Linux Audio

Check our new training course

Loading...
Note: File does not exist in v3.1.
   1// SPDX-License-Identifier: GPL-2.0-or-later
   2
   3/*
   4 * zsmalloc memory allocator
   5 *
   6 * Copyright (C) 2011  Nitin Gupta
   7 * Copyright (C) 2012, 2013 Minchan Kim
   8 *
   9 * This code is released using a dual license strategy: BSD/GPL
  10 * You can choose the license that better fits your requirements.
  11 *
  12 * Released under the terms of 3-clause BSD License
  13 * Released under the terms of GNU General Public License Version 2.0
  14 */
  15
  16/*
  17 * Following is how we use various fields and flags of underlying
  18 * struct page(s) to form a zspage.
  19 *
  20 * Usage of struct page fields:
  21 *	page->private: points to zspage
  22 *	page->index: links together all component pages of a zspage
  23 *		For the huge page, this is always 0, so we use this field
  24 *		to store handle.
  25 *	page->page_type: PGTY_zsmalloc, lower 24 bits locate the first object
  26 *		offset in a subpage of a zspage
  27 *
  28 * Usage of struct page flags:
  29 *	PG_private: identifies the first component page
  30 *	PG_owner_priv_1: identifies the huge component page
  31 *
  32 */
  33
  34#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  35
  36/*
  37 * lock ordering:
  38 *	page_lock
  39 *	pool->migrate_lock
  40 *	class->lock
  41 *	zspage->lock
  42 */
  43
  44#include <linux/module.h>
  45#include <linux/kernel.h>
  46#include <linux/sched.h>
  47#include <linux/bitops.h>
  48#include <linux/errno.h>
  49#include <linux/highmem.h>
  50#include <linux/string.h>
  51#include <linux/slab.h>
  52#include <linux/pgtable.h>
  53#include <asm/tlbflush.h>
  54#include <linux/cpumask.h>
  55#include <linux/cpu.h>
  56#include <linux/vmalloc.h>
  57#include <linux/preempt.h>
  58#include <linux/spinlock.h>
  59#include <linux/sprintf.h>
  60#include <linux/shrinker.h>
  61#include <linux/types.h>
  62#include <linux/debugfs.h>
  63#include <linux/zsmalloc.h>
  64#include <linux/zpool.h>
  65#include <linux/migrate.h>
  66#include <linux/wait.h>
  67#include <linux/pagemap.h>
  68#include <linux/fs.h>
  69#include <linux/local_lock.h>
  70
  71#define ZSPAGE_MAGIC	0x58
  72
  73/*
  74 * This must be power of 2 and greater than or equal to sizeof(link_free).
  75 * These two conditions ensure that any 'struct link_free' itself doesn't
  76 * span more than 1 page which avoids complex case of mapping 2 pages simply
  77 * to restore link_free pointer values.
  78 */
  79#define ZS_ALIGN		8
  80
  81#define ZS_HANDLE_SIZE (sizeof(unsigned long))
  82
  83/*
  84 * Object location (<PFN>, <obj_idx>) is encoded as
  85 * a single (unsigned long) handle value.
  86 *
  87 * Note that object index <obj_idx> starts from 0.
  88 *
  89 * This is made more complicated by various memory models and PAE.
  90 */
  91
  92#ifndef MAX_POSSIBLE_PHYSMEM_BITS
  93#ifdef MAX_PHYSMEM_BITS
  94#define MAX_POSSIBLE_PHYSMEM_BITS MAX_PHYSMEM_BITS
  95#else
  96/*
  97 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
  98 * be PAGE_SHIFT
  99 */
 100#define MAX_POSSIBLE_PHYSMEM_BITS BITS_PER_LONG
 101#endif
 102#endif
 103
 104#define _PFN_BITS		(MAX_POSSIBLE_PHYSMEM_BITS - PAGE_SHIFT)
 105
 106/*
 107 * Head in allocated object should have OBJ_ALLOCATED_TAG
 108 * to identify the object was allocated or not.
 109 * It's okay to add the status bit in the least bit because
 110 * header keeps handle which is 4byte-aligned address so we
 111 * have room for two bit at least.
 112 */
 113#define OBJ_ALLOCATED_TAG 1
 114
 115#define OBJ_TAG_BITS	1
 116#define OBJ_TAG_MASK	OBJ_ALLOCATED_TAG
 117
 118#define OBJ_INDEX_BITS	(BITS_PER_LONG - _PFN_BITS)
 119#define OBJ_INDEX_MASK	((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
 120
 121#define HUGE_BITS	1
 122#define FULLNESS_BITS	4
 123#define CLASS_BITS	8
 124#define MAGIC_VAL_BITS	8
 125
 126#define ZS_MAX_PAGES_PER_ZSPAGE	(_AC(CONFIG_ZSMALLOC_CHAIN_SIZE, UL))
 127
 128/* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
 129#define ZS_MIN_ALLOC_SIZE \
 130	MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
 131/* each chunk includes extra space to keep handle */
 132#define ZS_MAX_ALLOC_SIZE	PAGE_SIZE
 133
 134/*
 135 * On systems with 4K page size, this gives 255 size classes! There is a
 136 * trader-off here:
 137 *  - Large number of size classes is potentially wasteful as free page are
 138 *    spread across these classes
 139 *  - Small number of size classes causes large internal fragmentation
 140 *  - Probably its better to use specific size classes (empirically
 141 *    determined). NOTE: all those class sizes must be set as multiple of
 142 *    ZS_ALIGN to make sure link_free itself never has to span 2 pages.
 143 *
 144 *  ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
 145 *  (reason above)
 146 */
 147#define ZS_SIZE_CLASS_DELTA	(PAGE_SIZE >> CLASS_BITS)
 148#define ZS_SIZE_CLASSES	(DIV_ROUND_UP(ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE, \
 149				      ZS_SIZE_CLASS_DELTA) + 1)
 150
 151/*
 152 * Pages are distinguished by the ratio of used memory (that is the ratio
 153 * of ->inuse objects to all objects that page can store). For example,
 154 * INUSE_RATIO_10 means that the ratio of used objects is > 0% and <= 10%.
 155 *
 156 * The number of fullness groups is not random. It allows us to keep
 157 * difference between the least busy page in the group (minimum permitted
 158 * number of ->inuse objects) and the most busy page (maximum permitted
 159 * number of ->inuse objects) at a reasonable value.
 160 */
 161enum fullness_group {
 162	ZS_INUSE_RATIO_0,
 163	ZS_INUSE_RATIO_10,
 164	/* NOTE: 8 more fullness groups here */
 165	ZS_INUSE_RATIO_99       = 10,
 166	ZS_INUSE_RATIO_100,
 167	NR_FULLNESS_GROUPS,
 168};
 169
 170enum class_stat_type {
 171	/* NOTE: stats for 12 fullness groups here: from inuse 0 to 100 */
 172	ZS_OBJS_ALLOCATED       = NR_FULLNESS_GROUPS,
 173	ZS_OBJS_INUSE,
 174	NR_CLASS_STAT_TYPES,
 175};
 176
 177struct zs_size_stat {
 178	unsigned long objs[NR_CLASS_STAT_TYPES];
 179};
 180
 181#ifdef CONFIG_ZSMALLOC_STAT
 182static struct dentry *zs_stat_root;
 183#endif
 184
 185static size_t huge_class_size;
 186
 187struct size_class {
 188	spinlock_t lock;
 189	struct list_head fullness_list[NR_FULLNESS_GROUPS];
 190	/*
 191	 * Size of objects stored in this class. Must be multiple
 192	 * of ZS_ALIGN.
 193	 */
 194	int size;
 195	int objs_per_zspage;
 196	/* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
 197	int pages_per_zspage;
 198
 199	unsigned int index;
 200	struct zs_size_stat stats;
 201};
 202
 203/*
 204 * Placed within free objects to form a singly linked list.
 205 * For every zspage, zspage->freeobj gives head of this list.
 206 *
 207 * This must be power of 2 and less than or equal to ZS_ALIGN
 208 */
 209struct link_free {
 210	union {
 211		/*
 212		 * Free object index;
 213		 * It's valid for non-allocated object
 214		 */
 215		unsigned long next;
 216		/*
 217		 * Handle of allocated object.
 218		 */
 219		unsigned long handle;
 220	};
 221};
 222
 223struct zs_pool {
 224	const char *name;
 225
 226	struct size_class *size_class[ZS_SIZE_CLASSES];
 227	struct kmem_cache *handle_cachep;
 228	struct kmem_cache *zspage_cachep;
 229
 230	atomic_long_t pages_allocated;
 231
 232	struct zs_pool_stats stats;
 233
 234	/* Compact classes */
 235	struct shrinker *shrinker;
 236
 237#ifdef CONFIG_ZSMALLOC_STAT
 238	struct dentry *stat_dentry;
 239#endif
 240#ifdef CONFIG_COMPACTION
 241	struct work_struct free_work;
 242#endif
 243	/* protect page/zspage migration */
 244	rwlock_t migrate_lock;
 245	atomic_t compaction_in_progress;
 246};
 247
 248struct zspage {
 249	struct {
 250		unsigned int huge:HUGE_BITS;
 251		unsigned int fullness:FULLNESS_BITS;
 252		unsigned int class:CLASS_BITS + 1;
 253		unsigned int magic:MAGIC_VAL_BITS;
 254	};
 255	unsigned int inuse;
 256	unsigned int freeobj;
 257	struct page *first_page;
 258	struct list_head list; /* fullness list */
 259	struct zs_pool *pool;
 260	rwlock_t lock;
 261};
 262
 263struct mapping_area {
 264	local_lock_t lock;
 265	char *vm_buf; /* copy buffer for objects that span pages */
 266	char *vm_addr; /* address of kmap_local_page()'ed pages */
 267	enum zs_mapmode vm_mm; /* mapping mode */
 268};
 269
 270/* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
 271static void SetZsHugePage(struct zspage *zspage)
 272{
 273	zspage->huge = 1;
 274}
 275
 276static bool ZsHugePage(struct zspage *zspage)
 277{
 278	return zspage->huge;
 279}
 280
 281static void migrate_lock_init(struct zspage *zspage);
 282static void migrate_read_lock(struct zspage *zspage);
 283static void migrate_read_unlock(struct zspage *zspage);
 284static void migrate_write_lock(struct zspage *zspage);
 285static void migrate_write_unlock(struct zspage *zspage);
 286
 287#ifdef CONFIG_COMPACTION
 288static void kick_deferred_free(struct zs_pool *pool);
 289static void init_deferred_free(struct zs_pool *pool);
 290static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage);
 291#else
 292static void kick_deferred_free(struct zs_pool *pool) {}
 293static void init_deferred_free(struct zs_pool *pool) {}
 294static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage) {}
 295#endif
 296
 297static int create_cache(struct zs_pool *pool)
 298{
 299	char *name;
 300
 301	name = kasprintf(GFP_KERNEL, "zs_handle-%s", pool->name);
 302	if (!name)
 303		return -ENOMEM;
 304	pool->handle_cachep = kmem_cache_create(name, ZS_HANDLE_SIZE,
 305						0, 0, NULL);
 306	kfree(name);
 307	if (!pool->handle_cachep)
 308		return -EINVAL;
 309
 310	name = kasprintf(GFP_KERNEL, "zspage-%s", pool->name);
 311	if (!name)
 312		return -ENOMEM;
 313	pool->zspage_cachep = kmem_cache_create(name, sizeof(struct zspage),
 314						0, 0, NULL);
 315	kfree(name);
 316	if (!pool->zspage_cachep) {
 317		kmem_cache_destroy(pool->handle_cachep);
 318		pool->handle_cachep = NULL;
 319		return -EINVAL;
 320	}
 321
 322	return 0;
 323}
 324
 325static void destroy_cache(struct zs_pool *pool)
 326{
 327	kmem_cache_destroy(pool->handle_cachep);
 328	kmem_cache_destroy(pool->zspage_cachep);
 329}
 330
 331static unsigned long cache_alloc_handle(struct zs_pool *pool, gfp_t gfp)
 332{
 333	return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
 334			gfp & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
 335}
 336
 337static void cache_free_handle(struct zs_pool *pool, unsigned long handle)
 338{
 339	kmem_cache_free(pool->handle_cachep, (void *)handle);
 340}
 341
 342static struct zspage *cache_alloc_zspage(struct zs_pool *pool, gfp_t flags)
 343{
 344	return kmem_cache_zalloc(pool->zspage_cachep,
 345			flags & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
 346}
 347
 348static void cache_free_zspage(struct zs_pool *pool, struct zspage *zspage)
 349{
 350	kmem_cache_free(pool->zspage_cachep, zspage);
 351}
 352
 353/* class->lock(which owns the handle) synchronizes races */
 354static void record_obj(unsigned long handle, unsigned long obj)
 355{
 356	*(unsigned long *)handle = obj;
 357}
 358
 359/* zpool driver */
 360
 361#ifdef CONFIG_ZPOOL
 362
 363static void *zs_zpool_create(const char *name, gfp_t gfp)
 364{
 365	/*
 366	 * Ignore global gfp flags: zs_malloc() may be invoked from
 367	 * different contexts and its caller must provide a valid
 368	 * gfp mask.
 369	 */
 370	return zs_create_pool(name);
 371}
 372
 373static void zs_zpool_destroy(void *pool)
 374{
 375	zs_destroy_pool(pool);
 376}
 377
 378static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
 379			unsigned long *handle)
 380{
 381	*handle = zs_malloc(pool, size, gfp);
 382
 383	if (IS_ERR_VALUE(*handle))
 384		return PTR_ERR((void *)*handle);
 385	return 0;
 386}
 387static void zs_zpool_free(void *pool, unsigned long handle)
 388{
 389	zs_free(pool, handle);
 390}
 391
 392static void *zs_zpool_map(void *pool, unsigned long handle,
 393			enum zpool_mapmode mm)
 394{
 395	enum zs_mapmode zs_mm;
 396
 397	switch (mm) {
 398	case ZPOOL_MM_RO:
 399		zs_mm = ZS_MM_RO;
 400		break;
 401	case ZPOOL_MM_WO:
 402		zs_mm = ZS_MM_WO;
 403		break;
 404	case ZPOOL_MM_RW:
 405	default:
 406		zs_mm = ZS_MM_RW;
 407		break;
 408	}
 409
 410	return zs_map_object(pool, handle, zs_mm);
 411}
 412static void zs_zpool_unmap(void *pool, unsigned long handle)
 413{
 414	zs_unmap_object(pool, handle);
 415}
 416
 417static u64 zs_zpool_total_pages(void *pool)
 418{
 419	return zs_get_total_pages(pool);
 420}
 421
 422static struct zpool_driver zs_zpool_driver = {
 423	.type =			  "zsmalloc",
 424	.owner =		  THIS_MODULE,
 425	.create =		  zs_zpool_create,
 426	.destroy =		  zs_zpool_destroy,
 427	.malloc_support_movable = true,
 428	.malloc =		  zs_zpool_malloc,
 429	.free =			  zs_zpool_free,
 430	.map =			  zs_zpool_map,
 431	.unmap =		  zs_zpool_unmap,
 432	.total_pages =		  zs_zpool_total_pages,
 433};
 434
 435MODULE_ALIAS("zpool-zsmalloc");
 436#endif /* CONFIG_ZPOOL */
 437
 438/* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
 439static DEFINE_PER_CPU(struct mapping_area, zs_map_area) = {
 440	.lock	= INIT_LOCAL_LOCK(lock),
 441};
 442
 443static __maybe_unused int is_first_page(struct page *page)
 444{
 445	return PagePrivate(page);
 446}
 447
 448/* Protected by class->lock */
 449static inline int get_zspage_inuse(struct zspage *zspage)
 450{
 451	return zspage->inuse;
 452}
 453
 454
 455static inline void mod_zspage_inuse(struct zspage *zspage, int val)
 456{
 457	zspage->inuse += val;
 458}
 459
 460static inline struct page *get_first_page(struct zspage *zspage)
 461{
 462	struct page *first_page = zspage->first_page;
 463
 464	VM_BUG_ON_PAGE(!is_first_page(first_page), first_page);
 465	return first_page;
 466}
 467
 468#define FIRST_OBJ_PAGE_TYPE_MASK	0xffffff
 469
 470static inline unsigned int get_first_obj_offset(struct page *page)
 471{
 472	VM_WARN_ON_ONCE(!PageZsmalloc(page));
 473	return page->page_type & FIRST_OBJ_PAGE_TYPE_MASK;
 474}
 475
 476static inline void set_first_obj_offset(struct page *page, unsigned int offset)
 477{
 478	/* With 24 bits available, we can support offsets into 16 MiB pages. */
 479	BUILD_BUG_ON(PAGE_SIZE > SZ_16M);
 480	VM_WARN_ON_ONCE(!PageZsmalloc(page));
 481	VM_WARN_ON_ONCE(offset & ~FIRST_OBJ_PAGE_TYPE_MASK);
 482	page->page_type &= ~FIRST_OBJ_PAGE_TYPE_MASK;
 483	page->page_type |= offset & FIRST_OBJ_PAGE_TYPE_MASK;
 484}
 485
 486static inline unsigned int get_freeobj(struct zspage *zspage)
 487{
 488	return zspage->freeobj;
 489}
 490
 491static inline void set_freeobj(struct zspage *zspage, unsigned int obj)
 492{
 493	zspage->freeobj = obj;
 494}
 495
 496static struct size_class *zspage_class(struct zs_pool *pool,
 497				       struct zspage *zspage)
 498{
 499	return pool->size_class[zspage->class];
 500}
 501
 502/*
 503 * zsmalloc divides the pool into various size classes where each
 504 * class maintains a list of zspages where each zspage is divided
 505 * into equal sized chunks. Each allocation falls into one of these
 506 * classes depending on its size. This function returns index of the
 507 * size class which has chunk size big enough to hold the given size.
 508 */
 509static int get_size_class_index(int size)
 510{
 511	int idx = 0;
 512
 513	if (likely(size > ZS_MIN_ALLOC_SIZE))
 514		idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
 515				ZS_SIZE_CLASS_DELTA);
 516
 517	return min_t(int, ZS_SIZE_CLASSES - 1, idx);
 518}
 519
 520static inline void class_stat_add(struct size_class *class, int type,
 521				  unsigned long cnt)
 522{
 523	class->stats.objs[type] += cnt;
 524}
 525
 526static inline void class_stat_sub(struct size_class *class, int type,
 527				  unsigned long cnt)
 528{
 529	class->stats.objs[type] -= cnt;
 530}
 531
 532static inline unsigned long class_stat_read(struct size_class *class, int type)
 533{
 534	return class->stats.objs[type];
 535}
 536
 537#ifdef CONFIG_ZSMALLOC_STAT
 538
 539static void __init zs_stat_init(void)
 540{
 541	if (!debugfs_initialized()) {
 542		pr_warn("debugfs not available, stat dir not created\n");
 543		return;
 544	}
 545
 546	zs_stat_root = debugfs_create_dir("zsmalloc", NULL);
 547}
 548
 549static void __exit zs_stat_exit(void)
 550{
 551	debugfs_remove_recursive(zs_stat_root);
 552}
 553
 554static unsigned long zs_can_compact(struct size_class *class);
 555
 556static int zs_stats_size_show(struct seq_file *s, void *v)
 557{
 558	int i, fg;
 559	struct zs_pool *pool = s->private;
 560	struct size_class *class;
 561	int objs_per_zspage;
 562	unsigned long obj_allocated, obj_used, pages_used, freeable;
 563	unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0;
 564	unsigned long total_freeable = 0;
 565	unsigned long inuse_totals[NR_FULLNESS_GROUPS] = {0, };
 566
 567	seq_printf(s, " %5s %5s %9s %9s %9s %9s %9s %9s %9s %9s %9s %9s %9s %13s %10s %10s %16s %8s\n",
 568			"class", "size", "10%", "20%", "30%", "40%",
 569			"50%", "60%", "70%", "80%", "90%", "99%", "100%",
 570			"obj_allocated", "obj_used", "pages_used",
 571			"pages_per_zspage", "freeable");
 572
 573	for (i = 0; i < ZS_SIZE_CLASSES; i++) {
 574
 575		class = pool->size_class[i];
 576
 577		if (class->index != i)
 578			continue;
 579
 580		spin_lock(&class->lock);
 581
 582		seq_printf(s, " %5u %5u ", i, class->size);
 583		for (fg = ZS_INUSE_RATIO_10; fg < NR_FULLNESS_GROUPS; fg++) {
 584			inuse_totals[fg] += class_stat_read(class, fg);
 585			seq_printf(s, "%9lu ", class_stat_read(class, fg));
 586		}
 587
 588		obj_allocated = class_stat_read(class, ZS_OBJS_ALLOCATED);
 589		obj_used = class_stat_read(class, ZS_OBJS_INUSE);
 590		freeable = zs_can_compact(class);
 591		spin_unlock(&class->lock);
 592
 593		objs_per_zspage = class->objs_per_zspage;
 594		pages_used = obj_allocated / objs_per_zspage *
 595				class->pages_per_zspage;
 596
 597		seq_printf(s, "%13lu %10lu %10lu %16d %8lu\n",
 598			   obj_allocated, obj_used, pages_used,
 599			   class->pages_per_zspage, freeable);
 600
 601		total_objs += obj_allocated;
 602		total_used_objs += obj_used;
 603		total_pages += pages_used;
 604		total_freeable += freeable;
 605	}
 606
 607	seq_puts(s, "\n");
 608	seq_printf(s, " %5s %5s ", "Total", "");
 609
 610	for (fg = ZS_INUSE_RATIO_10; fg < NR_FULLNESS_GROUPS; fg++)
 611		seq_printf(s, "%9lu ", inuse_totals[fg]);
 612
 613	seq_printf(s, "%13lu %10lu %10lu %16s %8lu\n",
 614		   total_objs, total_used_objs, total_pages, "",
 615		   total_freeable);
 616
 617	return 0;
 618}
 619DEFINE_SHOW_ATTRIBUTE(zs_stats_size);
 620
 621static void zs_pool_stat_create(struct zs_pool *pool, const char *name)
 622{
 623	if (!zs_stat_root) {
 624		pr_warn("no root stat dir, not creating <%s> stat dir\n", name);
 625		return;
 626	}
 627
 628	pool->stat_dentry = debugfs_create_dir(name, zs_stat_root);
 629
 630	debugfs_create_file("classes", S_IFREG | 0444, pool->stat_dentry, pool,
 631			    &zs_stats_size_fops);
 632}
 633
 634static void zs_pool_stat_destroy(struct zs_pool *pool)
 635{
 636	debugfs_remove_recursive(pool->stat_dentry);
 637}
 638
 639#else /* CONFIG_ZSMALLOC_STAT */
 640static void __init zs_stat_init(void)
 641{
 642}
 643
 644static void __exit zs_stat_exit(void)
 645{
 646}
 647
 648static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name)
 649{
 650}
 651
 652static inline void zs_pool_stat_destroy(struct zs_pool *pool)
 653{
 654}
 655#endif
 656
 657
 658/*
 659 * For each size class, zspages are divided into different groups
 660 * depending on their usage ratio. This function returns fullness
 661 * status of the given page.
 662 */
 663static int get_fullness_group(struct size_class *class, struct zspage *zspage)
 664{
 665	int inuse, objs_per_zspage, ratio;
 666
 667	inuse = get_zspage_inuse(zspage);
 668	objs_per_zspage = class->objs_per_zspage;
 669
 670	if (inuse == 0)
 671		return ZS_INUSE_RATIO_0;
 672	if (inuse == objs_per_zspage)
 673		return ZS_INUSE_RATIO_100;
 674
 675	ratio = 100 * inuse / objs_per_zspage;
 676	/*
 677	 * Take integer division into consideration: a page with one inuse
 678	 * object out of 127 possible, will end up having 0 usage ratio,
 679	 * which is wrong as it belongs in ZS_INUSE_RATIO_10 fullness group.
 680	 */
 681	return ratio / 10 + 1;
 682}
 683
 684/*
 685 * Each size class maintains various freelists and zspages are assigned
 686 * to one of these freelists based on the number of live objects they
 687 * have. This functions inserts the given zspage into the freelist
 688 * identified by <class, fullness_group>.
 689 */
 690static void insert_zspage(struct size_class *class,
 691				struct zspage *zspage,
 692				int fullness)
 693{
 694	class_stat_add(class, fullness, 1);
 695	list_add(&zspage->list, &class->fullness_list[fullness]);
 696	zspage->fullness = fullness;
 697}
 698
 699/*
 700 * This function removes the given zspage from the freelist identified
 701 * by <class, fullness_group>.
 702 */
 703static void remove_zspage(struct size_class *class, struct zspage *zspage)
 704{
 705	int fullness = zspage->fullness;
 706
 707	VM_BUG_ON(list_empty(&class->fullness_list[fullness]));
 708
 709	list_del_init(&zspage->list);
 710	class_stat_sub(class, fullness, 1);
 711}
 712
 713/*
 714 * Each size class maintains zspages in different fullness groups depending
 715 * on the number of live objects they contain. When allocating or freeing
 716 * objects, the fullness status of the page can change, for instance, from
 717 * INUSE_RATIO_80 to INUSE_RATIO_70 when freeing an object. This function
 718 * checks if such a status change has occurred for the given page and
 719 * accordingly moves the page from the list of the old fullness group to that
 720 * of the new fullness group.
 721 */
 722static int fix_fullness_group(struct size_class *class, struct zspage *zspage)
 723{
 724	int newfg;
 725
 726	newfg = get_fullness_group(class, zspage);
 727	if (newfg == zspage->fullness)
 728		goto out;
 729
 730	remove_zspage(class, zspage);
 731	insert_zspage(class, zspage, newfg);
 732out:
 733	return newfg;
 734}
 735
 736static struct zspage *get_zspage(struct page *page)
 737{
 738	struct zspage *zspage = (struct zspage *)page_private(page);
 739
 740	BUG_ON(zspage->magic != ZSPAGE_MAGIC);
 741	return zspage;
 742}
 743
 744static struct page *get_next_page(struct page *page)
 745{
 746	struct zspage *zspage = get_zspage(page);
 747
 748	if (unlikely(ZsHugePage(zspage)))
 749		return NULL;
 750
 751	return (struct page *)page->index;
 752}
 753
 754/**
 755 * obj_to_location - get (<page>, <obj_idx>) from encoded object value
 756 * @obj: the encoded object value
 757 * @page: page object resides in zspage
 758 * @obj_idx: object index
 759 */
 760static void obj_to_location(unsigned long obj, struct page **page,
 761				unsigned int *obj_idx)
 762{
 763	*page = pfn_to_page(obj >> OBJ_INDEX_BITS);
 764	*obj_idx = (obj & OBJ_INDEX_MASK);
 765}
 766
 767static void obj_to_page(unsigned long obj, struct page **page)
 768{
 769	*page = pfn_to_page(obj >> OBJ_INDEX_BITS);
 770}
 771
 772/**
 773 * location_to_obj - get obj value encoded from (<page>, <obj_idx>)
 774 * @page: page object resides in zspage
 775 * @obj_idx: object index
 776 */
 777static unsigned long location_to_obj(struct page *page, unsigned int obj_idx)
 778{
 779	unsigned long obj;
 780
 781	obj = page_to_pfn(page) << OBJ_INDEX_BITS;
 782	obj |= obj_idx & OBJ_INDEX_MASK;
 783
 784	return obj;
 785}
 786
 787static unsigned long handle_to_obj(unsigned long handle)
 788{
 789	return *(unsigned long *)handle;
 790}
 791
 792static inline bool obj_allocated(struct page *page, void *obj,
 793				 unsigned long *phandle)
 794{
 795	unsigned long handle;
 796	struct zspage *zspage = get_zspage(page);
 797
 798	if (unlikely(ZsHugePage(zspage))) {
 799		VM_BUG_ON_PAGE(!is_first_page(page), page);
 800		handle = page->index;
 801	} else
 802		handle = *(unsigned long *)obj;
 803
 804	if (!(handle & OBJ_ALLOCATED_TAG))
 805		return false;
 806
 807	/* Clear all tags before returning the handle */
 808	*phandle = handle & ~OBJ_TAG_MASK;
 809	return true;
 810}
 811
 812static void reset_page(struct page *page)
 813{
 814	__ClearPageMovable(page);
 815	ClearPagePrivate(page);
 816	set_page_private(page, 0);
 817	page->index = 0;
 818	__ClearPageZsmalloc(page);
 819}
 820
 821static int trylock_zspage(struct zspage *zspage)
 822{
 823	struct page *cursor, *fail;
 824
 825	for (cursor = get_first_page(zspage); cursor != NULL; cursor =
 826					get_next_page(cursor)) {
 827		if (!trylock_page(cursor)) {
 828			fail = cursor;
 829			goto unlock;
 830		}
 831	}
 832
 833	return 1;
 834unlock:
 835	for (cursor = get_first_page(zspage); cursor != fail; cursor =
 836					get_next_page(cursor))
 837		unlock_page(cursor);
 838
 839	return 0;
 840}
 841
 842static void __free_zspage(struct zs_pool *pool, struct size_class *class,
 843				struct zspage *zspage)
 844{
 845	struct page *page, *next;
 846
 847	assert_spin_locked(&class->lock);
 848
 849	VM_BUG_ON(get_zspage_inuse(zspage));
 850	VM_BUG_ON(zspage->fullness != ZS_INUSE_RATIO_0);
 851
 852	next = page = get_first_page(zspage);
 853	do {
 854		VM_BUG_ON_PAGE(!PageLocked(page), page);
 855		next = get_next_page(page);
 856		reset_page(page);
 857		unlock_page(page);
 858		dec_zone_page_state(page, NR_ZSPAGES);
 859		put_page(page);
 860		page = next;
 861	} while (page != NULL);
 862
 863	cache_free_zspage(pool, zspage);
 864
 865	class_stat_sub(class, ZS_OBJS_ALLOCATED, class->objs_per_zspage);
 866	atomic_long_sub(class->pages_per_zspage, &pool->pages_allocated);
 867}
 868
 869static void free_zspage(struct zs_pool *pool, struct size_class *class,
 870				struct zspage *zspage)
 871{
 872	VM_BUG_ON(get_zspage_inuse(zspage));
 873	VM_BUG_ON(list_empty(&zspage->list));
 874
 875	/*
 876	 * Since zs_free couldn't be sleepable, this function cannot call
 877	 * lock_page. The page locks trylock_zspage got will be released
 878	 * by __free_zspage.
 879	 */
 880	if (!trylock_zspage(zspage)) {
 881		kick_deferred_free(pool);
 882		return;
 883	}
 884
 885	remove_zspage(class, zspage);
 886	__free_zspage(pool, class, zspage);
 887}
 888
 889/* Initialize a newly allocated zspage */
 890static void init_zspage(struct size_class *class, struct zspage *zspage)
 891{
 892	unsigned int freeobj = 1;
 893	unsigned long off = 0;
 894	struct page *page = get_first_page(zspage);
 895
 896	while (page) {
 897		struct page *next_page;
 898		struct link_free *link;
 899		void *vaddr;
 900
 901		set_first_obj_offset(page, off);
 902
 903		vaddr = kmap_local_page(page);
 904		link = (struct link_free *)vaddr + off / sizeof(*link);
 905
 906		while ((off += class->size) < PAGE_SIZE) {
 907			link->next = freeobj++ << OBJ_TAG_BITS;
 908			link += class->size / sizeof(*link);
 909		}
 910
 911		/*
 912		 * We now come to the last (full or partial) object on this
 913		 * page, which must point to the first object on the next
 914		 * page (if present)
 915		 */
 916		next_page = get_next_page(page);
 917		if (next_page) {
 918			link->next = freeobj++ << OBJ_TAG_BITS;
 919		} else {
 920			/*
 921			 * Reset OBJ_TAG_BITS bit to last link to tell
 922			 * whether it's allocated object or not.
 923			 */
 924			link->next = -1UL << OBJ_TAG_BITS;
 925		}
 926		kunmap_local(vaddr);
 927		page = next_page;
 928		off %= PAGE_SIZE;
 929	}
 930
 931	set_freeobj(zspage, 0);
 932}
 933
 934static void create_page_chain(struct size_class *class, struct zspage *zspage,
 935				struct page *pages[])
 936{
 937	int i;
 938	struct page *page;
 939	struct page *prev_page = NULL;
 940	int nr_pages = class->pages_per_zspage;
 941
 942	/*
 943	 * Allocate individual pages and link them together as:
 944	 * 1. all pages are linked together using page->index
 945	 * 2. each sub-page point to zspage using page->private
 946	 *
 947	 * we set PG_private to identify the first page (i.e. no other sub-page
 948	 * has this flag set).
 949	 */
 950	for (i = 0; i < nr_pages; i++) {
 951		page = pages[i];
 952		set_page_private(page, (unsigned long)zspage);
 953		page->index = 0;
 954		if (i == 0) {
 955			zspage->first_page = page;
 956			SetPagePrivate(page);
 957			if (unlikely(class->objs_per_zspage == 1 &&
 958					class->pages_per_zspage == 1))
 959				SetZsHugePage(zspage);
 960		} else {
 961			prev_page->index = (unsigned long)page;
 962		}
 963		prev_page = page;
 964	}
 965}
 966
 967/*
 968 * Allocate a zspage for the given size class
 969 */
 970static struct zspage *alloc_zspage(struct zs_pool *pool,
 971					struct size_class *class,
 972					gfp_t gfp)
 973{
 974	int i;
 975	struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE];
 976	struct zspage *zspage = cache_alloc_zspage(pool, gfp);
 977
 978	if (!zspage)
 979		return NULL;
 980
 981	zspage->magic = ZSPAGE_MAGIC;
 982	migrate_lock_init(zspage);
 983
 984	for (i = 0; i < class->pages_per_zspage; i++) {
 985		struct page *page;
 986
 987		page = alloc_page(gfp);
 988		if (!page) {
 989			while (--i >= 0) {
 990				dec_zone_page_state(pages[i], NR_ZSPAGES);
 991				__ClearPageZsmalloc(pages[i]);
 992				__free_page(pages[i]);
 993			}
 994			cache_free_zspage(pool, zspage);
 995			return NULL;
 996		}
 997		__SetPageZsmalloc(page);
 998
 999		inc_zone_page_state(page, NR_ZSPAGES);
1000		pages[i] = page;
1001	}
1002
1003	create_page_chain(class, zspage, pages);
1004	init_zspage(class, zspage);
1005	zspage->pool = pool;
1006	zspage->class = class->index;
1007
1008	return zspage;
1009}
1010
1011static struct zspage *find_get_zspage(struct size_class *class)
1012{
1013	int i;
1014	struct zspage *zspage;
1015
1016	for (i = ZS_INUSE_RATIO_99; i >= ZS_INUSE_RATIO_0; i--) {
1017		zspage = list_first_entry_or_null(&class->fullness_list[i],
1018						  struct zspage, list);
1019		if (zspage)
1020			break;
1021	}
1022
1023	return zspage;
1024}
1025
1026static inline int __zs_cpu_up(struct mapping_area *area)
1027{
1028	/*
1029	 * Make sure we don't leak memory if a cpu UP notification
1030	 * and zs_init() race and both call zs_cpu_up() on the same cpu
1031	 */
1032	if (area->vm_buf)
1033		return 0;
1034	area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
1035	if (!area->vm_buf)
1036		return -ENOMEM;
1037	return 0;
1038}
1039
1040static inline void __zs_cpu_down(struct mapping_area *area)
1041{
1042	kfree(area->vm_buf);
1043	area->vm_buf = NULL;
1044}
1045
1046static void *__zs_map_object(struct mapping_area *area,
1047			struct page *pages[2], int off, int size)
1048{
1049	size_t sizes[2];
1050	char *buf = area->vm_buf;
1051
1052	/* disable page faults to match kmap_local_page() return conditions */
1053	pagefault_disable();
1054
1055	/* no read fastpath */
1056	if (area->vm_mm == ZS_MM_WO)
1057		goto out;
1058
1059	sizes[0] = PAGE_SIZE - off;
1060	sizes[1] = size - sizes[0];
1061
1062	/* copy object to per-cpu buffer */
1063	memcpy_from_page(buf, pages[0], off, sizes[0]);
1064	memcpy_from_page(buf + sizes[0], pages[1], 0, sizes[1]);
1065out:
1066	return area->vm_buf;
1067}
1068
1069static void __zs_unmap_object(struct mapping_area *area,
1070			struct page *pages[2], int off, int size)
1071{
1072	size_t sizes[2];
1073	char *buf;
1074
1075	/* no write fastpath */
1076	if (area->vm_mm == ZS_MM_RO)
1077		goto out;
1078
1079	buf = area->vm_buf;
1080	buf = buf + ZS_HANDLE_SIZE;
1081	size -= ZS_HANDLE_SIZE;
1082	off += ZS_HANDLE_SIZE;
1083
1084	sizes[0] = PAGE_SIZE - off;
1085	sizes[1] = size - sizes[0];
1086
1087	/* copy per-cpu buffer to object */
1088	memcpy_to_page(pages[0], off, buf, sizes[0]);
1089	memcpy_to_page(pages[1], 0, buf + sizes[0], sizes[1]);
1090
1091out:
1092	/* enable page faults to match kunmap_local() return conditions */
1093	pagefault_enable();
1094}
1095
1096static int zs_cpu_prepare(unsigned int cpu)
1097{
1098	struct mapping_area *area;
1099
1100	area = &per_cpu(zs_map_area, cpu);
1101	return __zs_cpu_up(area);
1102}
1103
1104static int zs_cpu_dead(unsigned int cpu)
1105{
1106	struct mapping_area *area;
1107
1108	area = &per_cpu(zs_map_area, cpu);
1109	__zs_cpu_down(area);
1110	return 0;
1111}
1112
1113static bool can_merge(struct size_class *prev, int pages_per_zspage,
1114					int objs_per_zspage)
1115{
1116	if (prev->pages_per_zspage == pages_per_zspage &&
1117		prev->objs_per_zspage == objs_per_zspage)
1118		return true;
1119
1120	return false;
1121}
1122
1123static bool zspage_full(struct size_class *class, struct zspage *zspage)
1124{
1125	return get_zspage_inuse(zspage) == class->objs_per_zspage;
1126}
1127
1128static bool zspage_empty(struct zspage *zspage)
1129{
1130	return get_zspage_inuse(zspage) == 0;
1131}
1132
1133/**
1134 * zs_lookup_class_index() - Returns index of the zsmalloc &size_class
1135 * that hold objects of the provided size.
1136 * @pool: zsmalloc pool to use
1137 * @size: object size
1138 *
1139 * Context: Any context.
1140 *
1141 * Return: the index of the zsmalloc &size_class that hold objects of the
1142 * provided size.
1143 */
1144unsigned int zs_lookup_class_index(struct zs_pool *pool, unsigned int size)
1145{
1146	struct size_class *class;
1147
1148	class = pool->size_class[get_size_class_index(size)];
1149
1150	return class->index;
1151}
1152EXPORT_SYMBOL_GPL(zs_lookup_class_index);
1153
1154unsigned long zs_get_total_pages(struct zs_pool *pool)
1155{
1156	return atomic_long_read(&pool->pages_allocated);
1157}
1158EXPORT_SYMBOL_GPL(zs_get_total_pages);
1159
1160/**
1161 * zs_map_object - get address of allocated object from handle.
1162 * @pool: pool from which the object was allocated
1163 * @handle: handle returned from zs_malloc
1164 * @mm: mapping mode to use
1165 *
1166 * Before using an object allocated from zs_malloc, it must be mapped using
1167 * this function. When done with the object, it must be unmapped using
1168 * zs_unmap_object.
1169 *
1170 * Only one object can be mapped per cpu at a time. There is no protection
1171 * against nested mappings.
1172 *
1173 * This function returns with preemption and page faults disabled.
1174 */
1175void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1176			enum zs_mapmode mm)
1177{
1178	struct zspage *zspage;
1179	struct page *page;
1180	unsigned long obj, off;
1181	unsigned int obj_idx;
1182
1183	struct size_class *class;
1184	struct mapping_area *area;
1185	struct page *pages[2];
1186	void *ret;
1187
1188	/*
1189	 * Because we use per-cpu mapping areas shared among the
1190	 * pools/users, we can't allow mapping in interrupt context
1191	 * because it can corrupt another users mappings.
1192	 */
1193	BUG_ON(in_interrupt());
1194
1195	/* It guarantees it can get zspage from handle safely */
1196	read_lock(&pool->migrate_lock);
1197	obj = handle_to_obj(handle);
1198	obj_to_location(obj, &page, &obj_idx);
1199	zspage = get_zspage(page);
1200
1201	/*
1202	 * migration cannot move any zpages in this zspage. Here, class->lock
1203	 * is too heavy since callers would take some time until they calls
1204	 * zs_unmap_object API so delegate the locking from class to zspage
1205	 * which is smaller granularity.
1206	 */
1207	migrate_read_lock(zspage);
1208	read_unlock(&pool->migrate_lock);
1209
1210	class = zspage_class(pool, zspage);
1211	off = offset_in_page(class->size * obj_idx);
1212
1213	local_lock(&zs_map_area.lock);
1214	area = this_cpu_ptr(&zs_map_area);
1215	area->vm_mm = mm;
1216	if (off + class->size <= PAGE_SIZE) {
1217		/* this object is contained entirely within a page */
1218		area->vm_addr = kmap_local_page(page);
1219		ret = area->vm_addr + off;
1220		goto out;
1221	}
1222
1223	/* this object spans two pages */
1224	pages[0] = page;
1225	pages[1] = get_next_page(page);
1226	BUG_ON(!pages[1]);
1227
1228	ret = __zs_map_object(area, pages, off, class->size);
1229out:
1230	if (likely(!ZsHugePage(zspage)))
1231		ret += ZS_HANDLE_SIZE;
1232
1233	return ret;
1234}
1235EXPORT_SYMBOL_GPL(zs_map_object);
1236
1237void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1238{
1239	struct zspage *zspage;
1240	struct page *page;
1241	unsigned long obj, off;
1242	unsigned int obj_idx;
1243
1244	struct size_class *class;
1245	struct mapping_area *area;
1246
1247	obj = handle_to_obj(handle);
1248	obj_to_location(obj, &page, &obj_idx);
1249	zspage = get_zspage(page);
1250	class = zspage_class(pool, zspage);
1251	off = offset_in_page(class->size * obj_idx);
1252
1253	area = this_cpu_ptr(&zs_map_area);
1254	if (off + class->size <= PAGE_SIZE)
1255		kunmap_local(area->vm_addr);
1256	else {
1257		struct page *pages[2];
1258
1259		pages[0] = page;
1260		pages[1] = get_next_page(page);
1261		BUG_ON(!pages[1]);
1262
1263		__zs_unmap_object(area, pages, off, class->size);
1264	}
1265	local_unlock(&zs_map_area.lock);
1266
1267	migrate_read_unlock(zspage);
1268}
1269EXPORT_SYMBOL_GPL(zs_unmap_object);
1270
1271/**
1272 * zs_huge_class_size() - Returns the size (in bytes) of the first huge
1273 *                        zsmalloc &size_class.
1274 * @pool: zsmalloc pool to use
1275 *
1276 * The function returns the size of the first huge class - any object of equal
1277 * or bigger size will be stored in zspage consisting of a single physical
1278 * page.
1279 *
1280 * Context: Any context.
1281 *
1282 * Return: the size (in bytes) of the first huge zsmalloc &size_class.
1283 */
1284size_t zs_huge_class_size(struct zs_pool *pool)
1285{
1286	return huge_class_size;
1287}
1288EXPORT_SYMBOL_GPL(zs_huge_class_size);
1289
1290static unsigned long obj_malloc(struct zs_pool *pool,
1291				struct zspage *zspage, unsigned long handle)
1292{
1293	int i, nr_page, offset;
1294	unsigned long obj;
1295	struct link_free *link;
1296	struct size_class *class;
1297
1298	struct page *m_page;
1299	unsigned long m_offset;
1300	void *vaddr;
1301
1302	class = pool->size_class[zspage->class];
1303	obj = get_freeobj(zspage);
1304
1305	offset = obj * class->size;
1306	nr_page = offset >> PAGE_SHIFT;
1307	m_offset = offset_in_page(offset);
1308	m_page = get_first_page(zspage);
1309
1310	for (i = 0; i < nr_page; i++)
1311		m_page = get_next_page(m_page);
1312
1313	vaddr = kmap_local_page(m_page);
1314	link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1315	set_freeobj(zspage, link->next >> OBJ_TAG_BITS);
1316	if (likely(!ZsHugePage(zspage)))
1317		/* record handle in the header of allocated chunk */
1318		link->handle = handle | OBJ_ALLOCATED_TAG;
1319	else
1320		/* record handle to page->index */
1321		zspage->first_page->index = handle | OBJ_ALLOCATED_TAG;
1322
1323	kunmap_local(vaddr);
1324	mod_zspage_inuse(zspage, 1);
1325
1326	obj = location_to_obj(m_page, obj);
1327	record_obj(handle, obj);
1328
1329	return obj;
1330}
1331
1332
1333/**
1334 * zs_malloc - Allocate block of given size from pool.
1335 * @pool: pool to allocate from
1336 * @size: size of block to allocate
1337 * @gfp: gfp flags when allocating object
1338 *
1339 * On success, handle to the allocated object is returned,
1340 * otherwise an ERR_PTR().
1341 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1342 */
1343unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp)
1344{
1345	unsigned long handle;
1346	struct size_class *class;
1347	int newfg;
1348	struct zspage *zspage;
1349
1350	if (unlikely(!size))
1351		return (unsigned long)ERR_PTR(-EINVAL);
1352
1353	if (unlikely(size > ZS_MAX_ALLOC_SIZE))
1354		return (unsigned long)ERR_PTR(-ENOSPC);
1355
1356	handle = cache_alloc_handle(pool, gfp);
1357	if (!handle)
1358		return (unsigned long)ERR_PTR(-ENOMEM);
1359
1360	/* extra space in chunk to keep the handle */
1361	size += ZS_HANDLE_SIZE;
1362	class = pool->size_class[get_size_class_index(size)];
1363
1364	/* class->lock effectively protects the zpage migration */
1365	spin_lock(&class->lock);
1366	zspage = find_get_zspage(class);
1367	if (likely(zspage)) {
1368		obj_malloc(pool, zspage, handle);
1369		/* Now move the zspage to another fullness group, if required */
1370		fix_fullness_group(class, zspage);
1371		class_stat_add(class, ZS_OBJS_INUSE, 1);
1372
1373		goto out;
1374	}
1375
1376	spin_unlock(&class->lock);
1377
1378	zspage = alloc_zspage(pool, class, gfp);
1379	if (!zspage) {
1380		cache_free_handle(pool, handle);
1381		return (unsigned long)ERR_PTR(-ENOMEM);
1382	}
1383
1384	spin_lock(&class->lock);
1385	obj_malloc(pool, zspage, handle);
1386	newfg = get_fullness_group(class, zspage);
1387	insert_zspage(class, zspage, newfg);
1388	atomic_long_add(class->pages_per_zspage, &pool->pages_allocated);
1389	class_stat_add(class, ZS_OBJS_ALLOCATED, class->objs_per_zspage);
1390	class_stat_add(class, ZS_OBJS_INUSE, 1);
1391
1392	/* We completely set up zspage so mark them as movable */
1393	SetZsPageMovable(pool, zspage);
1394out:
1395	spin_unlock(&class->lock);
1396
1397	return handle;
1398}
1399EXPORT_SYMBOL_GPL(zs_malloc);
1400
1401static void obj_free(int class_size, unsigned long obj)
1402{
1403	struct link_free *link;
1404	struct zspage *zspage;
1405	struct page *f_page;
1406	unsigned long f_offset;
1407	unsigned int f_objidx;
1408	void *vaddr;
1409
1410	obj_to_location(obj, &f_page, &f_objidx);
1411	f_offset = offset_in_page(class_size * f_objidx);
1412	zspage = get_zspage(f_page);
1413
1414	vaddr = kmap_local_page(f_page);
1415	link = (struct link_free *)(vaddr + f_offset);
1416
1417	/* Insert this object in containing zspage's freelist */
1418	if (likely(!ZsHugePage(zspage)))
1419		link->next = get_freeobj(zspage) << OBJ_TAG_BITS;
1420	else
1421		f_page->index = 0;
1422	set_freeobj(zspage, f_objidx);
1423
1424	kunmap_local(vaddr);
1425	mod_zspage_inuse(zspage, -1);
1426}
1427
1428void zs_free(struct zs_pool *pool, unsigned long handle)
1429{
1430	struct zspage *zspage;
1431	struct page *f_page;
1432	unsigned long obj;
1433	struct size_class *class;
1434	int fullness;
1435
1436	if (IS_ERR_OR_NULL((void *)handle))
1437		return;
1438
1439	/*
1440	 * The pool->migrate_lock protects the race with zpage's migration
1441	 * so it's safe to get the page from handle.
1442	 */
1443	read_lock(&pool->migrate_lock);
1444	obj = handle_to_obj(handle);
1445	obj_to_page(obj, &f_page);
1446	zspage = get_zspage(f_page);
1447	class = zspage_class(pool, zspage);
1448	spin_lock(&class->lock);
1449	read_unlock(&pool->migrate_lock);
1450
1451	class_stat_sub(class, ZS_OBJS_INUSE, 1);
1452	obj_free(class->size, obj);
1453
1454	fullness = fix_fullness_group(class, zspage);
1455	if (fullness == ZS_INUSE_RATIO_0)
1456		free_zspage(pool, class, zspage);
1457
1458	spin_unlock(&class->lock);
1459	cache_free_handle(pool, handle);
1460}
1461EXPORT_SYMBOL_GPL(zs_free);
1462
1463static void zs_object_copy(struct size_class *class, unsigned long dst,
1464				unsigned long src)
1465{
1466	struct page *s_page, *d_page;
1467	unsigned int s_objidx, d_objidx;
1468	unsigned long s_off, d_off;
1469	void *s_addr, *d_addr;
1470	int s_size, d_size, size;
1471	int written = 0;
1472
1473	s_size = d_size = class->size;
1474
1475	obj_to_location(src, &s_page, &s_objidx);
1476	obj_to_location(dst, &d_page, &d_objidx);
1477
1478	s_off = offset_in_page(class->size * s_objidx);
1479	d_off = offset_in_page(class->size * d_objidx);
1480
1481	if (s_off + class->size > PAGE_SIZE)
1482		s_size = PAGE_SIZE - s_off;
1483
1484	if (d_off + class->size > PAGE_SIZE)
1485		d_size = PAGE_SIZE - d_off;
1486
1487	s_addr = kmap_local_page(s_page);
1488	d_addr = kmap_local_page(d_page);
1489
1490	while (1) {
1491		size = min(s_size, d_size);
1492		memcpy(d_addr + d_off, s_addr + s_off, size);
1493		written += size;
1494
1495		if (written == class->size)
1496			break;
1497
1498		s_off += size;
1499		s_size -= size;
1500		d_off += size;
1501		d_size -= size;
1502
1503		/*
1504		 * Calling kunmap_local(d_addr) is necessary. kunmap_local()
1505		 * calls must occurs in reverse order of calls to kmap_local_page().
1506		 * So, to call kunmap_local(s_addr) we should first call
1507		 * kunmap_local(d_addr). For more details see
1508		 * Documentation/mm/highmem.rst.
1509		 */
1510		if (s_off >= PAGE_SIZE) {
1511			kunmap_local(d_addr);
1512			kunmap_local(s_addr);
1513			s_page = get_next_page(s_page);
1514			s_addr = kmap_local_page(s_page);
1515			d_addr = kmap_local_page(d_page);
1516			s_size = class->size - written;
1517			s_off = 0;
1518		}
1519
1520		if (d_off >= PAGE_SIZE) {
1521			kunmap_local(d_addr);
1522			d_page = get_next_page(d_page);
1523			d_addr = kmap_local_page(d_page);
1524			d_size = class->size - written;
1525			d_off = 0;
1526		}
1527	}
1528
1529	kunmap_local(d_addr);
1530	kunmap_local(s_addr);
1531}
1532
1533/*
1534 * Find alloced object in zspage from index object and
1535 * return handle.
1536 */
1537static unsigned long find_alloced_obj(struct size_class *class,
1538				      struct page *page, int *obj_idx)
1539{
1540	unsigned int offset;
1541	int index = *obj_idx;
1542	unsigned long handle = 0;
1543	void *addr = kmap_local_page(page);
1544
1545	offset = get_first_obj_offset(page);
1546	offset += class->size * index;
1547
1548	while (offset < PAGE_SIZE) {
1549		if (obj_allocated(page, addr + offset, &handle))
1550			break;
1551
1552		offset += class->size;
1553		index++;
1554	}
1555
1556	kunmap_local(addr);
1557
1558	*obj_idx = index;
1559
1560	return handle;
1561}
1562
1563static void migrate_zspage(struct zs_pool *pool, struct zspage *src_zspage,
1564			   struct zspage *dst_zspage)
1565{
1566	unsigned long used_obj, free_obj;
1567	unsigned long handle;
1568	int obj_idx = 0;
1569	struct page *s_page = get_first_page(src_zspage);
1570	struct size_class *class = pool->size_class[src_zspage->class];
1571
1572	while (1) {
1573		handle = find_alloced_obj(class, s_page, &obj_idx);
1574		if (!handle) {
1575			s_page = get_next_page(s_page);
1576			if (!s_page)
1577				break;
1578			obj_idx = 0;
1579			continue;
1580		}
1581
1582		used_obj = handle_to_obj(handle);
1583		free_obj = obj_malloc(pool, dst_zspage, handle);
1584		zs_object_copy(class, free_obj, used_obj);
1585		obj_idx++;
1586		obj_free(class->size, used_obj);
1587
1588		/* Stop if there is no more space */
1589		if (zspage_full(class, dst_zspage))
1590			break;
1591
1592		/* Stop if there are no more objects to migrate */
1593		if (zspage_empty(src_zspage))
1594			break;
1595	}
1596}
1597
1598static struct zspage *isolate_src_zspage(struct size_class *class)
1599{
1600	struct zspage *zspage;
1601	int fg;
1602
1603	for (fg = ZS_INUSE_RATIO_10; fg <= ZS_INUSE_RATIO_99; fg++) {
1604		zspage = list_first_entry_or_null(&class->fullness_list[fg],
1605						  struct zspage, list);
1606		if (zspage) {
1607			remove_zspage(class, zspage);
1608			return zspage;
1609		}
1610	}
1611
1612	return zspage;
1613}
1614
1615static struct zspage *isolate_dst_zspage(struct size_class *class)
1616{
1617	struct zspage *zspage;
1618	int fg;
1619
1620	for (fg = ZS_INUSE_RATIO_99; fg >= ZS_INUSE_RATIO_10; fg--) {
1621		zspage = list_first_entry_or_null(&class->fullness_list[fg],
1622						  struct zspage, list);
1623		if (zspage) {
1624			remove_zspage(class, zspage);
1625			return zspage;
1626		}
1627	}
1628
1629	return zspage;
1630}
1631
1632/*
1633 * putback_zspage - add @zspage into right class's fullness list
1634 * @class: destination class
1635 * @zspage: target page
1636 *
1637 * Return @zspage's fullness status
1638 */
1639static int putback_zspage(struct size_class *class, struct zspage *zspage)
1640{
1641	int fullness;
1642
1643	fullness = get_fullness_group(class, zspage);
1644	insert_zspage(class, zspage, fullness);
1645
1646	return fullness;
1647}
1648
1649#ifdef CONFIG_COMPACTION
1650/*
1651 * To prevent zspage destroy during migration, zspage freeing should
1652 * hold locks of all pages in the zspage.
1653 */
1654static void lock_zspage(struct zspage *zspage)
1655{
1656	struct page *curr_page, *page;
1657
1658	/*
1659	 * Pages we haven't locked yet can be migrated off the list while we're
1660	 * trying to lock them, so we need to be careful and only attempt to
1661	 * lock each page under migrate_read_lock(). Otherwise, the page we lock
1662	 * may no longer belong to the zspage. This means that we may wait for
1663	 * the wrong page to unlock, so we must take a reference to the page
1664	 * prior to waiting for it to unlock outside migrate_read_lock().
1665	 */
1666	while (1) {
1667		migrate_read_lock(zspage);
1668		page = get_first_page(zspage);
1669		if (trylock_page(page))
1670			break;
1671		get_page(page);
1672		migrate_read_unlock(zspage);
1673		wait_on_page_locked(page);
1674		put_page(page);
1675	}
1676
1677	curr_page = page;
1678	while ((page = get_next_page(curr_page))) {
1679		if (trylock_page(page)) {
1680			curr_page = page;
1681		} else {
1682			get_page(page);
1683			migrate_read_unlock(zspage);
1684			wait_on_page_locked(page);
1685			put_page(page);
1686			migrate_read_lock(zspage);
1687		}
1688	}
1689	migrate_read_unlock(zspage);
1690}
1691#endif /* CONFIG_COMPACTION */
1692
1693static void migrate_lock_init(struct zspage *zspage)
1694{
1695	rwlock_init(&zspage->lock);
1696}
1697
1698static void migrate_read_lock(struct zspage *zspage) __acquires(&zspage->lock)
1699{
1700	read_lock(&zspage->lock);
1701}
1702
1703static void migrate_read_unlock(struct zspage *zspage) __releases(&zspage->lock)
1704{
1705	read_unlock(&zspage->lock);
1706}
1707
1708static void migrate_write_lock(struct zspage *zspage)
1709{
1710	write_lock(&zspage->lock);
1711}
1712
1713static void migrate_write_unlock(struct zspage *zspage)
1714{
1715	write_unlock(&zspage->lock);
1716}
1717
1718#ifdef CONFIG_COMPACTION
1719
1720static const struct movable_operations zsmalloc_mops;
1721
1722static void replace_sub_page(struct size_class *class, struct zspage *zspage,
1723				struct page *newpage, struct page *oldpage)
1724{
1725	struct page *page;
1726	struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE] = {NULL, };
1727	int idx = 0;
1728
1729	page = get_first_page(zspage);
1730	do {
1731		if (page == oldpage)
1732			pages[idx] = newpage;
1733		else
1734			pages[idx] = page;
1735		idx++;
1736	} while ((page = get_next_page(page)) != NULL);
1737
1738	create_page_chain(class, zspage, pages);
1739	set_first_obj_offset(newpage, get_first_obj_offset(oldpage));
1740	if (unlikely(ZsHugePage(zspage)))
1741		newpage->index = oldpage->index;
1742	__SetPageMovable(newpage, &zsmalloc_mops);
1743}
1744
1745static bool zs_page_isolate(struct page *page, isolate_mode_t mode)
1746{
1747	/*
1748	 * Page is locked so zspage couldn't be destroyed. For detail, look at
1749	 * lock_zspage in free_zspage.
1750	 */
1751	VM_BUG_ON_PAGE(PageIsolated(page), page);
1752
1753	return true;
1754}
1755
1756static int zs_page_migrate(struct page *newpage, struct page *page,
1757		enum migrate_mode mode)
1758{
1759	struct zs_pool *pool;
1760	struct size_class *class;
1761	struct zspage *zspage;
1762	struct page *dummy;
1763	void *s_addr, *d_addr, *addr;
1764	unsigned int offset;
1765	unsigned long handle;
1766	unsigned long old_obj, new_obj;
1767	unsigned int obj_idx;
1768
1769	VM_BUG_ON_PAGE(!PageIsolated(page), page);
1770
1771	/* We're committed, tell the world that this is a Zsmalloc page. */
1772	__SetPageZsmalloc(newpage);
1773
1774	/* The page is locked, so this pointer must remain valid */
1775	zspage = get_zspage(page);
1776	pool = zspage->pool;
1777
1778	/*
1779	 * The pool migrate_lock protects the race between zpage migration
1780	 * and zs_free.
1781	 */
1782	write_lock(&pool->migrate_lock);
1783	class = zspage_class(pool, zspage);
1784
1785	/*
1786	 * the class lock protects zpage alloc/free in the zspage.
1787	 */
1788	spin_lock(&class->lock);
1789	/* the migrate_write_lock protects zpage access via zs_map_object */
1790	migrate_write_lock(zspage);
1791
1792	offset = get_first_obj_offset(page);
1793	s_addr = kmap_local_page(page);
1794
1795	/*
1796	 * Here, any user cannot access all objects in the zspage so let's move.
1797	 */
1798	d_addr = kmap_local_page(newpage);
1799	copy_page(d_addr, s_addr);
1800	kunmap_local(d_addr);
1801
1802	for (addr = s_addr + offset; addr < s_addr + PAGE_SIZE;
1803					addr += class->size) {
1804		if (obj_allocated(page, addr, &handle)) {
1805
1806			old_obj = handle_to_obj(handle);
1807			obj_to_location(old_obj, &dummy, &obj_idx);
1808			new_obj = (unsigned long)location_to_obj(newpage,
1809								obj_idx);
1810			record_obj(handle, new_obj);
1811		}
1812	}
1813	kunmap_local(s_addr);
1814
1815	replace_sub_page(class, zspage, newpage, page);
1816	/*
1817	 * Since we complete the data copy and set up new zspage structure,
1818	 * it's okay to release migration_lock.
1819	 */
1820	write_unlock(&pool->migrate_lock);
1821	spin_unlock(&class->lock);
1822	migrate_write_unlock(zspage);
1823
1824	get_page(newpage);
1825	if (page_zone(newpage) != page_zone(page)) {
1826		dec_zone_page_state(page, NR_ZSPAGES);
1827		inc_zone_page_state(newpage, NR_ZSPAGES);
1828	}
1829
1830	reset_page(page);
1831	put_page(page);
1832
1833	return MIGRATEPAGE_SUCCESS;
1834}
1835
1836static void zs_page_putback(struct page *page)
1837{
1838	VM_BUG_ON_PAGE(!PageIsolated(page), page);
1839}
1840
1841static const struct movable_operations zsmalloc_mops = {
1842	.isolate_page = zs_page_isolate,
1843	.migrate_page = zs_page_migrate,
1844	.putback_page = zs_page_putback,
1845};
1846
1847/*
1848 * Caller should hold page_lock of all pages in the zspage
1849 * In here, we cannot use zspage meta data.
1850 */
1851static void async_free_zspage(struct work_struct *work)
1852{
1853	int i;
1854	struct size_class *class;
1855	struct zspage *zspage, *tmp;
1856	LIST_HEAD(free_pages);
1857	struct zs_pool *pool = container_of(work, struct zs_pool,
1858					free_work);
1859
1860	for (i = 0; i < ZS_SIZE_CLASSES; i++) {
1861		class = pool->size_class[i];
1862		if (class->index != i)
1863			continue;
1864
1865		spin_lock(&class->lock);
1866		list_splice_init(&class->fullness_list[ZS_INUSE_RATIO_0],
1867				 &free_pages);
1868		spin_unlock(&class->lock);
1869	}
1870
1871	list_for_each_entry_safe(zspage, tmp, &free_pages, list) {
1872		list_del(&zspage->list);
1873		lock_zspage(zspage);
1874
1875		class = zspage_class(pool, zspage);
1876		spin_lock(&class->lock);
1877		class_stat_sub(class, ZS_INUSE_RATIO_0, 1);
1878		__free_zspage(pool, class, zspage);
1879		spin_unlock(&class->lock);
1880	}
1881};
1882
1883static void kick_deferred_free(struct zs_pool *pool)
1884{
1885	schedule_work(&pool->free_work);
1886}
1887
1888static void zs_flush_migration(struct zs_pool *pool)
1889{
1890	flush_work(&pool->free_work);
1891}
1892
1893static void init_deferred_free(struct zs_pool *pool)
1894{
1895	INIT_WORK(&pool->free_work, async_free_zspage);
1896}
1897
1898static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage)
1899{
1900	struct page *page = get_first_page(zspage);
1901
1902	do {
1903		WARN_ON(!trylock_page(page));
1904		__SetPageMovable(page, &zsmalloc_mops);
1905		unlock_page(page);
1906	} while ((page = get_next_page(page)) != NULL);
1907}
1908#else
1909static inline void zs_flush_migration(struct zs_pool *pool) { }
1910#endif
1911
1912/*
1913 *
1914 * Based on the number of unused allocated objects calculate
1915 * and return the number of pages that we can free.
1916 */
1917static unsigned long zs_can_compact(struct size_class *class)
1918{
1919	unsigned long obj_wasted;
1920	unsigned long obj_allocated = class_stat_read(class, ZS_OBJS_ALLOCATED);
1921	unsigned long obj_used = class_stat_read(class, ZS_OBJS_INUSE);
1922
1923	if (obj_allocated <= obj_used)
1924		return 0;
1925
1926	obj_wasted = obj_allocated - obj_used;
1927	obj_wasted /= class->objs_per_zspage;
1928
1929	return obj_wasted * class->pages_per_zspage;
1930}
1931
1932static unsigned long __zs_compact(struct zs_pool *pool,
1933				  struct size_class *class)
1934{
1935	struct zspage *src_zspage = NULL;
1936	struct zspage *dst_zspage = NULL;
1937	unsigned long pages_freed = 0;
1938
1939	/*
1940	 * protect the race between zpage migration and zs_free
1941	 * as well as zpage allocation/free
1942	 */
1943	write_lock(&pool->migrate_lock);
1944	spin_lock(&class->lock);
1945	while (zs_can_compact(class)) {
1946		int fg;
1947
1948		if (!dst_zspage) {
1949			dst_zspage = isolate_dst_zspage(class);
1950			if (!dst_zspage)
1951				break;
1952		}
1953
1954		src_zspage = isolate_src_zspage(class);
1955		if (!src_zspage)
1956			break;
1957
1958		migrate_write_lock(src_zspage);
1959		migrate_zspage(pool, src_zspage, dst_zspage);
1960		migrate_write_unlock(src_zspage);
1961
1962		fg = putback_zspage(class, src_zspage);
1963		if (fg == ZS_INUSE_RATIO_0) {
1964			free_zspage(pool, class, src_zspage);
1965			pages_freed += class->pages_per_zspage;
1966		}
1967		src_zspage = NULL;
1968
1969		if (get_fullness_group(class, dst_zspage) == ZS_INUSE_RATIO_100
1970		    || rwlock_is_contended(&pool->migrate_lock)) {
1971			putback_zspage(class, dst_zspage);
1972			dst_zspage = NULL;
1973
1974			spin_unlock(&class->lock);
1975			write_unlock(&pool->migrate_lock);
1976			cond_resched();
1977			write_lock(&pool->migrate_lock);
1978			spin_lock(&class->lock);
1979		}
1980	}
1981
1982	if (src_zspage)
1983		putback_zspage(class, src_zspage);
1984
1985	if (dst_zspage)
1986		putback_zspage(class, dst_zspage);
1987
1988	spin_unlock(&class->lock);
1989	write_unlock(&pool->migrate_lock);
1990
1991	return pages_freed;
1992}
1993
1994unsigned long zs_compact(struct zs_pool *pool)
1995{
1996	int i;
1997	struct size_class *class;
1998	unsigned long pages_freed = 0;
1999
2000	/*
2001	 * Pool compaction is performed under pool->migrate_lock so it is basically
2002	 * single-threaded. Having more than one thread in __zs_compact()
2003	 * will increase pool->migrate_lock contention, which will impact other
2004	 * zsmalloc operations that need pool->migrate_lock.
2005	 */
2006	if (atomic_xchg(&pool->compaction_in_progress, 1))
2007		return 0;
2008
2009	for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2010		class = pool->size_class[i];
2011		if (class->index != i)
2012			continue;
2013		pages_freed += __zs_compact(pool, class);
2014	}
2015	atomic_long_add(pages_freed, &pool->stats.pages_compacted);
2016	atomic_set(&pool->compaction_in_progress, 0);
2017
2018	return pages_freed;
2019}
2020EXPORT_SYMBOL_GPL(zs_compact);
2021
2022void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats)
2023{
2024	memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats));
2025}
2026EXPORT_SYMBOL_GPL(zs_pool_stats);
2027
2028static unsigned long zs_shrinker_scan(struct shrinker *shrinker,
2029		struct shrink_control *sc)
2030{
2031	unsigned long pages_freed;
2032	struct zs_pool *pool = shrinker->private_data;
2033
2034	/*
2035	 * Compact classes and calculate compaction delta.
2036	 * Can run concurrently with a manually triggered
2037	 * (by user) compaction.
2038	 */
2039	pages_freed = zs_compact(pool);
2040
2041	return pages_freed ? pages_freed : SHRINK_STOP;
2042}
2043
2044static unsigned long zs_shrinker_count(struct shrinker *shrinker,
2045		struct shrink_control *sc)
2046{
2047	int i;
2048	struct size_class *class;
2049	unsigned long pages_to_free = 0;
2050	struct zs_pool *pool = shrinker->private_data;
2051
2052	for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2053		class = pool->size_class[i];
2054		if (class->index != i)
2055			continue;
2056
2057		pages_to_free += zs_can_compact(class);
2058	}
2059
2060	return pages_to_free;
2061}
2062
2063static void zs_unregister_shrinker(struct zs_pool *pool)
2064{
2065	shrinker_free(pool->shrinker);
2066}
2067
2068static int zs_register_shrinker(struct zs_pool *pool)
2069{
2070	pool->shrinker = shrinker_alloc(0, "mm-zspool:%s", pool->name);
2071	if (!pool->shrinker)
2072		return -ENOMEM;
2073
2074	pool->shrinker->scan_objects = zs_shrinker_scan;
2075	pool->shrinker->count_objects = zs_shrinker_count;
2076	pool->shrinker->batch = 0;
2077	pool->shrinker->private_data = pool;
2078
2079	shrinker_register(pool->shrinker);
2080
2081	return 0;
2082}
2083
2084static int calculate_zspage_chain_size(int class_size)
2085{
2086	int i, min_waste = INT_MAX;
2087	int chain_size = 1;
2088
2089	if (is_power_of_2(class_size))
2090		return chain_size;
2091
2092	for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
2093		int waste;
2094
2095		waste = (i * PAGE_SIZE) % class_size;
2096		if (waste < min_waste) {
2097			min_waste = waste;
2098			chain_size = i;
2099		}
2100	}
2101
2102	return chain_size;
2103}
2104
2105/**
2106 * zs_create_pool - Creates an allocation pool to work from.
2107 * @name: pool name to be created
2108 *
2109 * This function must be called before anything when using
2110 * the zsmalloc allocator.
2111 *
2112 * On success, a pointer to the newly created pool is returned,
2113 * otherwise NULL.
2114 */
2115struct zs_pool *zs_create_pool(const char *name)
2116{
2117	int i;
2118	struct zs_pool *pool;
2119	struct size_class *prev_class = NULL;
2120
2121	pool = kzalloc(sizeof(*pool), GFP_KERNEL);
2122	if (!pool)
2123		return NULL;
2124
2125	init_deferred_free(pool);
2126	rwlock_init(&pool->migrate_lock);
2127	atomic_set(&pool->compaction_in_progress, 0);
2128
2129	pool->name = kstrdup(name, GFP_KERNEL);
2130	if (!pool->name)
2131		goto err;
2132
2133	if (create_cache(pool))
2134		goto err;
2135
2136	/*
2137	 * Iterate reversely, because, size of size_class that we want to use
2138	 * for merging should be larger or equal to current size.
2139	 */
2140	for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2141		int size;
2142		int pages_per_zspage;
2143		int objs_per_zspage;
2144		struct size_class *class;
2145		int fullness;
2146
2147		size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
2148		if (size > ZS_MAX_ALLOC_SIZE)
2149			size = ZS_MAX_ALLOC_SIZE;
2150		pages_per_zspage = calculate_zspage_chain_size(size);
2151		objs_per_zspage = pages_per_zspage * PAGE_SIZE / size;
2152
2153		/*
2154		 * We iterate from biggest down to smallest classes,
2155		 * so huge_class_size holds the size of the first huge
2156		 * class. Any object bigger than or equal to that will
2157		 * endup in the huge class.
2158		 */
2159		if (pages_per_zspage != 1 && objs_per_zspage != 1 &&
2160				!huge_class_size) {
2161			huge_class_size = size;
2162			/*
2163			 * The object uses ZS_HANDLE_SIZE bytes to store the
2164			 * handle. We need to subtract it, because zs_malloc()
2165			 * unconditionally adds handle size before it performs
2166			 * size class search - so object may be smaller than
2167			 * huge class size, yet it still can end up in the huge
2168			 * class because it grows by ZS_HANDLE_SIZE extra bytes
2169			 * right before class lookup.
2170			 */
2171			huge_class_size -= (ZS_HANDLE_SIZE - 1);
2172		}
2173
2174		/*
2175		 * size_class is used for normal zsmalloc operation such
2176		 * as alloc/free for that size. Although it is natural that we
2177		 * have one size_class for each size, there is a chance that we
2178		 * can get more memory utilization if we use one size_class for
2179		 * many different sizes whose size_class have same
2180		 * characteristics. So, we makes size_class point to
2181		 * previous size_class if possible.
2182		 */
2183		if (prev_class) {
2184			if (can_merge(prev_class, pages_per_zspage, objs_per_zspage)) {
2185				pool->size_class[i] = prev_class;
2186				continue;
2187			}
2188		}
2189
2190		class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
2191		if (!class)
2192			goto err;
2193
2194		class->size = size;
2195		class->index = i;
2196		class->pages_per_zspage = pages_per_zspage;
2197		class->objs_per_zspage = objs_per_zspage;
2198		spin_lock_init(&class->lock);
2199		pool->size_class[i] = class;
2200
2201		fullness = ZS_INUSE_RATIO_0;
2202		while (fullness < NR_FULLNESS_GROUPS) {
2203			INIT_LIST_HEAD(&class->fullness_list[fullness]);
2204			fullness++;
2205		}
2206
2207		prev_class = class;
2208	}
2209
2210	/* debug only, don't abort if it fails */
2211	zs_pool_stat_create(pool, name);
2212
2213	/*
2214	 * Not critical since shrinker is only used to trigger internal
2215	 * defragmentation of the pool which is pretty optional thing.  If
2216	 * registration fails we still can use the pool normally and user can
2217	 * trigger compaction manually. Thus, ignore return code.
2218	 */
2219	zs_register_shrinker(pool);
2220
2221	return pool;
2222
2223err:
2224	zs_destroy_pool(pool);
2225	return NULL;
2226}
2227EXPORT_SYMBOL_GPL(zs_create_pool);
2228
2229void zs_destroy_pool(struct zs_pool *pool)
2230{
2231	int i;
2232
2233	zs_unregister_shrinker(pool);
2234	zs_flush_migration(pool);
2235	zs_pool_stat_destroy(pool);
2236
2237	for (i = 0; i < ZS_SIZE_CLASSES; i++) {
2238		int fg;
2239		struct size_class *class = pool->size_class[i];
2240
2241		if (!class)
2242			continue;
2243
2244		if (class->index != i)
2245			continue;
2246
2247		for (fg = ZS_INUSE_RATIO_0; fg < NR_FULLNESS_GROUPS; fg++) {
2248			if (list_empty(&class->fullness_list[fg]))
2249				continue;
2250
2251			pr_err("Class-%d fullness group %d is not empty\n",
2252			       class->size, fg);
2253		}
2254		kfree(class);
2255	}
2256
2257	destroy_cache(pool);
2258	kfree(pool->name);
2259	kfree(pool);
2260}
2261EXPORT_SYMBOL_GPL(zs_destroy_pool);
2262
2263static int __init zs_init(void)
2264{
2265	int ret;
2266
2267	ret = cpuhp_setup_state(CPUHP_MM_ZS_PREPARE, "mm/zsmalloc:prepare",
2268				zs_cpu_prepare, zs_cpu_dead);
2269	if (ret)
2270		goto out;
2271
2272#ifdef CONFIG_ZPOOL
2273	zpool_register_driver(&zs_zpool_driver);
2274#endif
2275
2276	zs_stat_init();
2277
2278	return 0;
2279
2280out:
2281	return ret;
2282}
2283
2284static void __exit zs_exit(void)
2285{
2286#ifdef CONFIG_ZPOOL
2287	zpool_unregister_driver(&zs_zpool_driver);
2288#endif
2289	cpuhp_remove_state(CPUHP_MM_ZS_PREPARE);
2290
2291	zs_stat_exit();
2292}
2293
2294module_init(zs_init);
2295module_exit(zs_exit);
2296
2297MODULE_LICENSE("Dual BSD/GPL");
2298MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");
2299MODULE_DESCRIPTION("zsmalloc memory allocator");