
1668	size_t left_over = 0;
1669	int gfporder;
1670
1671	for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
1672		unsigned int num;
1673		size_t remainder;
1674
1675		num = cache_estimate(gfporder, size, flags, &remainder);
1676		if (!num)
1677			continue;
1678
1679		/* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1680		if (num > SLAB_OBJ_MAX_NUM)
1681			break;
1682
1683		if (flags & CFLGS_OFF_SLAB) {
1684			struct kmem_cache *freelist_cache;
1685			size_t freelist_size;
1686
1687			freelist_size = num * sizeof(freelist_idx_t);
1688			freelist_cache = kmalloc_slab(freelist_size, 0u);
1689			if (!freelist_cache)
1690				continue;
1691
1692			/*
1693			 * Needed to avoid possible looping condition
1694			 * in cache_grow_begin()
1695			 */
1696			if (OFF_SLAB(freelist_cache))
1697				continue;
1698
1699			/* check if off slab has enough benefit */
1700			if (freelist_cache->size > cachep->size / 2)
1701				continue;
1702		}
1703
1704		/* Found something acceptable - save it away */
1705		cachep->num = num;
1706		cachep->gfporder = gfporder;
1707		left_over = remainder;
1708
1709		/*
1710		 * A VFS-reclaimable slab tends to have most allocations
1711		 * as GFP_NOFS and we really don't want to have to be allocating
1712		 * higher-order pages when we are unable to shrink dcache.
1713		 */
1714		if (flags & SLAB_RECLAIM_ACCOUNT)
1715			break;
1716
1717		/*
1718		 * Large number of objects is good, but very large slabs are
1719		 * currently bad for the gfp()s.
1720		 */
1721		if (gfporder >= slab_max_order)
1722			break;
1723
1724		/*
1725		 * Acceptable internal fragmentation?
1726		 */
1727		if (left_over * 8 <= (PAGE_SIZE << gfporder))
1728			break;
1729	}
1730	return left_over;
1731}
1732
1733static struct array_cache __percpu *alloc_kmem_cache_cpus(
1734		struct kmem_cache *cachep, int entries, int batchcount)
1735{
1736	int cpu;
1737	size_t size;
1738	struct array_cache __percpu *cpu_cache;
1739
1740	size = sizeof(void *) * entries + sizeof(struct array_cache);
1741	cpu_cache = __alloc_percpu(size, sizeof(void *));
1742
1743	if (!cpu_cache)
1744		return NULL;
1745
1746	for_each_possible_cpu(cpu) {
1747		init_arraycache(per_cpu_ptr(cpu_cache, cpu),
1748				entries, batchcount);
1749	}
1750
1751	return cpu_cache;
1752}
1753
1754static int __ref setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
1755{
1756	if (slab_state >= FULL)
1757		return enable_cpucache(cachep, gfp);
1758
1759	cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1);
1760	if (!cachep->cpu_cache)
1761		return 1;
1762
1763	if (slab_state == DOWN) {
1764		/* Creation of first cache (kmem_cache). */
1765		set_up_node(kmem_cache, CACHE_CACHE);
1766	} else if (slab_state == PARTIAL) {
1767		/* For kmem_cache_node */
1768		set_up_node(cachep, SIZE_NODE);
1769	} else {
1770		int node;
1771
1772		for_each_online_node(node) {
1773			cachep->node[node] = kmalloc_node(
1774				sizeof(struct kmem_cache_node), gfp, node);
1775			BUG_ON(!cachep->node[node]);
1776			kmem_cache_node_init(cachep->node[node]);
1777		}
1778	}
1779
1780	cachep->node[numa_mem_id()]->next_reap =
1781			jiffies + REAPTIMEOUT_NODE +
1782			((unsigned long)cachep) % REAPTIMEOUT_NODE;
1783
1784	cpu_cache_get(cachep)->avail = 0;
1785	cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1786	cpu_cache_get(cachep)->batchcount = 1;
1787	cpu_cache_get(cachep)->touched = 0;
1788	cachep->batchcount = 1;
1789	cachep->limit = BOOT_CPUCACHE_ENTRIES;
1790	return 0;
1791}
1792
1793slab_flags_t kmem_cache_flags(unsigned int object_size,
1794	slab_flags_t flags, const char *name)
 
1795{
1796	return flags;
1797}
1798
1799struct kmem_cache *
1800__kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
1801		   slab_flags_t flags, void (*ctor)(void *))
1802{
1803	struct kmem_cache *cachep;
1804
1805	cachep = find_mergeable(size, align, flags, name, ctor);
1806	if (cachep) {
1807		cachep->refcount++;
1808
1809		/*
1810		 * Adjust the object sizes so that we clear
1811		 * the complete object on kzalloc.
1812		 */
1813		cachep->object_size = max_t(int, cachep->object_size, size);
1814	}
1815	return cachep;
1816}
1817
1818static bool set_objfreelist_slab_cache(struct kmem_cache *cachep,
1819			size_t size, slab_flags_t flags)
1820{
1821	size_t left;
1822
1823	cachep->num = 0;
1824
1825	/*
1826	 * If slab auto-initialization on free is enabled, store the freelist
1827	 * off-slab, so that its contents don't end up in one of the allocated
1828	 * objects.
1829	 */
1830	if (unlikely(slab_want_init_on_free(cachep)))
1831		return false;
1832
1833	if (cachep->ctor || flags & SLAB_TYPESAFE_BY_RCU)
1834		return false;
1835
1836	left = calculate_slab_order(cachep, size,
1837			flags | CFLGS_OBJFREELIST_SLAB);
1838	if (!cachep->num)
1839		return false;
1840
1841	if (cachep->num * sizeof(freelist_idx_t) > cachep->object_size)
1842		return false;
1843
1844	cachep->colour = left / cachep->colour_off;
1845
1846	return true;
1847}
1848
1849static bool set_off_slab_cache(struct kmem_cache *cachep,
1850			size_t size, slab_flags_t flags)
1851{
1852	size_t left;
1853
1854	cachep->num = 0;
1855
1856	/*
1857	 * Always use on-slab management when SLAB_NOLEAKTRACE
1858	 * to avoid recursive calls into kmemleak.
1859	 */
1860	if (flags & SLAB_NOLEAKTRACE)
1861		return false;
1862
1863	/*
1864	 * Size is large, assume best to place the slab management obj
1865	 * off-slab (should allow better packing of objs).
1866	 */
1867	left = calculate_slab_order(cachep, size, flags | CFLGS_OFF_SLAB);
1868	if (!cachep->num)
1869		return false;
1870
1871	/*
1872	 * If the slab has been placed off-slab, and we have enough space then
1873	 * move it on-slab. This is at the expense of any extra colouring.
1874	 */
1875	if (left >= cachep->num * sizeof(freelist_idx_t))
1876		return false;
1877
1878	cachep->colour = left / cachep->colour_off;
1879
1880	return true;
1881}
1882
1883static bool set_on_slab_cache(struct kmem_cache *cachep,
1884			size_t size, slab_flags_t flags)
1885{
1886	size_t left;
1887
1888	cachep->num = 0;
1889
1890	left = calculate_slab_order(cachep, size, flags);
1891	if (!cachep->num)
1892		return false;
1893
1894	cachep->colour = left / cachep->colour_off;
1895
1896	return true;
1897}
1898
1899/**
1900 * __kmem_cache_create - Create a cache.
1901 * @cachep: cache management descriptor
1902 * @flags: SLAB flags
1903 *
1904 * Returns a ptr to the cache on success, NULL on failure.
1905 * Cannot be called within a int, but can be interrupted.
1906 * The @ctor is run when new pages are allocated by the cache.
1907 *
1908 * The flags are
1909 *
1910 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1911 * to catch references to uninitialised memory.
1912 *
1913 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1914 * for buffer overruns.
1915 *
1916 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1917 * cacheline.  This can be beneficial if you're counting cycles as closely
1918 * as davem.
1919 *
1920 * Return: a pointer to the created cache or %NULL in case of error
1921 */
1922int __kmem_cache_create(struct kmem_cache *cachep, slab_flags_t flags)
1923{
1924	size_t ralign = BYTES_PER_WORD;
1925	gfp_t gfp;
1926	int err;
1927	unsigned int size = cachep->size;
1928
1929#if DEBUG
1930#if FORCED_DEBUG
1931	/*
1932	 * Enable redzoning and last user accounting, except for caches with
1933	 * large objects, if the increased size would increase the object size
1934	 * above the next power of two: caches with object sizes just above a
1935	 * power of two have a significant amount of internal fragmentation.
1936	 */
1937	if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
1938						2 * sizeof(unsigned long long)))
1939		flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
1940	if (!(flags & SLAB_TYPESAFE_BY_RCU))
1941		flags |= SLAB_POISON;
1942#endif
1943#endif
1944
1945	/*
1946	 * Check that size is in terms of words.  This is needed to avoid
1947	 * unaligned accesses for some archs when redzoning is used, and makes
1948	 * sure any on-slab bufctl's are also correctly aligned.
1949	 */
1950	size = ALIGN(size, BYTES_PER_WORD);
1951
1952	if (flags & SLAB_RED_ZONE) {
1953		ralign = REDZONE_ALIGN;
1954		/* If redzoning, ensure that the second redzone is suitably
1955		 * aligned, by adjusting the object size accordingly. */
1956		size = ALIGN(size, REDZONE_ALIGN);
1957	}
1958
1959	/* 3) caller mandated alignment */
1960	if (ralign < cachep->align) {
1961		ralign = cachep->align;
1962	}
1963	/* disable debug if necessary */
1964	if (ralign > __alignof__(unsigned long long))
1965		flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
1966	/*
1967	 * 4) Store it.
1968	 */
1969	cachep->align = ralign;
1970	cachep->colour_off = cache_line_size();
1971	/* Offset must be a multiple of the alignment. */
1972	if (cachep->colour_off < cachep->align)
1973		cachep->colour_off = cachep->align;
1974
1975	if (slab_is_available())
1976		gfp = GFP_KERNEL;
1977	else
1978		gfp = GFP_NOWAIT;
1979
1980#if DEBUG
1981
1982	/*
1983	 * Both debugging options require word-alignment which is calculated
1984	 * into align above.
1985	 */
1986	if (flags & SLAB_RED_ZONE) {
1987		/* add space for red zone words */
1988		cachep->obj_offset += sizeof(unsigned long long);
1989		size += 2 * sizeof(unsigned long long);
1990	}
1991	if (flags & SLAB_STORE_USER) {
1992		/* user store requires one word storage behind the end of
1993		 * the real object. But if the second red zone needs to be
1994		 * aligned to 64 bits, we must allow that much space.
1995		 */
1996		if (flags & SLAB_RED_ZONE)
1997			size += REDZONE_ALIGN;
1998		else
1999			size += BYTES_PER_WORD;
2000	}
2001#endif
2002
2003	kasan_cache_create(cachep, &size, &flags);
2004
2005	size = ALIGN(size, cachep->align);
2006	/*
2007	 * We should restrict the number of objects in a slab to implement
2008	 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2009	 */
2010	if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2011		size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
2012
2013#if DEBUG
2014	/*
2015	 * To activate debug pagealloc, off-slab management is necessary
2016	 * requirement. In early phase of initialization, small sized slab
2017	 * doesn't get initialized so it would not be possible. So, we need
2018	 * to check size >= 256. It guarantees that all necessary small
2019	 * sized slab is initialized in current slab initialization sequence.
2020	 */
2021	if (debug_pagealloc_enabled_static() && (flags & SLAB_POISON) &&
2022		size >= 256 && cachep->object_size > cache_line_size()) {
2023		if (size < PAGE_SIZE || size % PAGE_SIZE == 0) {
2024			size_t tmp_size = ALIGN(size, PAGE_SIZE);
2025
2026			if (set_off_slab_cache(cachep, tmp_size, flags)) {
2027				flags |= CFLGS_OFF_SLAB;
2028				cachep->obj_offset += tmp_size - size;
2029				size = tmp_size;
2030				goto done;
2031			}
2032		}
2033	}
2034#endif
2035
2036	if (set_objfreelist_slab_cache(cachep, size, flags)) {
2037		flags |= CFLGS_OBJFREELIST_SLAB;
2038		goto done;
2039	}
2040
2041	if (set_off_slab_cache(cachep, size, flags)) {
2042		flags |= CFLGS_OFF_SLAB;
2043		goto done;
2044	}
2045
2046	if (set_on_slab_cache(cachep, size, flags))
2047		goto done;
2048
2049	return -E2BIG;
2050
2051done:
2052	cachep->freelist_size = cachep->num * sizeof(freelist_idx_t);
2053	cachep->flags = flags;
2054	cachep->allocflags = __GFP_COMP;
2055	if (flags & SLAB_CACHE_DMA)
2056		cachep->allocflags |= GFP_DMA;
2057	if (flags & SLAB_CACHE_DMA32)
2058		cachep->allocflags |= GFP_DMA32;
2059	if (flags & SLAB_RECLAIM_ACCOUNT)
2060		cachep->allocflags |= __GFP_RECLAIMABLE;
2061	cachep->size = size;
2062	cachep->reciprocal_buffer_size = reciprocal_value(size);
2063
2064#if DEBUG
2065	/*
2066	 * If we're going to use the generic kernel_map_pages()
2067	 * poisoning, then it's going to smash the contents of
2068	 * the redzone and userword anyhow, so switch them off.
2069	 */
2070	if (IS_ENABLED(CONFIG_PAGE_POISONING) &&
2071		(cachep->flags & SLAB_POISON) &&
2072		is_debug_pagealloc_cache(cachep))
2073		cachep->flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2074#endif
2075
2076	if (OFF_SLAB(cachep)) {
2077		cachep->freelist_cache =
2078			kmalloc_slab(cachep->freelist_size, 0u);
2079	}
2080
2081	err = setup_cpu_cache(cachep, gfp);
2082	if (err) {
2083		__kmem_cache_release(cachep);
2084		return err;
2085	}
2086
2087	return 0;
2088}
2089
2090#if DEBUG
2091static void check_irq_off(void)
2092{
2093	BUG_ON(!irqs_disabled());
2094}
2095
2096static void check_irq_on(void)
2097{
2098	BUG_ON(irqs_disabled());
2099}
2100
2101static void check_mutex_acquired(void)
2102{
2103	BUG_ON(!mutex_is_locked(&slab_mutex));
2104}
2105
2106static void check_spinlock_acquired(struct kmem_cache *cachep)
2107{
2108#ifdef CONFIG_SMP
2109	check_irq_off();
2110	assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
2111#endif
2112}
2113
2114static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2115{
2116#ifdef CONFIG_SMP
2117	check_irq_off();
2118	assert_spin_locked(&get_node(cachep, node)->list_lock);
2119#endif
2120}
2121
2122#else
2123#define check_irq_off()	do { } while(0)
2124#define check_irq_on()	do { } while(0)
2125#define check_mutex_acquired()	do { } while(0)
2126#define check_spinlock_acquired(x) do { } while(0)
2127#define check_spinlock_acquired_node(x, y) do { } while(0)
2128#endif
2129
2130static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac,
2131				int node, bool free_all, struct list_head *list)
2132{
2133	int tofree;
2134
2135	if (!ac || !ac->avail)
2136		return;
2137
2138	tofree = free_all ? ac->avail : (ac->limit + 4) / 5;
2139	if (tofree > ac->avail)
2140		tofree = (ac->avail + 1) / 2;
2141
2142	free_block(cachep, ac->entry, tofree, node, list);
2143	ac->avail -= tofree;
2144	memmove(ac->entry, &(ac->entry[tofree]), sizeof(void *) * ac->avail);
2145}
2146
2147static void do_drain(void *arg)
2148{
2149	struct kmem_cache *cachep = arg;
2150	struct array_cache *ac;
2151	int node = numa_mem_id();
2152	struct kmem_cache_node *n;
2153	LIST_HEAD(list);
2154
2155	check_irq_off();
2156	ac = cpu_cache_get(cachep);
2157	n = get_node(cachep, node);
2158	spin_lock(&n->list_lock);
2159	free_block(cachep, ac->entry, ac->avail, node, &list);
2160	spin_unlock(&n->list_lock);
 
2161	ac->avail = 0;
2162	slabs_destroy(cachep, &list);
2163}
2164
2165static void drain_cpu_caches(struct kmem_cache *cachep)
2166{
2167	struct kmem_cache_node *n;
2168	int node;
2169	LIST_HEAD(list);
2170
2171	on_each_cpu(do_drain, cachep, 1);
2172	check_irq_on();
2173	for_each_kmem_cache_node(cachep, node, n)
2174		if (n->alien)
2175			drain_alien_cache(cachep, n->alien);
2176
2177	for_each_kmem_cache_node(cachep, node, n) {
2178		spin_lock_irq(&n->list_lock);
2179		drain_array_locked(cachep, n->shared, node, true, &list);
2180		spin_unlock_irq(&n->list_lock);
2181
2182		slabs_destroy(cachep, &list);
2183	}
2184}
2185
2186/*
2187 * Remove slabs from the list of free slabs.
2188 * Specify the number of slabs to drain in tofree.
2189 *
2190 * Returns the actual number of slabs released.
2191 */
2192static int drain_freelist(struct kmem_cache *cache,
2193			struct kmem_cache_node *n, int tofree)
2194{
2195	struct list_head *p;
2196	int nr_freed;
2197	struct page *page;
2198
2199	nr_freed = 0;
2200	while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
2201
2202		spin_lock_irq(&n->list_lock);
2203		p = n->slabs_free.prev;
2204		if (p == &n->slabs_free) {
2205			spin_unlock_irq(&n->list_lock);
2206			goto out;
2207		}
2208
2209		page = list_entry(p, struct page, slab_list);
2210		list_del(&page->slab_list);
2211		n->free_slabs--;
2212		n->total_slabs--;
2213		/*
2214		 * Safe to drop the lock. The slab is no longer linked
2215		 * to the cache.
2216		 */
2217		n->free_objects -= cache->num;
2218		spin_unlock_irq(&n->list_lock);
2219		slab_destroy(cache, page);
2220		nr_freed++;
2221	}
2222out:
2223	return nr_freed;
2224}
2225
2226bool __kmem_cache_empty(struct kmem_cache *s)
2227{
2228	int node;
2229	struct kmem_cache_node *n;
2230
2231	for_each_kmem_cache_node(s, node, n)
2232		if (!list_empty(&n->slabs_full) ||
2233		    !list_empty(&n->slabs_partial))
2234			return false;
2235	return true;
2236}
2237
2238int __kmem_cache_shrink(struct kmem_cache *cachep)
2239{
2240	int ret = 0;
2241	int node;
2242	struct kmem_cache_node *n;
2243
2244	drain_cpu_caches(cachep);
2245
2246	check_irq_on();
2247	for_each_kmem_cache_node(cachep, node, n) {
2248		drain_freelist(cachep, n, INT_MAX);
2249
2250		ret += !list_empty(&n->slabs_full) ||
2251			!list_empty(&n->slabs_partial);
2252	}
2253	return (ret ? 1 : 0);
2254}
2255
 
 
 
 
 
 
 
2256int __kmem_cache_shutdown(struct kmem_cache *cachep)
2257{
2258	return __kmem_cache_shrink(cachep);
2259}
2260
2261void __kmem_cache_release(struct kmem_cache *cachep)
2262{
2263	int i;
2264	struct kmem_cache_node *n;
2265
2266	cache_random_seq_destroy(cachep);
2267
2268	free_percpu(cachep->cpu_cache);
2269
2270	/* NUMA: free the node structures */
2271	for_each_kmem_cache_node(cachep, i, n) {
2272		kfree(n->shared);
2273		free_alien_cache(n->alien);
2274		kfree(n);
2275		cachep->node[i] = NULL;
2276	}
2277}
2278
2279/*
2280 * Get the memory for a slab management obj.
2281 *
2282 * For a slab cache when the slab descriptor is off-slab, the
2283 * slab descriptor can't come from the same cache which is being created,
2284 * Because if it is the case, that means we defer the creation of
2285 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2286 * And we eventually call down to __kmem_cache_create(), which
2287 * in turn looks up in the kmalloc_{dma,}_caches for the desired-size one.
2288 * This is a "chicken-and-egg" problem.
2289 *
2290 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2291 * which are all initialized during kmem_cache_init().
2292 */
2293static void *alloc_slabmgmt(struct kmem_cache *cachep,
2294				   struct page *page, int colour_off,
2295				   gfp_t local_flags, int nodeid)
2296{
2297	void *freelist;
2298	void *addr = page_address(page);
2299
2300	page->s_mem = addr + colour_off;
2301	page->active = 0;
2302
2303	if (OBJFREELIST_SLAB(cachep))
2304		freelist = NULL;
2305	else if (OFF_SLAB(cachep)) {
2306		/* Slab management obj is off-slab. */
2307		freelist = kmem_cache_alloc_node(cachep->freelist_cache,
2308					      local_flags, nodeid);
 
 
2309	} else {
2310		/* We will use last bytes at the slab for freelist */
2311		freelist = addr + (PAGE_SIZE << cachep->gfporder) -
2312				cachep->freelist_size;
2313	}
2314
2315	return freelist;
2316}
2317
2318static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
2319{
2320	return ((freelist_idx_t *)page->freelist)[idx];
2321}
2322
2323static inline void set_free_obj(struct page *page,
2324					unsigned int idx, freelist_idx_t val)
2325{
2326	((freelist_idx_t *)(page->freelist))[idx] = val;
2327}
2328
2329static void cache_init_objs_debug(struct kmem_cache *cachep, struct page *page)
2330{
2331#if DEBUG
2332	int i;
2333
2334	for (i = 0; i < cachep->num; i++) {
2335		void *objp = index_to_obj(cachep, page, i);
2336
2337		if (cachep->flags & SLAB_STORE_USER)
2338			*dbg_userword(cachep, objp) = NULL;
2339
2340		if (cachep->flags & SLAB_RED_ZONE) {
2341			*dbg_redzone1(cachep, objp) = RED_INACTIVE;
2342			*dbg_redzone2(cachep, objp) = RED_INACTIVE;
2343		}
2344		/*
2345		 * Constructors are not allowed to allocate memory from the same
2346		 * cache which they are a constructor for.  Otherwise, deadlock.
2347		 * They must also be threaded.
2348		 */
2349		if (cachep->ctor && !(cachep->flags & SLAB_POISON)) {
2350			kasan_unpoison_object_data(cachep,
2351						   objp + obj_offset(cachep));
2352			cachep->ctor(objp + obj_offset(cachep));
2353			kasan_poison_object_data(
2354				cachep, objp + obj_offset(cachep));
2355		}
2356
2357		if (cachep->flags & SLAB_RED_ZONE) {
2358			if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2359				slab_error(cachep, "constructor overwrote the end of an object");
2360			if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2361				slab_error(cachep, "constructor overwrote the start of an object");
2362		}
2363		/* need to poison the objs? */
2364		if (cachep->flags & SLAB_POISON) {
2365			poison_obj(cachep, objp, POISON_FREE);
2366			slab_kernel_map(cachep, objp, 0);
2367		}
2368	}
2369#endif
2370}
2371
2372#ifdef CONFIG_SLAB_FREELIST_RANDOM
2373/* Hold information during a freelist initialization */
2374union freelist_init_state {
2375	struct {
2376		unsigned int pos;
2377		unsigned int *list;
2378		unsigned int count;
2379	};
2380	struct rnd_state rnd_state;
2381};
2382
2383/*
2384 * Initialize the state based on the randomization method available.
2385 * return true if the pre-computed list is available, false otherwise.
2386 */
2387static bool freelist_state_initialize(union freelist_init_state *state,
2388				struct kmem_cache *cachep,
2389				unsigned int count)
2390{
2391	bool ret;
2392	unsigned int rand;
2393
2394	/* Use best entropy available to define a random shift */
2395	rand = get_random_int();
2396
2397	/* Use a random state if the pre-computed list is not available */
2398	if (!cachep->random_seq) {
2399		prandom_seed_state(&state->rnd_state, rand);
2400		ret = false;
2401	} else {
2402		state->list = cachep->random_seq;
2403		state->count = count;
2404		state->pos = rand % count;
2405		ret = true;
2406	}
2407	return ret;
2408}
2409
2410/* Get the next entry on the list and randomize it using a random shift */
2411static freelist_idx_t next_random_slot(union freelist_init_state *state)
2412{
2413	if (state->pos >= state->count)
2414		state->pos = 0;
2415	return state->list[state->pos++];
2416}
2417
2418/* Swap two freelist entries */
2419static void swap_free_obj(struct page *page, unsigned int a, unsigned int b)
2420{
2421	swap(((freelist_idx_t *)page->freelist)[a],
2422		((freelist_idx_t *)page->freelist)[b]);
2423}
2424
2425/*
2426 * Shuffle the freelist initialization state based on pre-computed lists.
2427 * return true if the list was successfully shuffled, false otherwise.
2428 */
2429static bool shuffle_freelist(struct kmem_cache *cachep, struct page *page)
2430{
2431	unsigned int objfreelist = 0, i, rand, count = cachep->num;
2432	union freelist_init_state state;
2433	bool precomputed;
2434
2435	if (count < 2)
2436		return false;
2437
2438	precomputed = freelist_state_initialize(&state, cachep, count);
2439
2440	/* Take a random entry as the objfreelist */
2441	if (OBJFREELIST_SLAB(cachep)) {
2442		if (!precomputed)
2443			objfreelist = count - 1;
2444		else
2445			objfreelist = next_random_slot(&state);
2446		page->freelist = index_to_obj(cachep, page, objfreelist) +
2447						obj_offset(cachep);
2448		count--;
2449	}
2450
2451	/*
2452	 * On early boot, generate the list dynamically.
2453	 * Later use a pre-computed list for speed.
2454	 */
2455	if (!precomputed) {
2456		for (i = 0; i < count; i++)
2457			set_free_obj(page, i, i);
2458
2459		/* Fisher-Yates shuffle */
2460		for (i = count - 1; i > 0; i--) {
2461			rand = prandom_u32_state(&state.rnd_state);
2462			rand %= (i + 1);
2463			swap_free_obj(page, i, rand);
2464		}
2465	} else {
2466		for (i = 0; i < count; i++)
2467			set_free_obj(page, i, next_random_slot(&state));
2468	}
2469
2470	if (OBJFREELIST_SLAB(cachep))
2471		set_free_obj(page, cachep->num - 1, objfreelist);
2472
2473	return true;
2474}
2475#else
2476static inline bool shuffle_freelist(struct kmem_cache *cachep,
2477				struct page *page)
2478{
2479	return false;
2480}
2481#endif /* CONFIG_SLAB_FREELIST_RANDOM */
2482
2483static void cache_init_objs(struct kmem_cache *cachep,
2484			    struct page *page)
2485{
2486	int i;
2487	void *objp;
2488	bool shuffled;
2489
2490	cache_init_objs_debug(cachep, page);
2491
2492	/* Try to randomize the freelist if enabled */
2493	shuffled = shuffle_freelist(cachep, page);
2494
2495	if (!shuffled && OBJFREELIST_SLAB(cachep)) {
2496		page->freelist = index_to_obj(cachep, page, cachep->num - 1) +
2497						obj_offset(cachep);
2498	}
2499
2500	for (i = 0; i < cachep->num; i++) {
2501		objp = index_to_obj(cachep, page, i);
2502		objp = kasan_init_slab_obj(cachep, objp);
2503
2504		/* constructor could break poison info */
2505		if (DEBUG == 0 && cachep->ctor) {
2506			kasan_unpoison_object_data(cachep, objp);
2507			cachep->ctor(objp);
2508			kasan_poison_object_data(cachep, objp);
2509		}
2510
2511		if (!shuffled)
2512			set_free_obj(page, i, i);
2513	}
2514}
2515
2516static void *slab_get_obj(struct kmem_cache *cachep, struct page *page)
2517{
2518	void *objp;
2519
2520	objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
2521	page->active++;
2522
 
 
 
 
 
2523	return objp;
2524}
2525
2526static void slab_put_obj(struct kmem_cache *cachep,
2527			struct page *page, void *objp)
2528{
2529	unsigned int objnr = obj_to_index(cachep, page, objp);
2530#if DEBUG
2531	unsigned int i;
2532
2533	/* Verify double free bug */
2534	for (i = page->active; i < cachep->num; i++) {
2535		if (get_free_obj(page, i) == objnr) {
2536			pr_err("slab: double free detected in cache '%s', objp %px\n",
2537			       cachep->name, objp);
2538			BUG();
2539		}
2540	}
2541#endif
2542	page->active--;
2543	if (!page->freelist)
2544		page->freelist = objp + obj_offset(cachep);
2545
2546	set_free_obj(page, page->active, objnr);
2547}
2548
2549/*
2550 * Map pages beginning at addr to the given cache and slab. This is required
2551 * for the slab allocator to be able to lookup the cache and slab of a
2552 * virtual address for kfree, ksize, and slab debugging.
2553 */
2554static void slab_map_pages(struct kmem_cache *cache, struct page *page,
2555			   void *freelist)
2556{
2557	page->slab_cache = cache;
2558	page->freelist = freelist;
2559}
2560
2561/*
2562 * Grow (by 1) the number of slabs within a cache.  This is called by
2563 * kmem_cache_alloc() when there are no active objs left in a cache.
2564 */
2565static struct page *cache_grow_begin(struct kmem_cache *cachep,
2566				gfp_t flags, int nodeid)
2567{
2568	void *freelist;
2569	size_t offset;
2570	gfp_t local_flags;
2571	int page_node;
2572	struct kmem_cache_node *n;
2573	struct page *page;
2574
2575	/*
2576	 * Be lazy and only check for valid flags here,  keeping it out of the
2577	 * critical path in kmem_cache_alloc().
2578	 */
2579	if (unlikely(flags & GFP_SLAB_BUG_MASK))
2580		flags = kmalloc_fix_flags(flags);
2581
2582	WARN_ON_ONCE(cachep->ctor && (flags & __GFP_ZERO));
 
 
 
2583	local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2584
2585	check_irq_off();
2586	if (gfpflags_allow_blocking(local_flags))
2587		local_irq_enable();
2588
2589	/*
2590	 * Get mem for the objs.  Attempt to allocate a physical page from
2591	 * 'nodeid'.
2592	 */
2593	page = kmem_getpages(cachep, local_flags, nodeid);
2594	if (!page)
2595		goto failed;
2596
2597	page_node = page_to_nid(page);
2598	n = get_node(cachep, page_node);
2599
2600	/* Get colour for the slab, and cal the next value. */
2601	n->colour_next++;
2602	if (n->colour_next >= cachep->colour)
2603		n->colour_next = 0;
2604
2605	offset = n->colour_next;
2606	if (offset >= cachep->colour)
2607		offset = 0;
2608
2609	offset *= cachep->colour_off;
2610
2611	/*
2612	 * Call kasan_poison_slab() before calling alloc_slabmgmt(), so
2613	 * page_address() in the latter returns a non-tagged pointer,
2614	 * as it should be for slab pages.
2615	 */
2616	kasan_poison_slab(page);
2617
2618	/* Get slab management. */
2619	freelist = alloc_slabmgmt(cachep, page, offset,
2620			local_flags & ~GFP_CONSTRAINT_MASK, page_node);
2621	if (OFF_SLAB(cachep) && !freelist)
2622		goto opps1;
2623
2624	slab_map_pages(cachep, page, freelist);
2625
 
2626	cache_init_objs(cachep, page);
2627
2628	if (gfpflags_allow_blocking(local_flags))
2629		local_irq_disable();
2630
2631	return page;
2632
2633opps1:
2634	kmem_freepages(cachep, page);
2635failed:
2636	if (gfpflags_allow_blocking(local_flags))
2637		local_irq_disable();
2638	return NULL;
2639}
2640
2641static void cache_grow_end(struct kmem_cache *cachep, struct page *page)
2642{
2643	struct kmem_cache_node *n;
2644	void *list = NULL;
2645
2646	check_irq_off();
2647
2648	if (!page)
2649		return;
2650
2651	INIT_LIST_HEAD(&page->slab_list);
2652	n = get_node(cachep, page_to_nid(page));
2653
2654	spin_lock(&n->list_lock);
2655	n->total_slabs++;
2656	if (!page->active) {
2657		list_add_tail(&page->slab_list, &n->slabs_free);
2658		n->free_slabs++;
2659	} else
2660		fixup_slab_list(cachep, n, page, &list);
2661
2662	STATS_INC_GROWN(cachep);
2663	n->free_objects += cachep->num - page->active;
2664	spin_unlock(&n->list_lock);
2665
2666	fixup_objfreelist_debug(cachep, &list);
2667}
2668
2669#if DEBUG
2670
2671/*
2672 * Perform extra freeing checks:
2673 * - detect bad pointers.
2674 * - POISON/RED_ZONE checking
2675 */
2676static void kfree_debugcheck(const void *objp)
2677{
2678	if (!virt_addr_valid(objp)) {
2679		pr_err("kfree_debugcheck: out of range ptr %lxh\n",
2680		       (unsigned long)objp);
2681		BUG();
2682	}
2683}
2684
2685static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2686{
2687	unsigned long long redzone1, redzone2;
2688
2689	redzone1 = *dbg_redzone1(cache, obj);
2690	redzone2 = *dbg_redzone2(cache, obj);
2691
2692	/*
2693	 * Redzone is ok.
2694	 */
2695	if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2696		return;
2697
2698	if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2699		slab_error(cache, "double free detected");
2700	else
2701		slab_error(cache, "memory outside object was overwritten");
2702
2703	pr_err("%px: redzone 1:0x%llx, redzone 2:0x%llx\n",
2704	       obj, redzone1, redzone2);
2705}
2706
2707static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2708				   unsigned long caller)
2709{
2710	unsigned int objnr;
2711	struct page *page;
2712
2713	BUG_ON(virt_to_cache(objp) != cachep);
2714
2715	objp -= obj_offset(cachep);
2716	kfree_debugcheck(objp);
2717	page = virt_to_head_page(objp);
2718
2719	if (cachep->flags & SLAB_RED_ZONE) {
2720		verify_redzone_free(cachep, objp);
2721		*dbg_redzone1(cachep, objp) = RED_INACTIVE;
2722		*dbg_redzone2(cachep, objp) = RED_INACTIVE;
2723	}
2724	if (cachep->flags & SLAB_STORE_USER)
 
2725		*dbg_userword(cachep, objp) = (void *)caller;
 
2726
2727	objnr = obj_to_index(cachep, page, objp);
2728
2729	BUG_ON(objnr >= cachep->num);
2730	BUG_ON(objp != index_to_obj(cachep, page, objnr));
2731
2732	if (cachep->flags & SLAB_POISON) {
2733		poison_obj(cachep, objp, POISON_FREE);
2734		slab_kernel_map(cachep, objp, 0);
2735	}
2736	return objp;
2737}
2738
2739#else
2740#define kfree_debugcheck(x) do { } while(0)
2741#define cache_free_debugcheck(x, objp, z) (objp)
2742#endif
2743
2744static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
2745						void **list)
2746{
2747#if DEBUG
2748	void *next = *list;
2749	void *objp;
2750
2751	while (next) {
2752		objp = next - obj_offset(cachep);
2753		next = *(void **)next;
2754		poison_obj(cachep, objp, POISON_FREE);
2755	}
2756#endif
2757}
2758
2759static inline void fixup_slab_list(struct kmem_cache *cachep,
2760				struct kmem_cache_node *n, struct page *page,
2761				void **list)
2762{
2763	/* move slabp to correct slabp list: */
2764	list_del(&page->slab_list);
2765	if (page->active == cachep->num) {
2766		list_add(&page->slab_list, &n->slabs_full);
2767		if (OBJFREELIST_SLAB(cachep)) {
2768#if DEBUG
2769			/* Poisoning will be done without holding the lock */
2770			if (cachep->flags & SLAB_POISON) {
2771				void **objp = page->freelist;
2772
2773				*objp = *list;
2774				*list = objp;
2775			}
2776#endif
2777			page->freelist = NULL;
2778		}
2779	} else
2780		list_add(&page->slab_list, &n->slabs_partial);
2781}
2782
2783/* Try to find non-pfmemalloc slab if needed */
2784static noinline struct page *get_valid_first_slab(struct kmem_cache_node *n,
2785					struct page *page, bool pfmemalloc)
2786{
2787	if (!page)
2788		return NULL;
2789
2790	if (pfmemalloc)
2791		return page;
2792
2793	if (!PageSlabPfmemalloc(page))
2794		return page;
2795
2796	/* No need to keep pfmemalloc slab if we have enough free objects */
2797	if (n->free_objects > n->free_limit) {
2798		ClearPageSlabPfmemalloc(page);
2799		return page;
2800	}
2801
2802	/* Move pfmemalloc slab to the end of list to speed up next search */
2803	list_del(&page->slab_list);
2804	if (!page->active) {
2805		list_add_tail(&page->slab_list, &n->slabs_free);
2806		n->free_slabs++;
2807	} else
2808		list_add_tail(&page->slab_list, &n->slabs_partial);
2809
2810	list_for_each_entry(page, &n->slabs_partial, slab_list) {
2811		if (!PageSlabPfmemalloc(page))
2812			return page;
2813	}
2814
2815	n->free_touched = 1;
2816	list_for_each_entry(page, &n->slabs_free, slab_list) {
2817		if (!PageSlabPfmemalloc(page)) {
2818			n->free_slabs--;
2819			return page;
2820		}
2821	}
2822
2823	return NULL;
2824}
2825
2826static struct page *get_first_slab(struct kmem_cache_node *n, bool pfmemalloc)
2827{
2828	struct page *page;
2829
2830	assert_spin_locked(&n->list_lock);
2831	page = list_first_entry_or_null(&n->slabs_partial, struct page,
2832					slab_list);
2833	if (!page) {
2834		n->free_touched = 1;
2835		page = list_first_entry_or_null(&n->slabs_free, struct page,
2836						slab_list);
2837		if (page)
2838			n->free_slabs--;
2839	}
2840
2841	if (sk_memalloc_socks())
2842		page = get_valid_first_slab(n, page, pfmemalloc);
2843
2844	return page;
2845}
2846
2847static noinline void *cache_alloc_pfmemalloc(struct kmem_cache *cachep,
2848				struct kmem_cache_node *n, gfp_t flags)
2849{
2850	struct page *page;
2851	void *obj;
2852	void *list = NULL;
2853
2854	if (!gfp_pfmemalloc_allowed(flags))
2855		return NULL;
2856
2857	spin_lock(&n->list_lock);
2858	page = get_first_slab(n, true);
2859	if (!page) {
2860		spin_unlock(&n->list_lock);
2861		return NULL;
2862	}
2863
2864	obj = slab_get_obj(cachep, page);
2865	n->free_objects--;
2866
2867	fixup_slab_list(cachep, n, page, &list);
2868
2869	spin_unlock(&n->list_lock);
2870	fixup_objfreelist_debug(cachep, &list);
2871
2872	return obj;
2873}
2874
2875/*
2876 * Slab list should be fixed up by fixup_slab_list() for existing slab
2877 * or cache_grow_end() for new slab
2878 */
2879static __always_inline int alloc_block(struct kmem_cache *cachep,
2880		struct array_cache *ac, struct page *page, int batchcount)
2881{
2882	/*
2883	 * There must be at least one object available for
2884	 * allocation.
2885	 */
2886	BUG_ON(page->active >= cachep->num);
2887
2888	while (page->active < cachep->num && batchcount--) {
2889		STATS_INC_ALLOCED(cachep);
2890		STATS_INC_ACTIVE(cachep);
2891		STATS_SET_HIGH(cachep);
2892
2893		ac->entry[ac->avail++] = slab_get_obj(cachep, page);
2894	}
2895
2896	return batchcount;
2897}
2898
2899static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2900{
2901	int batchcount;
2902	struct kmem_cache_node *n;
2903	struct array_cache *ac, *shared;
2904	int node;
2905	void *list = NULL;
2906	struct page *page;
2907
2908	check_irq_off();
2909	node = numa_mem_id();
2910
2911	ac = cpu_cache_get(cachep);
2912	batchcount = ac->batchcount;
2913	if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2914		/*
2915		 * If there was little recent activity on this cache, then
2916		 * perform only a partial refill.  Otherwise we could generate
2917		 * refill bouncing.
2918		 */
2919		batchcount = BATCHREFILL_LIMIT;
2920	}
2921	n = get_node(cachep, node);
2922
2923	BUG_ON(ac->avail > 0 || !n);
2924	shared = READ_ONCE(n->shared);
2925	if (!n->free_objects && (!shared || !shared->avail))
2926		goto direct_grow;
2927
2928	spin_lock(&n->list_lock);
2929	shared = READ_ONCE(n->shared);
2930
2931	/* See if we can refill from the shared array */
2932	if (shared && transfer_objects(ac, shared, batchcount)) {
2933		shared->touched = 1;
2934		goto alloc_done;
2935	}
2936
2937	while (batchcount > 0) {
2938		/* Get slab alloc is to come from. */
2939		page = get_first_slab(n, false);
2940		if (!page)
2941			goto must_grow;
2942
2943		check_spinlock_acquired(cachep);
2944
2945		batchcount = alloc_block(cachep, ac, page, batchcount);
2946		fixup_slab_list(cachep, n, page, &list);
2947	}
2948
2949must_grow:
2950	n->free_objects -= ac->avail;
2951alloc_done:
2952	spin_unlock(&n->list_lock);
2953	fixup_objfreelist_debug(cachep, &list);
2954
2955direct_grow:
2956	if (unlikely(!ac->avail)) {
2957		/* Check if we can use obj in pfmemalloc slab */
2958		if (sk_memalloc_socks()) {
2959			void *obj = cache_alloc_pfmemalloc(cachep, n, flags);
2960
2961			if (obj)
2962				return obj;
2963		}
2964
2965		page = cache_grow_begin(cachep, gfp_exact_node(flags), node);
2966
2967		/*
2968		 * cache_grow_begin() can reenable interrupts,
2969		 * then ac could change.
2970		 */
2971		ac = cpu_cache_get(cachep);
2972		if (!ac->avail && page)
2973			alloc_block(cachep, ac, page, batchcount);
2974		cache_grow_end(cachep, page);
2975
2976		if (!ac->avail)
2977			return NULL;
2978	}
2979	ac->touched = 1;
2980
2981	return ac->entry[--ac->avail];
2982}
2983
2984static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2985						gfp_t flags)
2986{
2987	might_sleep_if(gfpflags_allow_blocking(flags));
2988}
2989
2990#if DEBUG
2991static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2992				gfp_t flags, void *objp, unsigned long caller)
2993{
2994	WARN_ON_ONCE(cachep->ctor && (flags & __GFP_ZERO));
2995	if (!objp || is_kfence_address(objp))
2996		return objp;
2997	if (cachep->flags & SLAB_POISON) {
2998		check_poison_obj(cachep, objp);
2999		slab_kernel_map(cachep, objp, 1);
3000		poison_obj(cachep, objp, POISON_INUSE);
3001	}
3002	if (cachep->flags & SLAB_STORE_USER)
3003		*dbg_userword(cachep, objp) = (void *)caller;
3004
3005	if (cachep->flags & SLAB_RED_ZONE) {
3006		if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3007				*dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3008			slab_error(cachep, "double free, or memory outside object was overwritten");
3009			pr_err("%px: redzone 1:0x%llx, redzone 2:0x%llx\n",
3010			       objp, *dbg_redzone1(cachep, objp),
3011			       *dbg_redzone2(cachep, objp));
3012		}
3013		*dbg_redzone1(cachep, objp) = RED_ACTIVE;
3014		*dbg_redzone2(cachep, objp) = RED_ACTIVE;
3015	}
3016
3017	objp += obj_offset(cachep);
3018	if (cachep->ctor && cachep->flags & SLAB_POISON)
3019		cachep->ctor(objp);
3020	if (ARCH_SLAB_MINALIGN &&
3021	    ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3022		pr_err("0x%px: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3023		       objp, (int)ARCH_SLAB_MINALIGN);
3024	}
3025	return objp;
3026}
3027#else
3028#define cache_alloc_debugcheck_after(a, b, objp, d) (objp)
3029#endif
3030
3031static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3032{
3033	void *objp;
3034	struct array_cache *ac;
3035
3036	check_irq_off();
3037
3038	ac = cpu_cache_get(cachep);
3039	if (likely(ac->avail)) {
3040		ac->touched = 1;
3041		objp = ac->entry[--ac->avail];
3042
3043		STATS_INC_ALLOCHIT(cachep);
3044		goto out;
3045	}
3046
3047	STATS_INC_ALLOCMISS(cachep);
3048	objp = cache_alloc_refill(cachep, flags);
3049	/*
3050	 * the 'ac' may be updated by cache_alloc_refill(),
3051	 * and kmemleak_erase() requires its correct value.
3052	 */
3053	ac = cpu_cache_get(cachep);
3054
3055out:
3056	/*
3057	 * To avoid a false negative, if an object that is in one of the
3058	 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3059	 * treat the array pointers as a reference to the object.
3060	 */
3061	if (objp)
3062		kmemleak_erase(&ac->entry[ac->avail]);
3063	return objp;
3064}
3065
3066#ifdef CONFIG_NUMA
3067/*
3068 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
3069 *
3070 * If we are in_interrupt, then process context, including cpusets and
3071 * mempolicy, may not apply and should not be used for allocation policy.
3072 */
3073static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3074{
3075	int nid_alloc, nid_here;
3076
3077	if (in_interrupt() || (flags & __GFP_THISNODE))
3078		return NULL;
3079	nid_alloc = nid_here = numa_mem_id();
3080	if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3081		nid_alloc = cpuset_slab_spread_node();
3082	else if (current->mempolicy)
3083		nid_alloc = mempolicy_slab_node();
3084	if (nid_alloc != nid_here)
3085		return ____cache_alloc_node(cachep, flags, nid_alloc);
3086	return NULL;
3087}
3088
3089/*
3090 * Fallback function if there was no memory available and no objects on a
3091 * certain node and fall back is permitted. First we scan all the
3092 * available node for available objects. If that fails then we
3093 * perform an allocation without specifying a node. This allows the page
3094 * allocator to do its reclaim / fallback magic. We then insert the
3095 * slab into the proper nodelist and then allocate from it.
3096 */
3097static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3098{
3099	struct zonelist *zonelist;
3100	struct zoneref *z;
3101	struct zone *zone;
3102	enum zone_type highest_zoneidx = gfp_zone(flags);
3103	void *obj = NULL;
3104	struct page *page;
3105	int nid;
3106	unsigned int cpuset_mems_cookie;
3107
3108	if (flags & __GFP_THISNODE)
3109		return NULL;
3110
3111retry_cpuset:
3112	cpuset_mems_cookie = read_mems_allowed_begin();
3113	zonelist = node_zonelist(mempolicy_slab_node(), flags);
3114
3115retry:
3116	/*
3117	 * Look through allowed nodes for objects available
3118	 * from existing per node queues.
3119	 */
3120	for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) {
3121		nid = zone_to_nid(zone);
3122
3123		if (cpuset_zone_allowed(zone, flags) &&
3124			get_node(cache, nid) &&
3125			get_node(cache, nid)->free_objects) {
3126				obj = ____cache_alloc_node(cache,
3127					gfp_exact_node(flags), nid);
3128				if (obj)
3129					break;
3130		}
3131	}
3132
3133	if (!obj) {
3134		/*
3135		 * This allocation will be performed within the constraints
3136		 * of the current cpuset / memory policy requirements.
3137		 * We may trigger various forms of reclaim on the allowed
3138		 * set and go into memory reserves if necessary.
3139		 */
3140		page = cache_grow_begin(cache, flags, numa_mem_id());
3141		cache_grow_end(cache, page);
3142		if (page) {
3143			nid = page_to_nid(page);
3144			obj = ____cache_alloc_node(cache,
3145				gfp_exact_node(flags), nid);
3146
3147			/*
3148			 * Another processor may allocate the objects in
3149			 * the slab since we are not holding any locks.
3150			 */
3151			if (!obj)
3152				goto retry;
3153		}
3154	}
3155
3156	if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
3157		goto retry_cpuset;
3158	return obj;
3159}
3160
3161/*
3162 * A interface to enable slab creation on nodeid
3163 */
3164static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3165				int nodeid)
3166{
3167	struct page *page;
3168	struct kmem_cache_node *n;
3169	void *obj = NULL;
3170	void *list = NULL;
3171
3172	VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES);
3173	n = get_node(cachep, nodeid);
3174	BUG_ON(!n);
3175
3176	check_irq_off();
3177	spin_lock(&n->list_lock);
3178	page = get_first_slab(n, false);
3179	if (!page)
3180		goto must_grow;
3181
3182	check_spinlock_acquired_node(cachep, nodeid);
3183
3184	STATS_INC_NODEALLOCS(cachep);
3185	STATS_INC_ACTIVE(cachep);
3186	STATS_SET_HIGH(cachep);
3187
3188	BUG_ON(page->active == cachep->num);
3189
3190	obj = slab_get_obj(cachep, page);
3191	n->free_objects--;
3192
3193	fixup_slab_list(cachep, n, page, &list);
3194
3195	spin_unlock(&n->list_lock);
3196	fixup_objfreelist_debug(cachep, &list);
3197	return obj;
3198
3199must_grow:
3200	spin_unlock(&n->list_lock);
3201	page = cache_grow_begin(cachep, gfp_exact_node(flags), nodeid);
3202	if (page) {
3203		/* This slab isn't counted yet so don't update free_objects */
3204		obj = slab_get_obj(cachep, page);
3205	}
3206	cache_grow_end(cachep, page);
3207
3208	return obj ? obj : fallback_alloc(cachep, flags);
3209}
3210
3211static __always_inline void *
3212slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid, size_t orig_size,
3213		   unsigned long caller)
3214{
3215	unsigned long save_flags;
3216	void *ptr;
3217	int slab_node = numa_mem_id();
3218	struct obj_cgroup *objcg = NULL;
3219	bool init = false;
3220
3221	flags &= gfp_allowed_mask;
3222	cachep = slab_pre_alloc_hook(cachep, &objcg, 1, flags);
3223	if (unlikely(!cachep))
3224		return NULL;
3225
3226	ptr = kfence_alloc(cachep, orig_size, flags);
3227	if (unlikely(ptr))
3228		goto out_hooks;
3229
3230	cache_alloc_debugcheck_before(cachep, flags);
3231	local_irq_save(save_flags);
3232
3233	if (nodeid == NUMA_NO_NODE)
3234		nodeid = slab_node;
3235
3236	if (unlikely(!get_node(cachep, nodeid))) {
3237		/* Node not bootstrapped yet */
3238		ptr = fallback_alloc(cachep, flags);
3239		goto out;
3240	}
3241
3242	if (nodeid == slab_node) {
3243		/*
3244		 * Use the locally cached objects if possible.
3245		 * However ____cache_alloc does not allow fallback
3246		 * to other nodes. It may fail while we still have
3247		 * objects on other nodes available.
3248		 */
3249		ptr = ____cache_alloc(cachep, flags);
3250		if (ptr)
3251			goto out;
3252	}
3253	/* ___cache_alloc_node can fall back to other nodes */
3254	ptr = ____cache_alloc_node(cachep, flags, nodeid);
3255  out:
3256	local_irq_restore(save_flags);
3257	ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3258	init = slab_want_init_on_alloc(flags, cachep);
3259
3260out_hooks:
3261	slab_post_alloc_hook(cachep, objcg, flags, 1, &ptr, init);
 
 
3262	return ptr;
3263}
3264
3265static __always_inline void *
3266__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3267{
3268	void *objp;
3269
3270	if (current->mempolicy || cpuset_do_slab_mem_spread()) {
3271		objp = alternate_node_alloc(cache, flags);
3272		if (objp)
3273			goto out;
3274	}
3275	objp = ____cache_alloc(cache, flags);
3276
3277	/*
3278	 * We may just have run out of memory on the local node.
3279	 * ____cache_alloc_node() knows how to locate memory on other nodes
3280	 */
3281	if (!objp)
3282		objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3283
3284  out:
3285	return objp;
3286}
3287#else
3288
3289static __always_inline void *
3290__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3291{
3292	return ____cache_alloc(cachep, flags);
3293}
3294
3295#endif /* CONFIG_NUMA */
3296
3297static __always_inline void *
3298slab_alloc(struct kmem_cache *cachep, gfp_t flags, size_t orig_size, unsigned long caller)
3299{
3300	unsigned long save_flags;
3301	void *objp;
3302	struct obj_cgroup *objcg = NULL;
3303	bool init = false;
3304
3305	flags &= gfp_allowed_mask;
3306	cachep = slab_pre_alloc_hook(cachep, &objcg, 1, flags);
3307	if (unlikely(!cachep))
3308		return NULL;
3309
3310	objp = kfence_alloc(cachep, orig_size, flags);
3311	if (unlikely(objp))
3312		goto out;
3313
3314	cache_alloc_debugcheck_before(cachep, flags);
3315	local_irq_save(save_flags);
3316	objp = __do_cache_alloc(cachep, flags);
3317	local_irq_restore(save_flags);
3318	objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3319	prefetchw(objp);
3320	init = slab_want_init_on_alloc(flags, cachep);
3321
3322out:
3323	slab_post_alloc_hook(cachep, objcg, flags, 1, &objp, init);
 
 
3324	return objp;
3325}
3326
3327/*
3328 * Caller needs to acquire correct kmem_cache_node's list_lock
3329 * @list: List of detached free slabs should be freed by caller
3330 */
3331static void free_block(struct kmem_cache *cachep, void **objpp,
3332			int nr_objects, int node, struct list_head *list)
3333{
3334	int i;
3335	struct kmem_cache_node *n = get_node(cachep, node);
3336	struct page *page;
3337
3338	n->free_objects += nr_objects;
3339
3340	for (i = 0; i < nr_objects; i++) {
3341		void *objp;
3342		struct page *page;
3343
3344		objp = objpp[i];
3345
3346		page = virt_to_head_page(objp);
3347		list_del(&page->slab_list);
3348		check_spinlock_acquired_node(cachep, node);
3349		slab_put_obj(cachep, page, objp);
3350		STATS_DEC_ACTIVE(cachep);
3351
3352		/* fixup slab chains */
3353		if (page->active == 0) {
3354			list_add(&page->slab_list, &n->slabs_free);
3355			n->free_slabs++;
3356		} else {
3357			/* Unconditionally move a slab to the end of the
3358			 * partial list on free - maximum time for the
3359			 * other objects to be freed, too.
3360			 */
3361			list_add_tail(&page->slab_list, &n->slabs_partial);
3362		}
3363	}
3364
3365	while (n->free_objects > n->free_limit && !list_empty(&n->slabs_free)) {
3366		n->free_objects -= cachep->num;
3367
3368		page = list_last_entry(&n->slabs_free, struct page, slab_list);
3369		list_move(&page->slab_list, list);
3370		n->free_slabs--;
3371		n->total_slabs--;
3372	}
3373}
3374
3375static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3376{
3377	int batchcount;
3378	struct kmem_cache_node *n;
3379	int node = numa_mem_id();
3380	LIST_HEAD(list);
3381
3382	batchcount = ac->batchcount;
3383
3384	check_irq_off();
3385	n = get_node(cachep, node);
3386	spin_lock(&n->list_lock);
3387	if (n->shared) {
3388		struct array_cache *shared_array = n->shared;
3389		int max = shared_array->limit - shared_array->avail;
3390		if (max) {
3391			if (batchcount > max)
3392				batchcount = max;
3393			memcpy(&(shared_array->entry[shared_array->avail]),
3394			       ac->entry, sizeof(void *) * batchcount);
3395			shared_array->avail += batchcount;
3396			goto free_done;
3397		}
3398	}
3399
3400	free_block(cachep, ac->entry, batchcount, node, &list);
3401free_done:
3402#if STATS
3403	{
3404		int i = 0;
3405		struct page *page;
3406
3407		list_for_each_entry(page, &n->slabs_free, slab_list) {
3408			BUG_ON(page->active);
3409
3410			i++;
3411		}
3412		STATS_SET_FREEABLE(cachep, i);
3413	}
3414#endif
3415	spin_unlock(&n->list_lock);
 
3416	ac->avail -= batchcount;
3417	memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3418	slabs_destroy(cachep, &list);
3419}
3420
3421/*
3422 * Release an obj back to its cache. If the obj has a constructed state, it must
3423 * be in this state _before_ it is released.  Called with disabled ints.
3424 */
3425static __always_inline void __cache_free(struct kmem_cache *cachep, void *objp,
3426					 unsigned long caller)
3427{
3428	bool init;
3429
3430	if (is_kfence_address(objp)) {
3431		kmemleak_free_recursive(objp, cachep->flags);
3432		__kfence_free(objp);
3433		return;
3434	}
3435
3436	/*
3437	 * As memory initialization might be integrated into KASAN,
3438	 * kasan_slab_free and initialization memset must be
3439	 * kept together to avoid discrepancies in behavior.
3440	 */
3441	init = slab_want_init_on_free(cachep);
3442	if (init && !kasan_has_integrated_init())
3443		memset(objp, 0, cachep->object_size);
3444	/* KASAN might put objp into memory quarantine, delaying its reuse. */
3445	if (kasan_slab_free(cachep, objp, init))
3446		return;
3447
3448	/* Use KCSAN to help debug racy use-after-free. */
3449	if (!(cachep->flags & SLAB_TYPESAFE_BY_RCU))
3450		__kcsan_check_access(objp, cachep->object_size,
3451				     KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT);
3452
3453	___cache_free(cachep, objp, caller);
3454}
3455
3456void ___cache_free(struct kmem_cache *cachep, void *objp,
3457		unsigned long caller)
3458{
3459	struct array_cache *ac = cpu_cache_get(cachep);
3460
3461	check_irq_off();
3462	kmemleak_free_recursive(objp, cachep->flags);
3463	objp = cache_free_debugcheck(cachep, objp, caller);
3464	memcg_slab_free_hook(cachep, &objp, 1);
3465
3466	/*
3467	 * Skip calling cache_free_alien() when the platform is not numa.
3468	 * This will avoid cache misses that happen while accessing slabp (which
3469	 * is per page memory  reference) to get nodeid. Instead use a global
3470	 * variable to skip the call, which is mostly likely to be present in
3471	 * the cache.
3472	 */
3473	if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3474		return;
3475
3476	if (ac->avail < ac->limit) {
3477		STATS_INC_FREEHIT(cachep);
3478	} else {
3479		STATS_INC_FREEMISS(cachep);
3480		cache_flusharray(cachep, ac);
3481	}
3482
3483	if (sk_memalloc_socks()) {
3484		struct page *page = virt_to_head_page(objp);
3485
3486		if (unlikely(PageSlabPfmemalloc(page))) {
3487			cache_free_pfmemalloc(cachep, page, objp);
3488			return;
3489		}
3490	}
3491
3492	__free_one(ac, objp);
3493}
3494
3495/**
3496 * kmem_cache_alloc - Allocate an object
3497 * @cachep: The cache to allocate from.
3498 * @flags: See kmalloc().
3499 *
3500 * Allocate an object from this cache.  The flags are only relevant
3501 * if the cache has no available objects.
3502 *
3503 * Return: pointer to the new object or %NULL in case of error
3504 */
3505void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3506{
3507	void *ret = slab_alloc(cachep, flags, cachep->object_size, _RET_IP_);
3508
 
3509	trace_kmem_cache_alloc(_RET_IP_, ret,
3510			       cachep->object_size, cachep->size, flags);
3511
3512	return ret;
3513}
3514EXPORT_SYMBOL(kmem_cache_alloc);
3515
3516static __always_inline void
3517cache_alloc_debugcheck_after_bulk(struct kmem_cache *s, gfp_t flags,
3518				  size_t size, void **p, unsigned long caller)
3519{
3520	size_t i;
3521
3522	for (i = 0; i < size; i++)
3523		p[i] = cache_alloc_debugcheck_after(s, flags, p[i], caller);
3524}
3525
3526int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3527			  void **p)
3528{
3529	size_t i;
3530	struct obj_cgroup *objcg = NULL;
3531
3532	s = slab_pre_alloc_hook(s, &objcg, size, flags);
3533	if (!s)
3534		return 0;
3535
3536	cache_alloc_debugcheck_before(s, flags);
3537
3538	local_irq_disable();
3539	for (i = 0; i < size; i++) {
3540		void *objp = kfence_alloc(s, s->object_size, flags) ?: __do_cache_alloc(s, flags);
3541
3542		if (unlikely(!objp))
3543			goto error;
3544		p[i] = objp;
3545	}
3546	local_irq_enable();
3547
3548	cache_alloc_debugcheck_after_bulk(s, flags, size, p, _RET_IP_);
3549
3550	/*
3551	 * memcg and kmem_cache debug support and memory initialization.
3552	 * Done outside of the IRQ disabled section.
3553	 */
3554	slab_post_alloc_hook(s, objcg, flags, size, p,
3555				slab_want_init_on_alloc(flags, s));
3556	/* FIXME: Trace call missing. Christoph would like a bulk variant */
3557	return size;
3558error:
3559	local_irq_enable();
3560	cache_alloc_debugcheck_after_bulk(s, flags, i, p, _RET_IP_);
3561	slab_post_alloc_hook(s, objcg, flags, i, p, false);
3562	__kmem_cache_free_bulk(s, i, p);
3563	return 0;
3564}
3565EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3566
3567#ifdef CONFIG_TRACING
3568void *
3569kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3570{
3571	void *ret;
3572
3573	ret = slab_alloc(cachep, flags, size, _RET_IP_);
3574
3575	ret = kasan_kmalloc(cachep, ret, size, flags);
3576	trace_kmalloc(_RET_IP_, ret,
3577		      size, cachep->size, flags);
3578	return ret;
3579}
3580EXPORT_SYMBOL(kmem_cache_alloc_trace);
3581#endif
3582
3583#ifdef CONFIG_NUMA
3584/**
3585 * kmem_cache_alloc_node - Allocate an object on the specified node
3586 * @cachep: The cache to allocate from.
3587 * @flags: See kmalloc().
3588 * @nodeid: node number of the target node.
3589 *
3590 * Identical to kmem_cache_alloc but it will allocate memory on the given
3591 * node, which can improve the performance for cpu bound structures.
3592 *
3593 * Fallback to other node is possible if __GFP_THISNODE is not set.
3594 *
3595 * Return: pointer to the new object or %NULL in case of error
3596 */
3597void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3598{
3599	void *ret = slab_alloc_node(cachep, flags, nodeid, cachep->object_size, _RET_IP_);
3600
 
3601	trace_kmem_cache_alloc_node(_RET_IP_, ret,
3602				    cachep->object_size, cachep->size,
3603				    flags, nodeid);
3604
3605	return ret;
3606}
3607EXPORT_SYMBOL(kmem_cache_alloc_node);
3608
3609#ifdef CONFIG_TRACING
3610void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3611				  gfp_t flags,
3612				  int nodeid,
3613				  size_t size)
3614{
3615	void *ret;
3616
3617	ret = slab_alloc_node(cachep, flags, nodeid, size, _RET_IP_);
3618
3619	ret = kasan_kmalloc(cachep, ret, size, flags);
3620	trace_kmalloc_node(_RET_IP_, ret,
3621			   size, cachep->size,
3622			   flags, nodeid);
3623	return ret;
3624}
3625EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3626#endif
3627
3628static __always_inline void *
3629__do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3630{
3631	struct kmem_cache *cachep;
3632	void *ret;
3633
3634	if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3635		return NULL;
3636	cachep = kmalloc_slab(size, flags);
3637	if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3638		return cachep;
3639	ret = kmem_cache_alloc_node_trace(cachep, flags, node, size);
3640	ret = kasan_kmalloc(cachep, ret, size, flags);
3641
3642	return ret;
3643}
3644
3645void *__kmalloc_node(size_t size, gfp_t flags, int node)
3646{
3647	return __do_kmalloc_node(size, flags, node, _RET_IP_);
3648}
3649EXPORT_SYMBOL(__kmalloc_node);
3650
3651void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3652		int node, unsigned long caller)
3653{
3654	return __do_kmalloc_node(size, flags, node, caller);
3655}
3656EXPORT_SYMBOL(__kmalloc_node_track_caller);
3657#endif /* CONFIG_NUMA */
3658
3659#ifdef CONFIG_PRINTK
3660void kmem_obj_info(struct kmem_obj_info *kpp, void *object, struct page *page)
3661{
3662	struct kmem_cache *cachep;
3663	unsigned int objnr;
3664	void *objp;
3665
3666	kpp->kp_ptr = object;
3667	kpp->kp_page = page;
3668	cachep = page->slab_cache;
3669	kpp->kp_slab_cache = cachep;
3670	objp = object - obj_offset(cachep);
3671	kpp->kp_data_offset = obj_offset(cachep);
3672	page = virt_to_head_page(objp);
3673	objnr = obj_to_index(cachep, page, objp);
3674	objp = index_to_obj(cachep, page, objnr);
3675	kpp->kp_objp = objp;
3676	if (DEBUG && cachep->flags & SLAB_STORE_USER)
3677		kpp->kp_ret = *dbg_userword(cachep, objp);
3678}
3679#endif
3680
3681/**
3682 * __do_kmalloc - allocate memory
3683 * @size: how many bytes of memory are required.
3684 * @flags: the type of memory to allocate (see kmalloc).
3685 * @caller: function caller for debug tracking of the caller
3686 *
3687 * Return: pointer to the allocated memory or %NULL in case of error
3688 */
3689static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3690					  unsigned long caller)
3691{
3692	struct kmem_cache *cachep;
3693	void *ret;
3694
3695	if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3696		return NULL;
3697	cachep = kmalloc_slab(size, flags);
3698	if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3699		return cachep;
3700	ret = slab_alloc(cachep, flags, size, caller);
3701
3702	ret = kasan_kmalloc(cachep, ret, size, flags);
3703	trace_kmalloc(caller, ret,
3704		      size, cachep->size, flags);
3705
3706	return ret;
3707}
3708
3709void *__kmalloc(size_t size, gfp_t flags)
3710{
3711	return __do_kmalloc(size, flags, _RET_IP_);
3712}
3713EXPORT_SYMBOL(__kmalloc);
3714
3715void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3716{
3717	return __do_kmalloc(size, flags, caller);
3718}
3719EXPORT_SYMBOL(__kmalloc_track_caller);
3720
3721/**
3722 * kmem_cache_free - Deallocate an object
3723 * @cachep: The cache the allocation was from.
3724 * @objp: The previously allocated object.
3725 *
3726 * Free an object which was previously allocated from this
3727 * cache.
3728 */
3729void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3730{
3731	unsigned long flags;
3732	cachep = cache_from_obj(cachep, objp);
3733	if (!cachep)
3734		return;
3735
3736	local_irq_save(flags);
3737	debug_check_no_locks_freed(objp, cachep->object_size);
3738	if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3739		debug_check_no_obj_freed(objp, cachep->object_size);
3740	__cache_free(cachep, objp, _RET_IP_);
3741	local_irq_restore(flags);
3742
3743	trace_kmem_cache_free(_RET_IP_, objp, cachep->name);
3744}
3745EXPORT_SYMBOL(kmem_cache_free);
3746
3747void kmem_cache_free_bulk(struct kmem_cache *orig_s, size_t size, void **p)
3748{
3749	struct kmem_cache *s;
3750	size_t i;
3751
3752	local_irq_disable();
3753	for (i = 0; i < size; i++) {
3754		void *objp = p[i];
3755
3756		if (!orig_s) /* called via kfree_bulk */
3757			s = virt_to_cache(objp);
3758		else
3759			s = cache_from_obj(orig_s, objp);
3760		if (!s)
3761			continue;
3762
3763		debug_check_no_locks_freed(objp, s->object_size);
3764		if (!(s->flags & SLAB_DEBUG_OBJECTS))
3765			debug_check_no_obj_freed(objp, s->object_size);
3766
3767		__cache_free(s, objp, _RET_IP_);
3768	}
3769	local_irq_enable();
3770
3771	/* FIXME: add tracing */
3772}
3773EXPORT_SYMBOL(kmem_cache_free_bulk);
3774
3775/**
3776 * kfree - free previously allocated memory
3777 * @objp: pointer returned by kmalloc.
3778 *
3779 * If @objp is NULL, no operation is performed.
3780 *
3781 * Don't free memory not originally allocated by kmalloc()
3782 * or you will run into trouble.
3783 */
3784void kfree(const void *objp)
3785{
3786	struct kmem_cache *c;
3787	unsigned long flags;
3788
3789	trace_kfree(_RET_IP_, objp);
3790
3791	if (unlikely(ZERO_OR_NULL_PTR(objp)))
3792		return;
3793	local_irq_save(flags);
3794	kfree_debugcheck(objp);
3795	c = virt_to_cache(objp);
3796	if (!c) {
3797		local_irq_restore(flags);
3798		return;
3799	}
3800	debug_check_no_locks_freed(objp, c->object_size);
3801
3802	debug_check_no_obj_freed(objp, c->object_size);
3803	__cache_free(c, (void *)objp, _RET_IP_);
3804	local_irq_restore(flags);
3805}
3806EXPORT_SYMBOL(kfree);
3807
3808/*
3809 * This initializes kmem_cache_node or resizes various caches for all nodes.
3810 */
3811static int setup_kmem_cache_nodes(struct kmem_cache *cachep, gfp_t gfp)
3812{
3813	int ret;
3814	int node;
3815	struct kmem_cache_node *n;
3816
3817	for_each_online_node(node) {
3818		ret = setup_kmem_cache_node(cachep, node, gfp, true);
3819		if (ret)
3820			goto fail;
3821
3822	}
3823
3824	return 0;
3825
3826fail:
3827	if (!cachep->list.next) {
3828		/* Cache is not active yet. Roll back what we did */
3829		node--;
3830		while (node >= 0) {
3831			n = get_node(cachep, node);
3832			if (n) {
3833				kfree(n->shared);
3834				free_alien_cache(n->alien);
3835				kfree(n);
3836				cachep->node[node] = NULL;
3837			}
3838			node--;
3839		}
3840	}
3841	return -ENOMEM;
3842}
3843
3844/* Always called with the slab_mutex held */
3845static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3846			    int batchcount, int shared, gfp_t gfp)
3847{
3848	struct array_cache __percpu *cpu_cache, *prev;
3849	int cpu;
3850
3851	cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount);
3852	if (!cpu_cache)
3853		return -ENOMEM;
3854
3855	prev = cachep->cpu_cache;
3856	cachep->cpu_cache = cpu_cache;
3857	/*
3858	 * Without a previous cpu_cache there's no need to synchronize remote
3859	 * cpus, so skip the IPIs.
3860	 */
3861	if (prev)
3862		kick_all_cpus_sync();
3863
3864	check_irq_on();
3865	cachep->batchcount = batchcount;
3866	cachep->limit = limit;
3867	cachep->shared = shared;
3868
3869	if (!prev)
3870		goto setup_node;
3871
3872	for_each_online_cpu(cpu) {
3873		LIST_HEAD(list);
3874		int node;
3875		struct kmem_cache_node *n;
3876		struct array_cache *ac = per_cpu_ptr(prev, cpu);
3877
3878		node = cpu_to_mem(cpu);
3879		n = get_node(cachep, node);
3880		spin_lock_irq(&n->list_lock);
3881		free_block(cachep, ac->entry, ac->avail, node, &list);
3882		spin_unlock_irq(&n->list_lock);
3883		slabs_destroy(cachep, &list);
3884	}
3885	free_percpu(prev);
3886
3887setup_node:
3888	return setup_kmem_cache_nodes(cachep, gfp);
3889}
3890
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3891/* Called with slab_mutex held always */
3892static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3893{
3894	int err;
3895	int limit = 0;
3896	int shared = 0;
3897	int batchcount = 0;
3898
3899	err = cache_random_seq_create(cachep, cachep->num, gfp);
3900	if (err)
3901		goto end;
3902
 
 
 
 
 
 
 
3903	if (limit && shared && batchcount)
3904		goto skip_setup;
3905	/*
3906	 * The head array serves three purposes:
3907	 * - create a LIFO ordering, i.e. return objects that are cache-warm
3908	 * - reduce the number of spinlock operations.
3909	 * - reduce the number of linked list operations on the slab and
3910	 *   bufctl chains: array operations are cheaper.
3911	 * The numbers are guessed, we should auto-tune as described by
3912	 * Bonwick.
3913	 */
3914	if (cachep->size > 131072)
3915		limit = 1;
3916	else if (cachep->size > PAGE_SIZE)
3917		limit = 8;
3918	else if (cachep->size > 1024)
3919		limit = 24;
3920	else if (cachep->size > 256)
3921		limit = 54;
3922	else
3923		limit = 120;
3924
3925	/*
3926	 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3927	 * allocation behaviour: Most allocs on one cpu, most free operations
3928	 * on another cpu. For these cases, an efficient object passing between
3929	 * cpus is necessary. This is provided by a shared array. The array
3930	 * replaces Bonwick's magazine layer.
3931	 * On uniprocessor, it's functionally equivalent (but less efficient)
3932	 * to a larger limit. Thus disabled by default.
3933	 */
3934	shared = 0;
3935	if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
3936		shared = 8;
3937
3938#if DEBUG
3939	/*
3940	 * With debugging enabled, large batchcount lead to excessively long
3941	 * periods with disabled local interrupts. Limit the batchcount
3942	 */
3943	if (limit > 32)
3944		limit = 32;
3945#endif
3946	batchcount = (limit + 1) / 2;
3947skip_setup:
3948	err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3949end:
3950	if (err)
3951		pr_err("enable_cpucache failed for %s, error %d\n",
3952		       cachep->name, -err);
3953	return err;
3954}
3955
3956/*
3957 * Drain an array if it contains any elements taking the node lock only if
3958 * necessary. Note that the node listlock also protects the array_cache
3959 * if drain_array() is used on the shared array.
3960 */
3961static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
3962			 struct array_cache *ac, int node)
3963{
3964	LIST_HEAD(list);
3965
3966	/* ac from n->shared can be freed if we don't hold the slab_mutex. */
3967	check_mutex_acquired();
3968
3969	if (!ac || !ac->avail)
3970		return;
3971
3972	if (ac->touched) {
3973		ac->touched = 0;
3974		return;
3975	}
3976
3977	spin_lock_irq(&n->list_lock);
3978	drain_array_locked(cachep, ac, node, false, &list);
3979	spin_unlock_irq(&n->list_lock);
3980
3981	slabs_destroy(cachep, &list);
3982}
3983
3984/**
3985 * cache_reap - Reclaim memory from caches.
3986 * @w: work descriptor
3987 *
3988 * Called from workqueue/eventd every few seconds.
3989 * Purpose:
3990 * - clear the per-cpu caches for this CPU.
3991 * - return freeable pages to the main free memory pool.
3992 *
3993 * If we cannot acquire the cache chain mutex then just give up - we'll try
3994 * again on the next iteration.
3995 */
3996static void cache_reap(struct work_struct *w)
3997{
3998	struct kmem_cache *searchp;
3999	struct kmem_cache_node *n;
4000	int node = numa_mem_id();
4001	struct delayed_work *work = to_delayed_work(w);
4002
4003	if (!mutex_trylock(&slab_mutex))
4004		/* Give up. Setup the next iteration. */
4005		goto out;
4006
4007	list_for_each_entry(searchp, &slab_caches, list) {
4008		check_irq_on();
4009
4010		/*
4011		 * We only take the node lock if absolutely necessary and we
4012		 * have established with reasonable certainty that
4013		 * we can do some work if the lock was obtained.
4014		 */
4015		n = get_node(searchp, node);
4016
4017		reap_alien(searchp, n);
4018
4019		drain_array(searchp, n, cpu_cache_get(searchp), node);
4020
4021		/*
4022		 * These are racy checks but it does not matter
4023		 * if we skip one check or scan twice.
4024		 */
4025		if (time_after(n->next_reap, jiffies))
4026			goto next;
4027
4028		n->next_reap = jiffies + REAPTIMEOUT_NODE;
4029
4030		drain_array(searchp, n, n->shared, node);
4031
4032		if (n->free_touched)
4033			n->free_touched = 0;
4034		else {
4035			int freed;
4036
4037			freed = drain_freelist(searchp, n, (n->free_limit +
4038				5 * searchp->num - 1) / (5 * searchp->num));
4039			STATS_ADD_REAPED(searchp, freed);
4040		}
4041next:
4042		cond_resched();
4043	}
4044	check_irq_on();
4045	mutex_unlock(&slab_mutex);
4046	next_reap_node();
4047out:
4048	/* Set up the next iteration */
4049	schedule_delayed_work_on(smp_processor_id(), work,
4050				round_jiffies_relative(REAPTIMEOUT_AC));
4051}
4052
4053void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
4054{
4055	unsigned long active_objs, num_objs, active_slabs;
4056	unsigned long total_slabs = 0, free_objs = 0, shared_avail = 0;
4057	unsigned long free_slabs = 0;
4058	int node;
4059	struct kmem_cache_node *n;
4060
4061	for_each_kmem_cache_node(cachep, node, n) {
4062		check_irq_on();
4063		spin_lock_irq(&n->list_lock);
4064
4065		total_slabs += n->total_slabs;
4066		free_slabs += n->free_slabs;
4067		free_objs += n->free_objects;
4068
4069		if (n->shared)
4070			shared_avail += n->shared->avail;
4071
4072		spin_unlock_irq(&n->list_lock);
4073	}
4074	num_objs = total_slabs * cachep->num;
4075	active_slabs = total_slabs - free_slabs;
4076	active_objs = num_objs - free_objs;
4077
4078	sinfo->active_objs = active_objs;
4079	sinfo->num_objs = num_objs;
4080	sinfo->active_slabs = active_slabs;
4081	sinfo->num_slabs = total_slabs;
4082	sinfo->shared_avail = shared_avail;
4083	sinfo->limit = cachep->limit;
4084	sinfo->batchcount = cachep->batchcount;
4085	sinfo->shared = cachep->shared;
4086	sinfo->objects_per_slab = cachep->num;
4087	sinfo->cache_order = cachep->gfporder;
4088}
4089
4090void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4091{
4092#if STATS
4093	{			/* node stats */
4094		unsigned long high = cachep->high_mark;
4095		unsigned long allocs = cachep->num_allocations;
4096		unsigned long grown = cachep->grown;
4097		unsigned long reaped = cachep->reaped;
4098		unsigned long errors = cachep->errors;
4099		unsigned long max_freeable = cachep->max_freeable;
4100		unsigned long node_allocs = cachep->node_allocs;
4101		unsigned long node_frees = cachep->node_frees;
4102		unsigned long overflows = cachep->node_overflow;
4103
4104		seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
4105			   allocs, high, grown,
4106			   reaped, errors, max_freeable, node_allocs,
4107			   node_frees, overflows);
4108	}
4109	/* cpu stats */
4110	{
4111		unsigned long allochit = atomic_read(&cachep->allochit);
4112		unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4113		unsigned long freehit = atomic_read(&cachep->freehit);
4114		unsigned long freemiss = atomic_read(&cachep->freemiss);
4115
4116		seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4117			   allochit, allocmiss, freehit, freemiss);
4118	}
4119#endif
4120}
4121
4122#define MAX_SLABINFO_WRITE 128
4123/**
4124 * slabinfo_write - Tuning for the slab allocator
4125 * @file: unused
4126 * @buffer: user buffer
4127 * @count: data length
4128 * @ppos: unused
4129 *
4130 * Return: %0 on success, negative error code otherwise.
4131 */
4132ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4133		       size_t count, loff_t *ppos)
4134{
4135	char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4136	int limit, batchcount, shared, res;
4137	struct kmem_cache *cachep;
4138
4139	if (count > MAX_SLABINFO_WRITE)
4140		return -EINVAL;
4141	if (copy_from_user(&kbuf, buffer, count))
4142		return -EFAULT;
4143	kbuf[MAX_SLABINFO_WRITE] = '\0';
4144
4145	tmp = strchr(kbuf, ' ');
4146	if (!tmp)
4147		return -EINVAL;
4148	*tmp = '\0';
4149	tmp++;
4150	if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4151		return -EINVAL;
4152
4153	/* Find the cache in the chain of caches. */
4154	mutex_lock(&slab_mutex);
4155	res = -EINVAL;
4156	list_for_each_entry(cachep, &slab_caches, list) {
4157		if (!strcmp(cachep->name, kbuf)) {
4158			if (limit < 1 || batchcount < 1 ||
4159					batchcount > limit || shared < 0) {
4160				res = 0;
4161			} else {
4162				res = do_tune_cpucache(cachep, limit,
4163						       batchcount, shared,
4164						       GFP_KERNEL);
4165			}
4166			break;
4167		}
4168	}
4169	mutex_unlock(&slab_mutex);
4170	if (res >= 0)
4171		res = count;
4172	return res;
4173}
4174
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4175#ifdef CONFIG_HARDENED_USERCOPY
4176/*
4177 * Rejects incorrectly sized objects and objects that are to be copied
4178 * to/from userspace but do not fall entirely within the containing slab
4179 * cache's usercopy region.
4180 *
4181 * Returns NULL if check passes, otherwise const char * to name of cache
4182 * to indicate an error.
4183 */
4184void __check_heap_object(const void *ptr, unsigned long n, struct page *page,
4185			 bool to_user)
4186{
4187	struct kmem_cache *cachep;
4188	unsigned int objnr;
4189	unsigned long offset;
4190
4191	ptr = kasan_reset_tag(ptr);
4192
4193	/* Find and validate object. */
4194	cachep = page->slab_cache;
4195	objnr = obj_to_index(cachep, page, (void *)ptr);
4196	BUG_ON(objnr >= cachep->num);
4197
4198	/* Find offset within object. */
4199	if (is_kfence_address(ptr))
4200		offset = ptr - kfence_object_start(ptr);
4201	else
4202		offset = ptr - index_to_obj(cachep, page, objnr) - obj_offset(cachep);
4203
4204	/* Allow address range falling entirely within usercopy region. */
4205	if (offset >= cachep->useroffset &&
4206	    offset - cachep->useroffset <= cachep->usersize &&
4207	    n <= cachep->useroffset - offset + cachep->usersize)
4208		return;
4209
4210	/*
4211	 * If the copy is still within the allocated object, produce
4212	 * a warning instead of rejecting the copy. This is intended
4213	 * to be a temporary method to find any missing usercopy
4214	 * whitelists.
4215	 */
4216	if (usercopy_fallback &&
4217	    offset <= cachep->object_size &&
4218	    n <= cachep->object_size - offset) {
4219		usercopy_warn("SLAB object", cachep->name, to_user, offset, n);
4220		return;
4221	}
4222
4223	usercopy_abort("SLAB object", cachep->name, to_user, offset, n);
4224}
4225#endif /* CONFIG_HARDENED_USERCOPY */
4226
4227/**
4228 * __ksize -- Uninstrumented ksize.
4229 * @objp: pointer to the object
4230 *
4231 * Unlike ksize(), __ksize() is uninstrumented, and does not provide the same
4232 * safety checks as ksize() with KASAN instrumentation enabled.
4233 *
4234 * Return: size of the actual memory used by @objp in bytes
 
 
 
 
 
 
4235 */
4236size_t __ksize(const void *objp)
4237{
4238	struct kmem_cache *c;
4239	size_t size;
4240
4241	BUG_ON(!objp);
4242	if (unlikely(objp == ZERO_SIZE_PTR))
4243		return 0;
4244
4245	c = virt_to_cache(objp);
4246	size = c ? c->object_size : 0;
 
 
 
4247
4248	return size;
4249}
4250EXPORT_SYMBOL(__ksize);