Linux Audio

Check our new training course

Embedded Linux training

Mar 10-20, 2025, special US time zones
Register
Loading...
Note: File does not exist in v3.1.
   1/*
   2 * zsmalloc memory allocator
   3 *
   4 * Copyright (C) 2011  Nitin Gupta
   5 * Copyright (C) 2012, 2013 Minchan Kim
   6 *
   7 * This code is released using a dual license strategy: BSD/GPL
   8 * You can choose the license that better fits your requirements.
   9 *
  10 * Released under the terms of 3-clause BSD License
  11 * Released under the terms of GNU General Public License Version 2.0
  12 */
  13
  14/*
  15 * Following is how we use various fields and flags of underlying
  16 * struct page(s) to form a zspage.
  17 *
  18 * Usage of struct page fields:
  19 *	page->private: points to zspage
  20 *	page->freelist(index): links together all component pages of a zspage
  21 *		For the huge page, this is always 0, so we use this field
  22 *		to store handle.
  23 *	page->units: first object offset in a subpage of zspage
  24 *
  25 * Usage of struct page flags:
  26 *	PG_private: identifies the first component page
  27 *	PG_owner_priv_1: identifies the huge component page
  28 *
  29 */
  30
  31#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  32
  33#include <linux/module.h>
  34#include <linux/kernel.h>
  35#include <linux/sched.h>
  36#include <linux/magic.h>
  37#include <linux/bitops.h>
  38#include <linux/errno.h>
  39#include <linux/highmem.h>
  40#include <linux/string.h>
  41#include <linux/slab.h>
  42#include <linux/pgtable.h>
  43#include <asm/tlbflush.h>
  44#include <linux/cpumask.h>
  45#include <linux/cpu.h>
  46#include <linux/vmalloc.h>
  47#include <linux/preempt.h>
  48#include <linux/spinlock.h>
  49#include <linux/shrinker.h>
  50#include <linux/types.h>
  51#include <linux/debugfs.h>
  52#include <linux/zsmalloc.h>
  53#include <linux/zpool.h>
  54#include <linux/mount.h>
  55#include <linux/pseudo_fs.h>
  56#include <linux/migrate.h>
  57#include <linux/wait.h>
  58#include <linux/pagemap.h>
  59#include <linux/fs.h>
  60
  61#define ZSPAGE_MAGIC	0x58
  62
  63/*
  64 * This must be power of 2 and greater than of equal to sizeof(link_free).
  65 * These two conditions ensure that any 'struct link_free' itself doesn't
  66 * span more than 1 page which avoids complex case of mapping 2 pages simply
  67 * to restore link_free pointer values.
  68 */
  69#define ZS_ALIGN		8
  70
  71/*
  72 * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
  73 * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
  74 */
  75#define ZS_MAX_ZSPAGE_ORDER 2
  76#define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
  77
  78#define ZS_HANDLE_SIZE (sizeof(unsigned long))
  79
  80/*
  81 * Object location (<PFN>, <obj_idx>) is encoded as
  82 * a single (unsigned long) handle value.
  83 *
  84 * Note that object index <obj_idx> starts from 0.
  85 *
  86 * This is made more complicated by various memory models and PAE.
  87 */
  88
  89#ifndef MAX_POSSIBLE_PHYSMEM_BITS
  90#ifdef MAX_PHYSMEM_BITS
  91#define MAX_POSSIBLE_PHYSMEM_BITS MAX_PHYSMEM_BITS
  92#else
  93/*
  94 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
  95 * be PAGE_SHIFT
  96 */
  97#define MAX_POSSIBLE_PHYSMEM_BITS BITS_PER_LONG
  98#endif
  99#endif
 100
 101#define _PFN_BITS		(MAX_POSSIBLE_PHYSMEM_BITS - PAGE_SHIFT)
 102
 103/*
 104 * Memory for allocating for handle keeps object position by
 105 * encoding <page, obj_idx> and the encoded value has a room
 106 * in least bit(ie, look at obj_to_location).
 107 * We use the bit to synchronize between object access by
 108 * user and migration.
 109 */
 110#define HANDLE_PIN_BIT	0
 111
 112/*
 113 * Head in allocated object should have OBJ_ALLOCATED_TAG
 114 * to identify the object was allocated or not.
 115 * It's okay to add the status bit in the least bit because
 116 * header keeps handle which is 4byte-aligned address so we
 117 * have room for two bit at least.
 118 */
 119#define OBJ_ALLOCATED_TAG 1
 120#define OBJ_TAG_BITS 1
 121#define OBJ_INDEX_BITS	(BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
 122#define OBJ_INDEX_MASK	((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
 123
 124#define FULLNESS_BITS	2
 125#define CLASS_BITS	8
 126#define ISOLATED_BITS	3
 127#define MAGIC_VAL_BITS	8
 128
 129#define MAX(a, b) ((a) >= (b) ? (a) : (b))
 130/* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
 131#define ZS_MIN_ALLOC_SIZE \
 132	MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
 133/* each chunk includes extra space to keep handle */
 134#define ZS_MAX_ALLOC_SIZE	PAGE_SIZE
 135
 136/*
 137 * On systems with 4K page size, this gives 255 size classes! There is a
 138 * trader-off here:
 139 *  - Large number of size classes is potentially wasteful as free page are
 140 *    spread across these classes
 141 *  - Small number of size classes causes large internal fragmentation
 142 *  - Probably its better to use specific size classes (empirically
 143 *    determined). NOTE: all those class sizes must be set as multiple of
 144 *    ZS_ALIGN to make sure link_free itself never has to span 2 pages.
 145 *
 146 *  ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
 147 *  (reason above)
 148 */
 149#define ZS_SIZE_CLASS_DELTA	(PAGE_SIZE >> CLASS_BITS)
 150#define ZS_SIZE_CLASSES	(DIV_ROUND_UP(ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE, \
 151				      ZS_SIZE_CLASS_DELTA) + 1)
 152
 153enum fullness_group {
 154	ZS_EMPTY,
 155	ZS_ALMOST_EMPTY,
 156	ZS_ALMOST_FULL,
 157	ZS_FULL,
 158	NR_ZS_FULLNESS,
 159};
 160
 161enum zs_stat_type {
 162	CLASS_EMPTY,
 163	CLASS_ALMOST_EMPTY,
 164	CLASS_ALMOST_FULL,
 165	CLASS_FULL,
 166	OBJ_ALLOCATED,
 167	OBJ_USED,
 168	NR_ZS_STAT_TYPE,
 169};
 170
 171struct zs_size_stat {
 172	unsigned long objs[NR_ZS_STAT_TYPE];
 173};
 174
 175#ifdef CONFIG_ZSMALLOC_STAT
 176static struct dentry *zs_stat_root;
 177#endif
 178
 179#ifdef CONFIG_COMPACTION
 180static struct vfsmount *zsmalloc_mnt;
 181#endif
 182
 183/*
 184 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
 185 *	n <= N / f, where
 186 * n = number of allocated objects
 187 * N = total number of objects zspage can store
 188 * f = fullness_threshold_frac
 189 *
 190 * Similarly, we assign zspage to:
 191 *	ZS_ALMOST_FULL	when n > N / f
 192 *	ZS_EMPTY	when n == 0
 193 *	ZS_FULL		when n == N
 194 *
 195 * (see: fix_fullness_group())
 196 */
 197static const int fullness_threshold_frac = 4;
 198static size_t huge_class_size;
 199
 200struct size_class {
 201	spinlock_t lock;
 202	struct list_head fullness_list[NR_ZS_FULLNESS];
 203	/*
 204	 * Size of objects stored in this class. Must be multiple
 205	 * of ZS_ALIGN.
 206	 */
 207	int size;
 208	int objs_per_zspage;
 209	/* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
 210	int pages_per_zspage;
 211
 212	unsigned int index;
 213	struct zs_size_stat stats;
 214};
 215
 216/* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
 217static void SetPageHugeObject(struct page *page)
 218{
 219	SetPageOwnerPriv1(page);
 220}
 221
 222static void ClearPageHugeObject(struct page *page)
 223{
 224	ClearPageOwnerPriv1(page);
 225}
 226
 227static int PageHugeObject(struct page *page)
 228{
 229	return PageOwnerPriv1(page);
 230}
 231
 232/*
 233 * Placed within free objects to form a singly linked list.
 234 * For every zspage, zspage->freeobj gives head of this list.
 235 *
 236 * This must be power of 2 and less than or equal to ZS_ALIGN
 237 */
 238struct link_free {
 239	union {
 240		/*
 241		 * Free object index;
 242		 * It's valid for non-allocated object
 243		 */
 244		unsigned long next;
 245		/*
 246		 * Handle of allocated object.
 247		 */
 248		unsigned long handle;
 249	};
 250};
 251
 252struct zs_pool {
 253	const char *name;
 254
 255	struct size_class *size_class[ZS_SIZE_CLASSES];
 256	struct kmem_cache *handle_cachep;
 257	struct kmem_cache *zspage_cachep;
 258
 259	atomic_long_t pages_allocated;
 260
 261	struct zs_pool_stats stats;
 262
 263	/* Compact classes */
 264	struct shrinker shrinker;
 265
 266#ifdef CONFIG_ZSMALLOC_STAT
 267	struct dentry *stat_dentry;
 268#endif
 269#ifdef CONFIG_COMPACTION
 270	struct inode *inode;
 271	struct work_struct free_work;
 272	/* A wait queue for when migration races with async_free_zspage() */
 273	struct wait_queue_head migration_wait;
 274	atomic_long_t isolated_pages;
 275	bool destroying;
 276#endif
 277};
 278
 279struct zspage {
 280	struct {
 281		unsigned int fullness:FULLNESS_BITS;
 282		unsigned int class:CLASS_BITS + 1;
 283		unsigned int isolated:ISOLATED_BITS;
 284		unsigned int magic:MAGIC_VAL_BITS;
 285	};
 286	unsigned int inuse;
 287	unsigned int freeobj;
 288	struct page *first_page;
 289	struct list_head list; /* fullness list */
 290#ifdef CONFIG_COMPACTION
 291	rwlock_t lock;
 292#endif
 293};
 294
 295struct mapping_area {
 296#ifdef CONFIG_ZSMALLOC_PGTABLE_MAPPING
 297	struct vm_struct *vm; /* vm area for mapping object that span pages */
 298#else
 299	char *vm_buf; /* copy buffer for objects that span pages */
 300#endif
 301	char *vm_addr; /* address of kmap_atomic()'ed pages */
 302	enum zs_mapmode vm_mm; /* mapping mode */
 303};
 304
 305#ifdef CONFIG_COMPACTION
 306static int zs_register_migration(struct zs_pool *pool);
 307static void zs_unregister_migration(struct zs_pool *pool);
 308static void migrate_lock_init(struct zspage *zspage);
 309static void migrate_read_lock(struct zspage *zspage);
 310static void migrate_read_unlock(struct zspage *zspage);
 311static void kick_deferred_free(struct zs_pool *pool);
 312static void init_deferred_free(struct zs_pool *pool);
 313static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage);
 314#else
 315static int zsmalloc_mount(void) { return 0; }
 316static void zsmalloc_unmount(void) {}
 317static int zs_register_migration(struct zs_pool *pool) { return 0; }
 318static void zs_unregister_migration(struct zs_pool *pool) {}
 319static void migrate_lock_init(struct zspage *zspage) {}
 320static void migrate_read_lock(struct zspage *zspage) {}
 321static void migrate_read_unlock(struct zspage *zspage) {}
 322static void kick_deferred_free(struct zs_pool *pool) {}
 323static void init_deferred_free(struct zs_pool *pool) {}
 324static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage) {}
 325#endif
 326
 327static int create_cache(struct zs_pool *pool)
 328{
 329	pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE,
 330					0, 0, NULL);
 331	if (!pool->handle_cachep)
 332		return 1;
 333
 334	pool->zspage_cachep = kmem_cache_create("zspage", sizeof(struct zspage),
 335					0, 0, NULL);
 336	if (!pool->zspage_cachep) {
 337		kmem_cache_destroy(pool->handle_cachep);
 338		pool->handle_cachep = NULL;
 339		return 1;
 340	}
 341
 342	return 0;
 343}
 344
 345static void destroy_cache(struct zs_pool *pool)
 346{
 347	kmem_cache_destroy(pool->handle_cachep);
 348	kmem_cache_destroy(pool->zspage_cachep);
 349}
 350
 351static unsigned long cache_alloc_handle(struct zs_pool *pool, gfp_t gfp)
 352{
 353	return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
 354			gfp & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
 355}
 356
 357static void cache_free_handle(struct zs_pool *pool, unsigned long handle)
 358{
 359	kmem_cache_free(pool->handle_cachep, (void *)handle);
 360}
 361
 362static struct zspage *cache_alloc_zspage(struct zs_pool *pool, gfp_t flags)
 363{
 364	return kmem_cache_alloc(pool->zspage_cachep,
 365			flags & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
 366}
 367
 368static void cache_free_zspage(struct zs_pool *pool, struct zspage *zspage)
 369{
 370	kmem_cache_free(pool->zspage_cachep, zspage);
 371}
 372
 373static void record_obj(unsigned long handle, unsigned long obj)
 374{
 375	/*
 376	 * lsb of @obj represents handle lock while other bits
 377	 * represent object value the handle is pointing so
 378	 * updating shouldn't do store tearing.
 379	 */
 380	WRITE_ONCE(*(unsigned long *)handle, obj);
 381}
 382
 383/* zpool driver */
 384
 385#ifdef CONFIG_ZPOOL
 386
 387static void *zs_zpool_create(const char *name, gfp_t gfp,
 388			     const struct zpool_ops *zpool_ops,
 389			     struct zpool *zpool)
 390{
 391	/*
 392	 * Ignore global gfp flags: zs_malloc() may be invoked from
 393	 * different contexts and its caller must provide a valid
 394	 * gfp mask.
 395	 */
 396	return zs_create_pool(name);
 397}
 398
 399static void zs_zpool_destroy(void *pool)
 400{
 401	zs_destroy_pool(pool);
 402}
 403
 404static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
 405			unsigned long *handle)
 406{
 407	*handle = zs_malloc(pool, size, gfp);
 408	return *handle ? 0 : -1;
 409}
 410static void zs_zpool_free(void *pool, unsigned long handle)
 411{
 412	zs_free(pool, handle);
 413}
 414
 415static void *zs_zpool_map(void *pool, unsigned long handle,
 416			enum zpool_mapmode mm)
 417{
 418	enum zs_mapmode zs_mm;
 419
 420	switch (mm) {
 421	case ZPOOL_MM_RO:
 422		zs_mm = ZS_MM_RO;
 423		break;
 424	case ZPOOL_MM_WO:
 425		zs_mm = ZS_MM_WO;
 426		break;
 427	case ZPOOL_MM_RW:
 428	default:
 429		zs_mm = ZS_MM_RW;
 430		break;
 431	}
 432
 433	return zs_map_object(pool, handle, zs_mm);
 434}
 435static void zs_zpool_unmap(void *pool, unsigned long handle)
 436{
 437	zs_unmap_object(pool, handle);
 438}
 439
 440static u64 zs_zpool_total_size(void *pool)
 441{
 442	return zs_get_total_pages(pool) << PAGE_SHIFT;
 443}
 444
 445static struct zpool_driver zs_zpool_driver = {
 446	.type =			  "zsmalloc",
 447	.owner =		  THIS_MODULE,
 448	.create =		  zs_zpool_create,
 449	.destroy =		  zs_zpool_destroy,
 450	.malloc_support_movable = true,
 451	.malloc =		  zs_zpool_malloc,
 452	.free =			  zs_zpool_free,
 453	.map =			  zs_zpool_map,
 454	.unmap =		  zs_zpool_unmap,
 455	.total_size =		  zs_zpool_total_size,
 456};
 457
 458MODULE_ALIAS("zpool-zsmalloc");
 459#endif /* CONFIG_ZPOOL */
 460
 461/* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
 462static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
 463
 464static bool is_zspage_isolated(struct zspage *zspage)
 465{
 466	return zspage->isolated;
 467}
 468
 469static __maybe_unused int is_first_page(struct page *page)
 470{
 471	return PagePrivate(page);
 472}
 473
 474/* Protected by class->lock */
 475static inline int get_zspage_inuse(struct zspage *zspage)
 476{
 477	return zspage->inuse;
 478}
 479
 480
 481static inline void mod_zspage_inuse(struct zspage *zspage, int val)
 482{
 483	zspage->inuse += val;
 484}
 485
 486static inline struct page *get_first_page(struct zspage *zspage)
 487{
 488	struct page *first_page = zspage->first_page;
 489
 490	VM_BUG_ON_PAGE(!is_first_page(first_page), first_page);
 491	return first_page;
 492}
 493
 494static inline int get_first_obj_offset(struct page *page)
 495{
 496	return page->units;
 497}
 498
 499static inline void set_first_obj_offset(struct page *page, int offset)
 500{
 501	page->units = offset;
 502}
 503
 504static inline unsigned int get_freeobj(struct zspage *zspage)
 505{
 506	return zspage->freeobj;
 507}
 508
 509static inline void set_freeobj(struct zspage *zspage, unsigned int obj)
 510{
 511	zspage->freeobj = obj;
 512}
 513
 514static void get_zspage_mapping(struct zspage *zspage,
 515				unsigned int *class_idx,
 516				enum fullness_group *fullness)
 517{
 518	BUG_ON(zspage->magic != ZSPAGE_MAGIC);
 519
 520	*fullness = zspage->fullness;
 521	*class_idx = zspage->class;
 522}
 523
 524static void set_zspage_mapping(struct zspage *zspage,
 525				unsigned int class_idx,
 526				enum fullness_group fullness)
 527{
 528	zspage->class = class_idx;
 529	zspage->fullness = fullness;
 530}
 531
 532/*
 533 * zsmalloc divides the pool into various size classes where each
 534 * class maintains a list of zspages where each zspage is divided
 535 * into equal sized chunks. Each allocation falls into one of these
 536 * classes depending on its size. This function returns index of the
 537 * size class which has chunk size big enough to hold the give size.
 538 */
 539static int get_size_class_index(int size)
 540{
 541	int idx = 0;
 542
 543	if (likely(size > ZS_MIN_ALLOC_SIZE))
 544		idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
 545				ZS_SIZE_CLASS_DELTA);
 546
 547	return min_t(int, ZS_SIZE_CLASSES - 1, idx);
 548}
 549
 550/* type can be of enum type zs_stat_type or fullness_group */
 551static inline void zs_stat_inc(struct size_class *class,
 552				int type, unsigned long cnt)
 553{
 554	class->stats.objs[type] += cnt;
 555}
 556
 557/* type can be of enum type zs_stat_type or fullness_group */
 558static inline void zs_stat_dec(struct size_class *class,
 559				int type, unsigned long cnt)
 560{
 561	class->stats.objs[type] -= cnt;
 562}
 563
 564/* type can be of enum type zs_stat_type or fullness_group */
 565static inline unsigned long zs_stat_get(struct size_class *class,
 566				int type)
 567{
 568	return class->stats.objs[type];
 569}
 570
 571#ifdef CONFIG_ZSMALLOC_STAT
 572
 573static void __init zs_stat_init(void)
 574{
 575	if (!debugfs_initialized()) {
 576		pr_warn("debugfs not available, stat dir not created\n");
 577		return;
 578	}
 579
 580	zs_stat_root = debugfs_create_dir("zsmalloc", NULL);
 581}
 582
 583static void __exit zs_stat_exit(void)
 584{
 585	debugfs_remove_recursive(zs_stat_root);
 586}
 587
 588static unsigned long zs_can_compact(struct size_class *class);
 589
 590static int zs_stats_size_show(struct seq_file *s, void *v)
 591{
 592	int i;
 593	struct zs_pool *pool = s->private;
 594	struct size_class *class;
 595	int objs_per_zspage;
 596	unsigned long class_almost_full, class_almost_empty;
 597	unsigned long obj_allocated, obj_used, pages_used, freeable;
 598	unsigned long total_class_almost_full = 0, total_class_almost_empty = 0;
 599	unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0;
 600	unsigned long total_freeable = 0;
 601
 602	seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n",
 603			"class", "size", "almost_full", "almost_empty",
 604			"obj_allocated", "obj_used", "pages_used",
 605			"pages_per_zspage", "freeable");
 606
 607	for (i = 0; i < ZS_SIZE_CLASSES; i++) {
 608		class = pool->size_class[i];
 609
 610		if (class->index != i)
 611			continue;
 612
 613		spin_lock(&class->lock);
 614		class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL);
 615		class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY);
 616		obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
 617		obj_used = zs_stat_get(class, OBJ_USED);
 618		freeable = zs_can_compact(class);
 619		spin_unlock(&class->lock);
 620
 621		objs_per_zspage = class->objs_per_zspage;
 622		pages_used = obj_allocated / objs_per_zspage *
 623				class->pages_per_zspage;
 624
 625		seq_printf(s, " %5u %5u %11lu %12lu %13lu"
 626				" %10lu %10lu %16d %8lu\n",
 627			i, class->size, class_almost_full, class_almost_empty,
 628			obj_allocated, obj_used, pages_used,
 629			class->pages_per_zspage, freeable);
 630
 631		total_class_almost_full += class_almost_full;
 632		total_class_almost_empty += class_almost_empty;
 633		total_objs += obj_allocated;
 634		total_used_objs += obj_used;
 635		total_pages += pages_used;
 636		total_freeable += freeable;
 637	}
 638
 639	seq_puts(s, "\n");
 640	seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n",
 641			"Total", "", total_class_almost_full,
 642			total_class_almost_empty, total_objs,
 643			total_used_objs, total_pages, "", total_freeable);
 644
 645	return 0;
 646}
 647DEFINE_SHOW_ATTRIBUTE(zs_stats_size);
 648
 649static void zs_pool_stat_create(struct zs_pool *pool, const char *name)
 650{
 651	if (!zs_stat_root) {
 652		pr_warn("no root stat dir, not creating <%s> stat dir\n", name);
 653		return;
 654	}
 655
 656	pool->stat_dentry = debugfs_create_dir(name, zs_stat_root);
 657
 658	debugfs_create_file("classes", S_IFREG | 0444, pool->stat_dentry, pool,
 659			    &zs_stats_size_fops);
 660}
 661
 662static void zs_pool_stat_destroy(struct zs_pool *pool)
 663{
 664	debugfs_remove_recursive(pool->stat_dentry);
 665}
 666
 667#else /* CONFIG_ZSMALLOC_STAT */
 668static void __init zs_stat_init(void)
 669{
 670}
 671
 672static void __exit zs_stat_exit(void)
 673{
 674}
 675
 676static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name)
 677{
 678}
 679
 680static inline void zs_pool_stat_destroy(struct zs_pool *pool)
 681{
 682}
 683#endif
 684
 685
 686/*
 687 * For each size class, zspages are divided into different groups
 688 * depending on how "full" they are. This was done so that we could
 689 * easily find empty or nearly empty zspages when we try to shrink
 690 * the pool (not yet implemented). This function returns fullness
 691 * status of the given page.
 692 */
 693static enum fullness_group get_fullness_group(struct size_class *class,
 694						struct zspage *zspage)
 695{
 696	int inuse, objs_per_zspage;
 697	enum fullness_group fg;
 698
 699	inuse = get_zspage_inuse(zspage);
 700	objs_per_zspage = class->objs_per_zspage;
 701
 702	if (inuse == 0)
 703		fg = ZS_EMPTY;
 704	else if (inuse == objs_per_zspage)
 705		fg = ZS_FULL;
 706	else if (inuse <= 3 * objs_per_zspage / fullness_threshold_frac)
 707		fg = ZS_ALMOST_EMPTY;
 708	else
 709		fg = ZS_ALMOST_FULL;
 710
 711	return fg;
 712}
 713
 714/*
 715 * Each size class maintains various freelists and zspages are assigned
 716 * to one of these freelists based on the number of live objects they
 717 * have. This functions inserts the given zspage into the freelist
 718 * identified by <class, fullness_group>.
 719 */
 720static void insert_zspage(struct size_class *class,
 721				struct zspage *zspage,
 722				enum fullness_group fullness)
 723{
 724	struct zspage *head;
 725
 726	zs_stat_inc(class, fullness, 1);
 727	head = list_first_entry_or_null(&class->fullness_list[fullness],
 728					struct zspage, list);
 729	/*
 730	 * We want to see more ZS_FULL pages and less almost empty/full.
 731	 * Put pages with higher ->inuse first.
 732	 */
 733	if (head) {
 734		if (get_zspage_inuse(zspage) < get_zspage_inuse(head)) {
 735			list_add(&zspage->list, &head->list);
 736			return;
 737		}
 738	}
 739	list_add(&zspage->list, &class->fullness_list[fullness]);
 740}
 741
 742/*
 743 * This function removes the given zspage from the freelist identified
 744 * by <class, fullness_group>.
 745 */
 746static void remove_zspage(struct size_class *class,
 747				struct zspage *zspage,
 748				enum fullness_group fullness)
 749{
 750	VM_BUG_ON(list_empty(&class->fullness_list[fullness]));
 751	VM_BUG_ON(is_zspage_isolated(zspage));
 752
 753	list_del_init(&zspage->list);
 754	zs_stat_dec(class, fullness, 1);
 755}
 756
 757/*
 758 * Each size class maintains zspages in different fullness groups depending
 759 * on the number of live objects they contain. When allocating or freeing
 760 * objects, the fullness status of the page can change, say, from ALMOST_FULL
 761 * to ALMOST_EMPTY when freeing an object. This function checks if such
 762 * a status change has occurred for the given page and accordingly moves the
 763 * page from the freelist of the old fullness group to that of the new
 764 * fullness group.
 765 */
 766static enum fullness_group fix_fullness_group(struct size_class *class,
 767						struct zspage *zspage)
 768{
 769	int class_idx;
 770	enum fullness_group currfg, newfg;
 771
 772	get_zspage_mapping(zspage, &class_idx, &currfg);
 773	newfg = get_fullness_group(class, zspage);
 774	if (newfg == currfg)
 775		goto out;
 776
 777	if (!is_zspage_isolated(zspage)) {
 778		remove_zspage(class, zspage, currfg);
 779		insert_zspage(class, zspage, newfg);
 780	}
 781
 782	set_zspage_mapping(zspage, class_idx, newfg);
 783
 784out:
 785	return newfg;
 786}
 787
 788/*
 789 * We have to decide on how many pages to link together
 790 * to form a zspage for each size class. This is important
 791 * to reduce wastage due to unusable space left at end of
 792 * each zspage which is given as:
 793 *     wastage = Zp % class_size
 794 *     usage = Zp - wastage
 795 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
 796 *
 797 * For example, for size class of 3/8 * PAGE_SIZE, we should
 798 * link together 3 PAGE_SIZE sized pages to form a zspage
 799 * since then we can perfectly fit in 8 such objects.
 800 */
 801static int get_pages_per_zspage(int class_size)
 802{
 803	int i, max_usedpc = 0;
 804	/* zspage order which gives maximum used size per KB */
 805	int max_usedpc_order = 1;
 806
 807	for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
 808		int zspage_size;
 809		int waste, usedpc;
 810
 811		zspage_size = i * PAGE_SIZE;
 812		waste = zspage_size % class_size;
 813		usedpc = (zspage_size - waste) * 100 / zspage_size;
 814
 815		if (usedpc > max_usedpc) {
 816			max_usedpc = usedpc;
 817			max_usedpc_order = i;
 818		}
 819	}
 820
 821	return max_usedpc_order;
 822}
 823
 824static struct zspage *get_zspage(struct page *page)
 825{
 826	struct zspage *zspage = (struct zspage *)page->private;
 827
 828	BUG_ON(zspage->magic != ZSPAGE_MAGIC);
 829	return zspage;
 830}
 831
 832static struct page *get_next_page(struct page *page)
 833{
 834	if (unlikely(PageHugeObject(page)))
 835		return NULL;
 836
 837	return page->freelist;
 838}
 839
 840/**
 841 * obj_to_location - get (<page>, <obj_idx>) from encoded object value
 842 * @obj: the encoded object value
 843 * @page: page object resides in zspage
 844 * @obj_idx: object index
 845 */
 846static void obj_to_location(unsigned long obj, struct page **page,
 847				unsigned int *obj_idx)
 848{
 849	obj >>= OBJ_TAG_BITS;
 850	*page = pfn_to_page(obj >> OBJ_INDEX_BITS);
 851	*obj_idx = (obj & OBJ_INDEX_MASK);
 852}
 853
 854/**
 855 * location_to_obj - get obj value encoded from (<page>, <obj_idx>)
 856 * @page: page object resides in zspage
 857 * @obj_idx: object index
 858 */
 859static unsigned long location_to_obj(struct page *page, unsigned int obj_idx)
 860{
 861	unsigned long obj;
 862
 863	obj = page_to_pfn(page) << OBJ_INDEX_BITS;
 864	obj |= obj_idx & OBJ_INDEX_MASK;
 865	obj <<= OBJ_TAG_BITS;
 866
 867	return obj;
 868}
 869
 870static unsigned long handle_to_obj(unsigned long handle)
 871{
 872	return *(unsigned long *)handle;
 873}
 874
 875static unsigned long obj_to_head(struct page *page, void *obj)
 876{
 877	if (unlikely(PageHugeObject(page))) {
 878		VM_BUG_ON_PAGE(!is_first_page(page), page);
 879		return page->index;
 880	} else
 881		return *(unsigned long *)obj;
 882}
 883
 884static inline int testpin_tag(unsigned long handle)
 885{
 886	return bit_spin_is_locked(HANDLE_PIN_BIT, (unsigned long *)handle);
 887}
 888
 889static inline int trypin_tag(unsigned long handle)
 890{
 891	return bit_spin_trylock(HANDLE_PIN_BIT, (unsigned long *)handle);
 892}
 893
 894static void pin_tag(unsigned long handle) __acquires(bitlock)
 895{
 896	bit_spin_lock(HANDLE_PIN_BIT, (unsigned long *)handle);
 897}
 898
 899static void unpin_tag(unsigned long handle) __releases(bitlock)
 900{
 901	bit_spin_unlock(HANDLE_PIN_BIT, (unsigned long *)handle);
 902}
 903
 904static void reset_page(struct page *page)
 905{
 906	__ClearPageMovable(page);
 907	ClearPagePrivate(page);
 908	set_page_private(page, 0);
 909	page_mapcount_reset(page);
 910	ClearPageHugeObject(page);
 911	page->freelist = NULL;
 912}
 913
 914static int trylock_zspage(struct zspage *zspage)
 915{
 916	struct page *cursor, *fail;
 917
 918	for (cursor = get_first_page(zspage); cursor != NULL; cursor =
 919					get_next_page(cursor)) {
 920		if (!trylock_page(cursor)) {
 921			fail = cursor;
 922			goto unlock;
 923		}
 924	}
 925
 926	return 1;
 927unlock:
 928	for (cursor = get_first_page(zspage); cursor != fail; cursor =
 929					get_next_page(cursor))
 930		unlock_page(cursor);
 931
 932	return 0;
 933}
 934
 935static void __free_zspage(struct zs_pool *pool, struct size_class *class,
 936				struct zspage *zspage)
 937{
 938	struct page *page, *next;
 939	enum fullness_group fg;
 940	unsigned int class_idx;
 941
 942	get_zspage_mapping(zspage, &class_idx, &fg);
 943
 944	assert_spin_locked(&class->lock);
 945
 946	VM_BUG_ON(get_zspage_inuse(zspage));
 947	VM_BUG_ON(fg != ZS_EMPTY);
 948
 949	next = page = get_first_page(zspage);
 950	do {
 951		VM_BUG_ON_PAGE(!PageLocked(page), page);
 952		next = get_next_page(page);
 953		reset_page(page);
 954		unlock_page(page);
 955		dec_zone_page_state(page, NR_ZSPAGES);
 956		put_page(page);
 957		page = next;
 958	} while (page != NULL);
 959
 960	cache_free_zspage(pool, zspage);
 961
 962	zs_stat_dec(class, OBJ_ALLOCATED, class->objs_per_zspage);
 963	atomic_long_sub(class->pages_per_zspage,
 964					&pool->pages_allocated);
 965}
 966
 967static void free_zspage(struct zs_pool *pool, struct size_class *class,
 968				struct zspage *zspage)
 969{
 970	VM_BUG_ON(get_zspage_inuse(zspage));
 971	VM_BUG_ON(list_empty(&zspage->list));
 972
 973	if (!trylock_zspage(zspage)) {
 974		kick_deferred_free(pool);
 975		return;
 976	}
 977
 978	remove_zspage(class, zspage, ZS_EMPTY);
 979	__free_zspage(pool, class, zspage);
 980}
 981
 982/* Initialize a newly allocated zspage */
 983static void init_zspage(struct size_class *class, struct zspage *zspage)
 984{
 985	unsigned int freeobj = 1;
 986	unsigned long off = 0;
 987	struct page *page = get_first_page(zspage);
 988
 989	while (page) {
 990		struct page *next_page;
 991		struct link_free *link;
 992		void *vaddr;
 993
 994		set_first_obj_offset(page, off);
 995
 996		vaddr = kmap_atomic(page);
 997		link = (struct link_free *)vaddr + off / sizeof(*link);
 998
 999		while ((off += class->size) < PAGE_SIZE) {
1000			link->next = freeobj++ << OBJ_TAG_BITS;
1001			link += class->size / sizeof(*link);
1002		}
1003
1004		/*
1005		 * We now come to the last (full or partial) object on this
1006		 * page, which must point to the first object on the next
1007		 * page (if present)
1008		 */
1009		next_page = get_next_page(page);
1010		if (next_page) {
1011			link->next = freeobj++ << OBJ_TAG_BITS;
1012		} else {
1013			/*
1014			 * Reset OBJ_TAG_BITS bit to last link to tell
1015			 * whether it's allocated object or not.
1016			 */
1017			link->next = -1UL << OBJ_TAG_BITS;
1018		}
1019		kunmap_atomic(vaddr);
1020		page = next_page;
1021		off %= PAGE_SIZE;
1022	}
1023
1024	set_freeobj(zspage, 0);
1025}
1026
1027static void create_page_chain(struct size_class *class, struct zspage *zspage,
1028				struct page *pages[])
1029{
1030	int i;
1031	struct page *page;
1032	struct page *prev_page = NULL;
1033	int nr_pages = class->pages_per_zspage;
1034
1035	/*
1036	 * Allocate individual pages and link them together as:
1037	 * 1. all pages are linked together using page->freelist
1038	 * 2. each sub-page point to zspage using page->private
1039	 *
1040	 * we set PG_private to identify the first page (i.e. no other sub-page
1041	 * has this flag set).
1042	 */
1043	for (i = 0; i < nr_pages; i++) {
1044		page = pages[i];
1045		set_page_private(page, (unsigned long)zspage);
1046		page->freelist = NULL;
1047		if (i == 0) {
1048			zspage->first_page = page;
1049			SetPagePrivate(page);
1050			if (unlikely(class->objs_per_zspage == 1 &&
1051					class->pages_per_zspage == 1))
1052				SetPageHugeObject(page);
1053		} else {
1054			prev_page->freelist = page;
1055		}
1056		prev_page = page;
1057	}
1058}
1059
1060/*
1061 * Allocate a zspage for the given size class
1062 */
1063static struct zspage *alloc_zspage(struct zs_pool *pool,
1064					struct size_class *class,
1065					gfp_t gfp)
1066{
1067	int i;
1068	struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE];
1069	struct zspage *zspage = cache_alloc_zspage(pool, gfp);
1070
1071	if (!zspage)
1072		return NULL;
1073
1074	memset(zspage, 0, sizeof(struct zspage));
1075	zspage->magic = ZSPAGE_MAGIC;
1076	migrate_lock_init(zspage);
1077
1078	for (i = 0; i < class->pages_per_zspage; i++) {
1079		struct page *page;
1080
1081		page = alloc_page(gfp);
1082		if (!page) {
1083			while (--i >= 0) {
1084				dec_zone_page_state(pages[i], NR_ZSPAGES);
1085				__free_page(pages[i]);
1086			}
1087			cache_free_zspage(pool, zspage);
1088			return NULL;
1089		}
1090
1091		inc_zone_page_state(page, NR_ZSPAGES);
1092		pages[i] = page;
1093	}
1094
1095	create_page_chain(class, zspage, pages);
1096	init_zspage(class, zspage);
1097
1098	return zspage;
1099}
1100
1101static struct zspage *find_get_zspage(struct size_class *class)
1102{
1103	int i;
1104	struct zspage *zspage;
1105
1106	for (i = ZS_ALMOST_FULL; i >= ZS_EMPTY; i--) {
1107		zspage = list_first_entry_or_null(&class->fullness_list[i],
1108				struct zspage, list);
1109		if (zspage)
1110			break;
1111	}
1112
1113	return zspage;
1114}
1115
1116#ifdef CONFIG_ZSMALLOC_PGTABLE_MAPPING
1117static inline int __zs_cpu_up(struct mapping_area *area)
1118{
1119	/*
1120	 * Make sure we don't leak memory if a cpu UP notification
1121	 * and zs_init() race and both call zs_cpu_up() on the same cpu
1122	 */
1123	if (area->vm)
1124		return 0;
1125	area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL);
1126	if (!area->vm)
1127		return -ENOMEM;
1128	return 0;
1129}
1130
1131static inline void __zs_cpu_down(struct mapping_area *area)
1132{
1133	if (area->vm)
1134		free_vm_area(area->vm);
1135	area->vm = NULL;
1136}
1137
1138static inline void *__zs_map_object(struct mapping_area *area,
1139				struct page *pages[2], int off, int size)
1140{
1141	unsigned long addr = (unsigned long)area->vm->addr;
1142
1143	BUG_ON(map_kernel_range(addr, PAGE_SIZE * 2, PAGE_KERNEL, pages) < 0);
1144	area->vm_addr = area->vm->addr;
1145	return area->vm_addr + off;
1146}
1147
1148static inline void __zs_unmap_object(struct mapping_area *area,
1149				struct page *pages[2], int off, int size)
1150{
1151	unsigned long addr = (unsigned long)area->vm_addr;
1152
1153	unmap_kernel_range(addr, PAGE_SIZE * 2);
1154}
1155
1156#else /* CONFIG_ZSMALLOC_PGTABLE_MAPPING */
1157
1158static inline int __zs_cpu_up(struct mapping_area *area)
1159{
1160	/*
1161	 * Make sure we don't leak memory if a cpu UP notification
1162	 * and zs_init() race and both call zs_cpu_up() on the same cpu
1163	 */
1164	if (area->vm_buf)
1165		return 0;
1166	area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
1167	if (!area->vm_buf)
1168		return -ENOMEM;
1169	return 0;
1170}
1171
1172static inline void __zs_cpu_down(struct mapping_area *area)
1173{
1174	kfree(area->vm_buf);
1175	area->vm_buf = NULL;
1176}
1177
1178static void *__zs_map_object(struct mapping_area *area,
1179			struct page *pages[2], int off, int size)
1180{
1181	int sizes[2];
1182	void *addr;
1183	char *buf = area->vm_buf;
1184
1185	/* disable page faults to match kmap_atomic() return conditions */
1186	pagefault_disable();
1187
1188	/* no read fastpath */
1189	if (area->vm_mm == ZS_MM_WO)
1190		goto out;
1191
1192	sizes[0] = PAGE_SIZE - off;
1193	sizes[1] = size - sizes[0];
1194
1195	/* copy object to per-cpu buffer */
1196	addr = kmap_atomic(pages[0]);
1197	memcpy(buf, addr + off, sizes[0]);
1198	kunmap_atomic(addr);
1199	addr = kmap_atomic(pages[1]);
1200	memcpy(buf + sizes[0], addr, sizes[1]);
1201	kunmap_atomic(addr);
1202out:
1203	return area->vm_buf;
1204}
1205
1206static void __zs_unmap_object(struct mapping_area *area,
1207			struct page *pages[2], int off, int size)
1208{
1209	int sizes[2];
1210	void *addr;
1211	char *buf;
1212
1213	/* no write fastpath */
1214	if (area->vm_mm == ZS_MM_RO)
1215		goto out;
1216
1217	buf = area->vm_buf;
1218	buf = buf + ZS_HANDLE_SIZE;
1219	size -= ZS_HANDLE_SIZE;
1220	off += ZS_HANDLE_SIZE;
1221
1222	sizes[0] = PAGE_SIZE - off;
1223	sizes[1] = size - sizes[0];
1224
1225	/* copy per-cpu buffer to object */
1226	addr = kmap_atomic(pages[0]);
1227	memcpy(addr + off, buf, sizes[0]);
1228	kunmap_atomic(addr);
1229	addr = kmap_atomic(pages[1]);
1230	memcpy(addr, buf + sizes[0], sizes[1]);
1231	kunmap_atomic(addr);
1232
1233out:
1234	/* enable page faults to match kunmap_atomic() return conditions */
1235	pagefault_enable();
1236}
1237
1238#endif /* CONFIG_ZSMALLOC_PGTABLE_MAPPING */
1239
1240static int zs_cpu_prepare(unsigned int cpu)
1241{
1242	struct mapping_area *area;
1243
1244	area = &per_cpu(zs_map_area, cpu);
1245	return __zs_cpu_up(area);
1246}
1247
1248static int zs_cpu_dead(unsigned int cpu)
1249{
1250	struct mapping_area *area;
1251
1252	area = &per_cpu(zs_map_area, cpu);
1253	__zs_cpu_down(area);
1254	return 0;
1255}
1256
1257static bool can_merge(struct size_class *prev, int pages_per_zspage,
1258					int objs_per_zspage)
1259{
1260	if (prev->pages_per_zspage == pages_per_zspage &&
1261		prev->objs_per_zspage == objs_per_zspage)
1262		return true;
1263
1264	return false;
1265}
1266
1267static bool zspage_full(struct size_class *class, struct zspage *zspage)
1268{
1269	return get_zspage_inuse(zspage) == class->objs_per_zspage;
1270}
1271
1272unsigned long zs_get_total_pages(struct zs_pool *pool)
1273{
1274	return atomic_long_read(&pool->pages_allocated);
1275}
1276EXPORT_SYMBOL_GPL(zs_get_total_pages);
1277
1278/**
1279 * zs_map_object - get address of allocated object from handle.
1280 * @pool: pool from which the object was allocated
1281 * @handle: handle returned from zs_malloc
1282 * @mm: maping mode to use
1283 *
1284 * Before using an object allocated from zs_malloc, it must be mapped using
1285 * this function. When done with the object, it must be unmapped using
1286 * zs_unmap_object.
1287 *
1288 * Only one object can be mapped per cpu at a time. There is no protection
1289 * against nested mappings.
1290 *
1291 * This function returns with preemption and page faults disabled.
1292 */
1293void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1294			enum zs_mapmode mm)
1295{
1296	struct zspage *zspage;
1297	struct page *page;
1298	unsigned long obj, off;
1299	unsigned int obj_idx;
1300
1301	unsigned int class_idx;
1302	enum fullness_group fg;
1303	struct size_class *class;
1304	struct mapping_area *area;
1305	struct page *pages[2];
1306	void *ret;
1307
1308	/*
1309	 * Because we use per-cpu mapping areas shared among the
1310	 * pools/users, we can't allow mapping in interrupt context
1311	 * because it can corrupt another users mappings.
1312	 */
1313	BUG_ON(in_interrupt());
1314
1315	/* From now on, migration cannot move the object */
1316	pin_tag(handle);
1317
1318	obj = handle_to_obj(handle);
1319	obj_to_location(obj, &page, &obj_idx);
1320	zspage = get_zspage(page);
1321
1322	/* migration cannot move any subpage in this zspage */
1323	migrate_read_lock(zspage);
1324
1325	get_zspage_mapping(zspage, &class_idx, &fg);
1326	class = pool->size_class[class_idx];
1327	off = (class->size * obj_idx) & ~PAGE_MASK;
1328
1329	area = &get_cpu_var(zs_map_area);
1330	area->vm_mm = mm;
1331	if (off + class->size <= PAGE_SIZE) {
1332		/* this object is contained entirely within a page */
1333		area->vm_addr = kmap_atomic(page);
1334		ret = area->vm_addr + off;
1335		goto out;
1336	}
1337
1338	/* this object spans two pages */
1339	pages[0] = page;
1340	pages[1] = get_next_page(page);
1341	BUG_ON(!pages[1]);
1342
1343	ret = __zs_map_object(area, pages, off, class->size);
1344out:
1345	if (likely(!PageHugeObject(page)))
1346		ret += ZS_HANDLE_SIZE;
1347
1348	return ret;
1349}
1350EXPORT_SYMBOL_GPL(zs_map_object);
1351
1352void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1353{
1354	struct zspage *zspage;
1355	struct page *page;
1356	unsigned long obj, off;
1357	unsigned int obj_idx;
1358
1359	unsigned int class_idx;
1360	enum fullness_group fg;
1361	struct size_class *class;
1362	struct mapping_area *area;
1363
1364	obj = handle_to_obj(handle);
1365	obj_to_location(obj, &page, &obj_idx);
1366	zspage = get_zspage(page);
1367	get_zspage_mapping(zspage, &class_idx, &fg);
1368	class = pool->size_class[class_idx];
1369	off = (class->size * obj_idx) & ~PAGE_MASK;
1370
1371	area = this_cpu_ptr(&zs_map_area);
1372	if (off + class->size <= PAGE_SIZE)
1373		kunmap_atomic(area->vm_addr);
1374	else {
1375		struct page *pages[2];
1376
1377		pages[0] = page;
1378		pages[1] = get_next_page(page);
1379		BUG_ON(!pages[1]);
1380
1381		__zs_unmap_object(area, pages, off, class->size);
1382	}
1383	put_cpu_var(zs_map_area);
1384
1385	migrate_read_unlock(zspage);
1386	unpin_tag(handle);
1387}
1388EXPORT_SYMBOL_GPL(zs_unmap_object);
1389
1390/**
1391 * zs_huge_class_size() - Returns the size (in bytes) of the first huge
1392 *                        zsmalloc &size_class.
1393 * @pool: zsmalloc pool to use
1394 *
1395 * The function returns the size of the first huge class - any object of equal
1396 * or bigger size will be stored in zspage consisting of a single physical
1397 * page.
1398 *
1399 * Context: Any context.
1400 *
1401 * Return: the size (in bytes) of the first huge zsmalloc &size_class.
1402 */
1403size_t zs_huge_class_size(struct zs_pool *pool)
1404{
1405	return huge_class_size;
1406}
1407EXPORT_SYMBOL_GPL(zs_huge_class_size);
1408
1409static unsigned long obj_malloc(struct size_class *class,
1410				struct zspage *zspage, unsigned long handle)
1411{
1412	int i, nr_page, offset;
1413	unsigned long obj;
1414	struct link_free *link;
1415
1416	struct page *m_page;
1417	unsigned long m_offset;
1418	void *vaddr;
1419
1420	handle |= OBJ_ALLOCATED_TAG;
1421	obj = get_freeobj(zspage);
1422
1423	offset = obj * class->size;
1424	nr_page = offset >> PAGE_SHIFT;
1425	m_offset = offset & ~PAGE_MASK;
1426	m_page = get_first_page(zspage);
1427
1428	for (i = 0; i < nr_page; i++)
1429		m_page = get_next_page(m_page);
1430
1431	vaddr = kmap_atomic(m_page);
1432	link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1433	set_freeobj(zspage, link->next >> OBJ_TAG_BITS);
1434	if (likely(!PageHugeObject(m_page)))
1435		/* record handle in the header of allocated chunk */
1436		link->handle = handle;
1437	else
1438		/* record handle to page->index */
1439		zspage->first_page->index = handle;
1440
1441	kunmap_atomic(vaddr);
1442	mod_zspage_inuse(zspage, 1);
1443	zs_stat_inc(class, OBJ_USED, 1);
1444
1445	obj = location_to_obj(m_page, obj);
1446
1447	return obj;
1448}
1449
1450
1451/**
1452 * zs_malloc - Allocate block of given size from pool.
1453 * @pool: pool to allocate from
1454 * @size: size of block to allocate
1455 * @gfp: gfp flags when allocating object
1456 *
1457 * On success, handle to the allocated object is returned,
1458 * otherwise 0.
1459 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1460 */
1461unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp)
1462{
1463	unsigned long handle, obj;
1464	struct size_class *class;
1465	enum fullness_group newfg;
1466	struct zspage *zspage;
1467
1468	if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
1469		return 0;
1470
1471	handle = cache_alloc_handle(pool, gfp);
1472	if (!handle)
1473		return 0;
1474
1475	/* extra space in chunk to keep the handle */
1476	size += ZS_HANDLE_SIZE;
1477	class = pool->size_class[get_size_class_index(size)];
1478
1479	spin_lock(&class->lock);
1480	zspage = find_get_zspage(class);
1481	if (likely(zspage)) {
1482		obj = obj_malloc(class, zspage, handle);
1483		/* Now move the zspage to another fullness group, if required */
1484		fix_fullness_group(class, zspage);
1485		record_obj(handle, obj);
1486		spin_unlock(&class->lock);
1487
1488		return handle;
1489	}
1490
1491	spin_unlock(&class->lock);
1492
1493	zspage = alloc_zspage(pool, class, gfp);
1494	if (!zspage) {
1495		cache_free_handle(pool, handle);
1496		return 0;
1497	}
1498
1499	spin_lock(&class->lock);
1500	obj = obj_malloc(class, zspage, handle);
1501	newfg = get_fullness_group(class, zspage);
1502	insert_zspage(class, zspage, newfg);
1503	set_zspage_mapping(zspage, class->index, newfg);
1504	record_obj(handle, obj);
1505	atomic_long_add(class->pages_per_zspage,
1506				&pool->pages_allocated);
1507	zs_stat_inc(class, OBJ_ALLOCATED, class->objs_per_zspage);
1508
1509	/* We completely set up zspage so mark them as movable */
1510	SetZsPageMovable(pool, zspage);
1511	spin_unlock(&class->lock);
1512
1513	return handle;
1514}
1515EXPORT_SYMBOL_GPL(zs_malloc);
1516
1517static void obj_free(struct size_class *class, unsigned long obj)
1518{
1519	struct link_free *link;
1520	struct zspage *zspage;
1521	struct page *f_page;
1522	unsigned long f_offset;
1523	unsigned int f_objidx;
1524	void *vaddr;
1525
1526	obj &= ~OBJ_ALLOCATED_TAG;
1527	obj_to_location(obj, &f_page, &f_objidx);
1528	f_offset = (class->size * f_objidx) & ~PAGE_MASK;
1529	zspage = get_zspage(f_page);
1530
1531	vaddr = kmap_atomic(f_page);
1532
1533	/* Insert this object in containing zspage's freelist */
1534	link = (struct link_free *)(vaddr + f_offset);
1535	link->next = get_freeobj(zspage) << OBJ_TAG_BITS;
1536	kunmap_atomic(vaddr);
1537	set_freeobj(zspage, f_objidx);
1538	mod_zspage_inuse(zspage, -1);
1539	zs_stat_dec(class, OBJ_USED, 1);
1540}
1541
1542void zs_free(struct zs_pool *pool, unsigned long handle)
1543{
1544	struct zspage *zspage;
1545	struct page *f_page;
1546	unsigned long obj;
1547	unsigned int f_objidx;
1548	int class_idx;
1549	struct size_class *class;
1550	enum fullness_group fullness;
1551	bool isolated;
1552
1553	if (unlikely(!handle))
1554		return;
1555
1556	pin_tag(handle);
1557	obj = handle_to_obj(handle);
1558	obj_to_location(obj, &f_page, &f_objidx);
1559	zspage = get_zspage(f_page);
1560
1561	migrate_read_lock(zspage);
1562
1563	get_zspage_mapping(zspage, &class_idx, &fullness);
1564	class = pool->size_class[class_idx];
1565
1566	spin_lock(&class->lock);
1567	obj_free(class, obj);
1568	fullness = fix_fullness_group(class, zspage);
1569	if (fullness != ZS_EMPTY) {
1570		migrate_read_unlock(zspage);
1571		goto out;
1572	}
1573
1574	isolated = is_zspage_isolated(zspage);
1575	migrate_read_unlock(zspage);
1576	/* If zspage is isolated, zs_page_putback will free the zspage */
1577	if (likely(!isolated))
1578		free_zspage(pool, class, zspage);
1579out:
1580
1581	spin_unlock(&class->lock);
1582	unpin_tag(handle);
1583	cache_free_handle(pool, handle);
1584}
1585EXPORT_SYMBOL_GPL(zs_free);
1586
1587static void zs_object_copy(struct size_class *class, unsigned long dst,
1588				unsigned long src)
1589{
1590	struct page *s_page, *d_page;
1591	unsigned int s_objidx, d_objidx;
1592	unsigned long s_off, d_off;
1593	void *s_addr, *d_addr;
1594	int s_size, d_size, size;
1595	int written = 0;
1596
1597	s_size = d_size = class->size;
1598
1599	obj_to_location(src, &s_page, &s_objidx);
1600	obj_to_location(dst, &d_page, &d_objidx);
1601
1602	s_off = (class->size * s_objidx) & ~PAGE_MASK;
1603	d_off = (class->size * d_objidx) & ~PAGE_MASK;
1604
1605	if (s_off + class->size > PAGE_SIZE)
1606		s_size = PAGE_SIZE - s_off;
1607
1608	if (d_off + class->size > PAGE_SIZE)
1609		d_size = PAGE_SIZE - d_off;
1610
1611	s_addr = kmap_atomic(s_page);
1612	d_addr = kmap_atomic(d_page);
1613
1614	while (1) {
1615		size = min(s_size, d_size);
1616		memcpy(d_addr + d_off, s_addr + s_off, size);
1617		written += size;
1618
1619		if (written == class->size)
1620			break;
1621
1622		s_off += size;
1623		s_size -= size;
1624		d_off += size;
1625		d_size -= size;
1626
1627		if (s_off >= PAGE_SIZE) {
1628			kunmap_atomic(d_addr);
1629			kunmap_atomic(s_addr);
1630			s_page = get_next_page(s_page);
1631			s_addr = kmap_atomic(s_page);
1632			d_addr = kmap_atomic(d_page);
1633			s_size = class->size - written;
1634			s_off = 0;
1635		}
1636
1637		if (d_off >= PAGE_SIZE) {
1638			kunmap_atomic(d_addr);
1639			d_page = get_next_page(d_page);
1640			d_addr = kmap_atomic(d_page);
1641			d_size = class->size - written;
1642			d_off = 0;
1643		}
1644	}
1645
1646	kunmap_atomic(d_addr);
1647	kunmap_atomic(s_addr);
1648}
1649
1650/*
1651 * Find alloced object in zspage from index object and
1652 * return handle.
1653 */
1654static unsigned long find_alloced_obj(struct size_class *class,
1655					struct page *page, int *obj_idx)
1656{
1657	unsigned long head;
1658	int offset = 0;
1659	int index = *obj_idx;
1660	unsigned long handle = 0;
1661	void *addr = kmap_atomic(page);
1662
1663	offset = get_first_obj_offset(page);
1664	offset += class->size * index;
1665
1666	while (offset < PAGE_SIZE) {
1667		head = obj_to_head(page, addr + offset);
1668		if (head & OBJ_ALLOCATED_TAG) {
1669			handle = head & ~OBJ_ALLOCATED_TAG;
1670			if (trypin_tag(handle))
1671				break;
1672			handle = 0;
1673		}
1674
1675		offset += class->size;
1676		index++;
1677	}
1678
1679	kunmap_atomic(addr);
1680
1681	*obj_idx = index;
1682
1683	return handle;
1684}
1685
1686struct zs_compact_control {
1687	/* Source spage for migration which could be a subpage of zspage */
1688	struct page *s_page;
1689	/* Destination page for migration which should be a first page
1690	 * of zspage. */
1691	struct page *d_page;
1692	 /* Starting object index within @s_page which used for live object
1693	  * in the subpage. */
1694	int obj_idx;
1695};
1696
1697static int migrate_zspage(struct zs_pool *pool, struct size_class *class,
1698				struct zs_compact_control *cc)
1699{
1700	unsigned long used_obj, free_obj;
1701	unsigned long handle;
1702	struct page *s_page = cc->s_page;
1703	struct page *d_page = cc->d_page;
1704	int obj_idx = cc->obj_idx;
1705	int ret = 0;
1706
1707	while (1) {
1708		handle = find_alloced_obj(class, s_page, &obj_idx);
1709		if (!handle) {
1710			s_page = get_next_page(s_page);
1711			if (!s_page)
1712				break;
1713			obj_idx = 0;
1714			continue;
1715		}
1716
1717		/* Stop if there is no more space */
1718		if (zspage_full(class, get_zspage(d_page))) {
1719			unpin_tag(handle);
1720			ret = -ENOMEM;
1721			break;
1722		}
1723
1724		used_obj = handle_to_obj(handle);
1725		free_obj = obj_malloc(class, get_zspage(d_page), handle);
1726		zs_object_copy(class, free_obj, used_obj);
1727		obj_idx++;
1728		/*
1729		 * record_obj updates handle's value to free_obj and it will
1730		 * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which
1731		 * breaks synchronization using pin_tag(e,g, zs_free) so
1732		 * let's keep the lock bit.
1733		 */
1734		free_obj |= BIT(HANDLE_PIN_BIT);
1735		record_obj(handle, free_obj);
1736		unpin_tag(handle);
1737		obj_free(class, used_obj);
1738	}
1739
1740	/* Remember last position in this iteration */
1741	cc->s_page = s_page;
1742	cc->obj_idx = obj_idx;
1743
1744	return ret;
1745}
1746
1747static struct zspage *isolate_zspage(struct size_class *class, bool source)
1748{
1749	int i;
1750	struct zspage *zspage;
1751	enum fullness_group fg[2] = {ZS_ALMOST_EMPTY, ZS_ALMOST_FULL};
1752
1753	if (!source) {
1754		fg[0] = ZS_ALMOST_FULL;
1755		fg[1] = ZS_ALMOST_EMPTY;
1756	}
1757
1758	for (i = 0; i < 2; i++) {
1759		zspage = list_first_entry_or_null(&class->fullness_list[fg[i]],
1760							struct zspage, list);
1761		if (zspage) {
1762			VM_BUG_ON(is_zspage_isolated(zspage));
1763			remove_zspage(class, zspage, fg[i]);
1764			return zspage;
1765		}
1766	}
1767
1768	return zspage;
1769}
1770
1771/*
1772 * putback_zspage - add @zspage into right class's fullness list
1773 * @class: destination class
1774 * @zspage: target page
1775 *
1776 * Return @zspage's fullness_group
1777 */
1778static enum fullness_group putback_zspage(struct size_class *class,
1779			struct zspage *zspage)
1780{
1781	enum fullness_group fullness;
1782
1783	VM_BUG_ON(is_zspage_isolated(zspage));
1784
1785	fullness = get_fullness_group(class, zspage);
1786	insert_zspage(class, zspage, fullness);
1787	set_zspage_mapping(zspage, class->index, fullness);
1788
1789	return fullness;
1790}
1791
1792#ifdef CONFIG_COMPACTION
1793/*
1794 * To prevent zspage destroy during migration, zspage freeing should
1795 * hold locks of all pages in the zspage.
1796 */
1797static void lock_zspage(struct zspage *zspage)
1798{
1799	struct page *page = get_first_page(zspage);
1800
1801	do {
1802		lock_page(page);
1803	} while ((page = get_next_page(page)) != NULL);
1804}
1805
1806static int zs_init_fs_context(struct fs_context *fc)
1807{
1808	return init_pseudo(fc, ZSMALLOC_MAGIC) ? 0 : -ENOMEM;
1809}
1810
1811static struct file_system_type zsmalloc_fs = {
1812	.name		= "zsmalloc",
1813	.init_fs_context = zs_init_fs_context,
1814	.kill_sb	= kill_anon_super,
1815};
1816
1817static int zsmalloc_mount(void)
1818{
1819	int ret = 0;
1820
1821	zsmalloc_mnt = kern_mount(&zsmalloc_fs);
1822	if (IS_ERR(zsmalloc_mnt))
1823		ret = PTR_ERR(zsmalloc_mnt);
1824
1825	return ret;
1826}
1827
1828static void zsmalloc_unmount(void)
1829{
1830	kern_unmount(zsmalloc_mnt);
1831}
1832
1833static void migrate_lock_init(struct zspage *zspage)
1834{
1835	rwlock_init(&zspage->lock);
1836}
1837
1838static void migrate_read_lock(struct zspage *zspage) __acquires(&zspage->lock)
1839{
1840	read_lock(&zspage->lock);
1841}
1842
1843static void migrate_read_unlock(struct zspage *zspage) __releases(&zspage->lock)
1844{
1845	read_unlock(&zspage->lock);
1846}
1847
1848static void migrate_write_lock(struct zspage *zspage)
1849{
1850	write_lock(&zspage->lock);
1851}
1852
1853static void migrate_write_unlock(struct zspage *zspage)
1854{
1855	write_unlock(&zspage->lock);
1856}
1857
1858/* Number of isolated subpage for *page migration* in this zspage */
1859static void inc_zspage_isolation(struct zspage *zspage)
1860{
1861	zspage->isolated++;
1862}
1863
1864static void dec_zspage_isolation(struct zspage *zspage)
1865{
1866	zspage->isolated--;
1867}
1868
1869static void putback_zspage_deferred(struct zs_pool *pool,
1870				    struct size_class *class,
1871				    struct zspage *zspage)
1872{
1873	enum fullness_group fg;
1874
1875	fg = putback_zspage(class, zspage);
1876	if (fg == ZS_EMPTY)
1877		schedule_work(&pool->free_work);
1878
1879}
1880
1881static inline void zs_pool_dec_isolated(struct zs_pool *pool)
1882{
1883	VM_BUG_ON(atomic_long_read(&pool->isolated_pages) <= 0);
1884	atomic_long_dec(&pool->isolated_pages);
1885	/*
1886	 * There's no possibility of racing, since wait_for_isolated_drain()
1887	 * checks the isolated count under &class->lock after enqueuing
1888	 * on migration_wait.
1889	 */
1890	if (atomic_long_read(&pool->isolated_pages) == 0 && pool->destroying)
1891		wake_up_all(&pool->migration_wait);
1892}
1893
1894static void replace_sub_page(struct size_class *class, struct zspage *zspage,
1895				struct page *newpage, struct page *oldpage)
1896{
1897	struct page *page;
1898	struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE] = {NULL, };
1899	int idx = 0;
1900
1901	page = get_first_page(zspage);
1902	do {
1903		if (page == oldpage)
1904			pages[idx] = newpage;
1905		else
1906			pages[idx] = page;
1907		idx++;
1908	} while ((page = get_next_page(page)) != NULL);
1909
1910	create_page_chain(class, zspage, pages);
1911	set_first_obj_offset(newpage, get_first_obj_offset(oldpage));
1912	if (unlikely(PageHugeObject(oldpage)))
1913		newpage->index = oldpage->index;
1914	__SetPageMovable(newpage, page_mapping(oldpage));
1915}
1916
1917static bool zs_page_isolate(struct page *page, isolate_mode_t mode)
1918{
1919	struct zs_pool *pool;
1920	struct size_class *class;
1921	int class_idx;
1922	enum fullness_group fullness;
1923	struct zspage *zspage;
1924	struct address_space *mapping;
1925
1926	/*
1927	 * Page is locked so zspage couldn't be destroyed. For detail, look at
1928	 * lock_zspage in free_zspage.
1929	 */
1930	VM_BUG_ON_PAGE(!PageMovable(page), page);
1931	VM_BUG_ON_PAGE(PageIsolated(page), page);
1932
1933	zspage = get_zspage(page);
1934
1935	/*
1936	 * Without class lock, fullness could be stale while class_idx is okay
1937	 * because class_idx is constant unless page is freed so we should get
1938	 * fullness again under class lock.
1939	 */
1940	get_zspage_mapping(zspage, &class_idx, &fullness);
1941	mapping = page_mapping(page);
1942	pool = mapping->private_data;
1943	class = pool->size_class[class_idx];
1944
1945	spin_lock(&class->lock);
1946	if (get_zspage_inuse(zspage) == 0) {
1947		spin_unlock(&class->lock);
1948		return false;
1949	}
1950
1951	/* zspage is isolated for object migration */
1952	if (list_empty(&zspage->list) && !is_zspage_isolated(zspage)) {
1953		spin_unlock(&class->lock);
1954		return false;
1955	}
1956
1957	/*
1958	 * If this is first time isolation for the zspage, isolate zspage from
1959	 * size_class to prevent further object allocation from the zspage.
1960	 */
1961	if (!list_empty(&zspage->list) && !is_zspage_isolated(zspage)) {
1962		get_zspage_mapping(zspage, &class_idx, &fullness);
1963		atomic_long_inc(&pool->isolated_pages);
1964		remove_zspage(class, zspage, fullness);
1965	}
1966
1967	inc_zspage_isolation(zspage);
1968	spin_unlock(&class->lock);
1969
1970	return true;
1971}
1972
1973static int zs_page_migrate(struct address_space *mapping, struct page *newpage,
1974		struct page *page, enum migrate_mode mode)
1975{
1976	struct zs_pool *pool;
1977	struct size_class *class;
1978	int class_idx;
1979	enum fullness_group fullness;
1980	struct zspage *zspage;
1981	struct page *dummy;
1982	void *s_addr, *d_addr, *addr;
1983	int offset, pos;
1984	unsigned long handle, head;
1985	unsigned long old_obj, new_obj;
1986	unsigned int obj_idx;
1987	int ret = -EAGAIN;
1988
1989	/*
1990	 * We cannot support the _NO_COPY case here, because copy needs to
1991	 * happen under the zs lock, which does not work with
1992	 * MIGRATE_SYNC_NO_COPY workflow.
1993	 */
1994	if (mode == MIGRATE_SYNC_NO_COPY)
1995		return -EINVAL;
1996
1997	VM_BUG_ON_PAGE(!PageMovable(page), page);
1998	VM_BUG_ON_PAGE(!PageIsolated(page), page);
1999
2000	zspage = get_zspage(page);
2001
2002	/* Concurrent compactor cannot migrate any subpage in zspage */
2003	migrate_write_lock(zspage);
2004	get_zspage_mapping(zspage, &class_idx, &fullness);
2005	pool = mapping->private_data;
2006	class = pool->size_class[class_idx];
2007	offset = get_first_obj_offset(page);
2008
2009	spin_lock(&class->lock);
2010	if (!get_zspage_inuse(zspage)) {
2011		/*
2012		 * Set "offset" to end of the page so that every loops
2013		 * skips unnecessary object scanning.
2014		 */
2015		offset = PAGE_SIZE;
2016	}
2017
2018	pos = offset;
2019	s_addr = kmap_atomic(page);
2020	while (pos < PAGE_SIZE) {
2021		head = obj_to_head(page, s_addr + pos);
2022		if (head & OBJ_ALLOCATED_TAG) {
2023			handle = head & ~OBJ_ALLOCATED_TAG;
2024			if (!trypin_tag(handle))
2025				goto unpin_objects;
2026		}
2027		pos += class->size;
2028	}
2029
2030	/*
2031	 * Here, any user cannot access all objects in the zspage so let's move.
2032	 */
2033	d_addr = kmap_atomic(newpage);
2034	memcpy(d_addr, s_addr, PAGE_SIZE);
2035	kunmap_atomic(d_addr);
2036
2037	for (addr = s_addr + offset; addr < s_addr + pos;
2038					addr += class->size) {
2039		head = obj_to_head(page, addr);
2040		if (head & OBJ_ALLOCATED_TAG) {
2041			handle = head & ~OBJ_ALLOCATED_TAG;
2042			if (!testpin_tag(handle))
2043				BUG();
2044
2045			old_obj = handle_to_obj(handle);
2046			obj_to_location(old_obj, &dummy, &obj_idx);
2047			new_obj = (unsigned long)location_to_obj(newpage,
2048								obj_idx);
2049			new_obj |= BIT(HANDLE_PIN_BIT);
2050			record_obj(handle, new_obj);
2051		}
2052	}
2053
2054	replace_sub_page(class, zspage, newpage, page);
2055	get_page(newpage);
2056
2057	dec_zspage_isolation(zspage);
2058
2059	/*
2060	 * Page migration is done so let's putback isolated zspage to
2061	 * the list if @page is final isolated subpage in the zspage.
2062	 */
2063	if (!is_zspage_isolated(zspage)) {
2064		/*
2065		 * We cannot race with zs_destroy_pool() here because we wait
2066		 * for isolation to hit zero before we start destroying.
2067		 * Also, we ensure that everyone can see pool->destroying before
2068		 * we start waiting.
2069		 */
2070		putback_zspage_deferred(pool, class, zspage);
2071		zs_pool_dec_isolated(pool);
2072	}
2073
2074	if (page_zone(newpage) != page_zone(page)) {
2075		dec_zone_page_state(page, NR_ZSPAGES);
2076		inc_zone_page_state(newpage, NR_ZSPAGES);
2077	}
2078
2079	reset_page(page);
2080	put_page(page);
2081	page = newpage;
2082
2083	ret = MIGRATEPAGE_SUCCESS;
2084unpin_objects:
2085	for (addr = s_addr + offset; addr < s_addr + pos;
2086						addr += class->size) {
2087		head = obj_to_head(page, addr);
2088		if (head & OBJ_ALLOCATED_TAG) {
2089			handle = head & ~OBJ_ALLOCATED_TAG;
2090			if (!testpin_tag(handle))
2091				BUG();
2092			unpin_tag(handle);
2093		}
2094	}
2095	kunmap_atomic(s_addr);
2096	spin_unlock(&class->lock);
2097	migrate_write_unlock(zspage);
2098
2099	return ret;
2100}
2101
2102static void zs_page_putback(struct page *page)
2103{
2104	struct zs_pool *pool;
2105	struct size_class *class;
2106	int class_idx;
2107	enum fullness_group fg;
2108	struct address_space *mapping;
2109	struct zspage *zspage;
2110
2111	VM_BUG_ON_PAGE(!PageMovable(page), page);
2112	VM_BUG_ON_PAGE(!PageIsolated(page), page);
2113
2114	zspage = get_zspage(page);
2115	get_zspage_mapping(zspage, &class_idx, &fg);
2116	mapping = page_mapping(page);
2117	pool = mapping->private_data;
2118	class = pool->size_class[class_idx];
2119
2120	spin_lock(&class->lock);
2121	dec_zspage_isolation(zspage);
2122	if (!is_zspage_isolated(zspage)) {
2123		/*
2124		 * Due to page_lock, we cannot free zspage immediately
2125		 * so let's defer.
2126		 */
2127		putback_zspage_deferred(pool, class, zspage);
2128		zs_pool_dec_isolated(pool);
2129	}
2130	spin_unlock(&class->lock);
2131}
2132
2133static const struct address_space_operations zsmalloc_aops = {
2134	.isolate_page = zs_page_isolate,
2135	.migratepage = zs_page_migrate,
2136	.putback_page = zs_page_putback,
2137};
2138
2139static int zs_register_migration(struct zs_pool *pool)
2140{
2141	pool->inode = alloc_anon_inode(zsmalloc_mnt->mnt_sb);
2142	if (IS_ERR(pool->inode)) {
2143		pool->inode = NULL;
2144		return 1;
2145	}
2146
2147	pool->inode->i_mapping->private_data = pool;
2148	pool->inode->i_mapping->a_ops = &zsmalloc_aops;
2149	return 0;
2150}
2151
2152static bool pool_isolated_are_drained(struct zs_pool *pool)
2153{
2154	return atomic_long_read(&pool->isolated_pages) == 0;
2155}
2156
2157/* Function for resolving migration */
2158static void wait_for_isolated_drain(struct zs_pool *pool)
2159{
2160
2161	/*
2162	 * We're in the process of destroying the pool, so there are no
2163	 * active allocations. zs_page_isolate() fails for completely free
2164	 * zspages, so we need only wait for the zs_pool's isolated
2165	 * count to hit zero.
2166	 */
2167	wait_event(pool->migration_wait,
2168		   pool_isolated_are_drained(pool));
2169}
2170
2171static void zs_unregister_migration(struct zs_pool *pool)
2172{
2173	pool->destroying = true;
2174	/*
2175	 * We need a memory barrier here to ensure global visibility of
2176	 * pool->destroying. Thus pool->isolated pages will either be 0 in which
2177	 * case we don't care, or it will be > 0 and pool->destroying will
2178	 * ensure that we wake up once isolation hits 0.
2179	 */
2180	smp_mb();
2181	wait_for_isolated_drain(pool); /* This can block */
2182	flush_work(&pool->free_work);
2183	iput(pool->inode);
2184}
2185
2186/*
2187 * Caller should hold page_lock of all pages in the zspage
2188 * In here, we cannot use zspage meta data.
2189 */
2190static void async_free_zspage(struct work_struct *work)
2191{
2192	int i;
2193	struct size_class *class;
2194	unsigned int class_idx;
2195	enum fullness_group fullness;
2196	struct zspage *zspage, *tmp;
2197	LIST_HEAD(free_pages);
2198	struct zs_pool *pool = container_of(work, struct zs_pool,
2199					free_work);
2200
2201	for (i = 0; i < ZS_SIZE_CLASSES; i++) {
2202		class = pool->size_class[i];
2203		if (class->index != i)
2204			continue;
2205
2206		spin_lock(&class->lock);
2207		list_splice_init(&class->fullness_list[ZS_EMPTY], &free_pages);
2208		spin_unlock(&class->lock);
2209	}
2210
2211
2212	list_for_each_entry_safe(zspage, tmp, &free_pages, list) {
2213		list_del(&zspage->list);
2214		lock_zspage(zspage);
2215
2216		get_zspage_mapping(zspage, &class_idx, &fullness);
2217		VM_BUG_ON(fullness != ZS_EMPTY);
2218		class = pool->size_class[class_idx];
2219		spin_lock(&class->lock);
2220		__free_zspage(pool, pool->size_class[class_idx], zspage);
2221		spin_unlock(&class->lock);
2222	}
2223};
2224
2225static void kick_deferred_free(struct zs_pool *pool)
2226{
2227	schedule_work(&pool->free_work);
2228}
2229
2230static void init_deferred_free(struct zs_pool *pool)
2231{
2232	INIT_WORK(&pool->free_work, async_free_zspage);
2233}
2234
2235static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage)
2236{
2237	struct page *page = get_first_page(zspage);
2238
2239	do {
2240		WARN_ON(!trylock_page(page));
2241		__SetPageMovable(page, pool->inode->i_mapping);
2242		unlock_page(page);
2243	} while ((page = get_next_page(page)) != NULL);
2244}
2245#endif
2246
2247/*
2248 *
2249 * Based on the number of unused allocated objects calculate
2250 * and return the number of pages that we can free.
2251 */
2252static unsigned long zs_can_compact(struct size_class *class)
2253{
2254	unsigned long obj_wasted;
2255	unsigned long obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
2256	unsigned long obj_used = zs_stat_get(class, OBJ_USED);
2257
2258	if (obj_allocated <= obj_used)
2259		return 0;
2260
2261	obj_wasted = obj_allocated - obj_used;
2262	obj_wasted /= class->objs_per_zspage;
2263
2264	return obj_wasted * class->pages_per_zspage;
2265}
2266
2267static void __zs_compact(struct zs_pool *pool, struct size_class *class)
2268{
2269	struct zs_compact_control cc;
2270	struct zspage *src_zspage;
2271	struct zspage *dst_zspage = NULL;
2272
2273	spin_lock(&class->lock);
2274	while ((src_zspage = isolate_zspage(class, true))) {
2275
2276		if (!zs_can_compact(class))
2277			break;
2278
2279		cc.obj_idx = 0;
2280		cc.s_page = get_first_page(src_zspage);
2281
2282		while ((dst_zspage = isolate_zspage(class, false))) {
2283			cc.d_page = get_first_page(dst_zspage);
2284			/*
2285			 * If there is no more space in dst_page, resched
2286			 * and see if anyone had allocated another zspage.
2287			 */
2288			if (!migrate_zspage(pool, class, &cc))
2289				break;
2290
2291			putback_zspage(class, dst_zspage);
2292		}
2293
2294		/* Stop if we couldn't find slot */
2295		if (dst_zspage == NULL)
2296			break;
2297
2298		putback_zspage(class, dst_zspage);
2299		if (putback_zspage(class, src_zspage) == ZS_EMPTY) {
2300			free_zspage(pool, class, src_zspage);
2301			pool->stats.pages_compacted += class->pages_per_zspage;
2302		}
2303		spin_unlock(&class->lock);
2304		cond_resched();
2305		spin_lock(&class->lock);
2306	}
2307
2308	if (src_zspage)
2309		putback_zspage(class, src_zspage);
2310
2311	spin_unlock(&class->lock);
2312}
2313
2314unsigned long zs_compact(struct zs_pool *pool)
2315{
2316	int i;
2317	struct size_class *class;
2318
2319	for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2320		class = pool->size_class[i];
2321		if (!class)
2322			continue;
2323		if (class->index != i)
2324			continue;
2325		__zs_compact(pool, class);
2326	}
2327
2328	return pool->stats.pages_compacted;
2329}
2330EXPORT_SYMBOL_GPL(zs_compact);
2331
2332void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats)
2333{
2334	memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats));
2335}
2336EXPORT_SYMBOL_GPL(zs_pool_stats);
2337
2338static unsigned long zs_shrinker_scan(struct shrinker *shrinker,
2339		struct shrink_control *sc)
2340{
2341	unsigned long pages_freed;
2342	struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2343			shrinker);
2344
2345	pages_freed = pool->stats.pages_compacted;
2346	/*
2347	 * Compact classes and calculate compaction delta.
2348	 * Can run concurrently with a manually triggered
2349	 * (by user) compaction.
2350	 */
2351	pages_freed = zs_compact(pool) - pages_freed;
2352
2353	return pages_freed ? pages_freed : SHRINK_STOP;
2354}
2355
2356static unsigned long zs_shrinker_count(struct shrinker *shrinker,
2357		struct shrink_control *sc)
2358{
2359	int i;
2360	struct size_class *class;
2361	unsigned long pages_to_free = 0;
2362	struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2363			shrinker);
2364
2365	for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2366		class = pool->size_class[i];
2367		if (!class)
2368			continue;
2369		if (class->index != i)
2370			continue;
2371
2372		pages_to_free += zs_can_compact(class);
2373	}
2374
2375	return pages_to_free;
2376}
2377
2378static void zs_unregister_shrinker(struct zs_pool *pool)
2379{
2380	unregister_shrinker(&pool->shrinker);
2381}
2382
2383static int zs_register_shrinker(struct zs_pool *pool)
2384{
2385	pool->shrinker.scan_objects = zs_shrinker_scan;
2386	pool->shrinker.count_objects = zs_shrinker_count;
2387	pool->shrinker.batch = 0;
2388	pool->shrinker.seeks = DEFAULT_SEEKS;
2389
2390	return register_shrinker(&pool->shrinker);
2391}
2392
2393/**
2394 * zs_create_pool - Creates an allocation pool to work from.
2395 * @name: pool name to be created
2396 *
2397 * This function must be called before anything when using
2398 * the zsmalloc allocator.
2399 *
2400 * On success, a pointer to the newly created pool is returned,
2401 * otherwise NULL.
2402 */
2403struct zs_pool *zs_create_pool(const char *name)
2404{
2405	int i;
2406	struct zs_pool *pool;
2407	struct size_class *prev_class = NULL;
2408
2409	pool = kzalloc(sizeof(*pool), GFP_KERNEL);
2410	if (!pool)
2411		return NULL;
2412
2413	init_deferred_free(pool);
2414
2415	pool->name = kstrdup(name, GFP_KERNEL);
2416	if (!pool->name)
2417		goto err;
2418
2419#ifdef CONFIG_COMPACTION
2420	init_waitqueue_head(&pool->migration_wait);
2421#endif
2422
2423	if (create_cache(pool))
2424		goto err;
2425
2426	/*
2427	 * Iterate reversely, because, size of size_class that we want to use
2428	 * for merging should be larger or equal to current size.
2429	 */
2430	for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2431		int size;
2432		int pages_per_zspage;
2433		int objs_per_zspage;
2434		struct size_class *class;
2435		int fullness = 0;
2436
2437		size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
2438		if (size > ZS_MAX_ALLOC_SIZE)
2439			size = ZS_MAX_ALLOC_SIZE;
2440		pages_per_zspage = get_pages_per_zspage(size);
2441		objs_per_zspage = pages_per_zspage * PAGE_SIZE / size;
2442
2443		/*
2444		 * We iterate from biggest down to smallest classes,
2445		 * so huge_class_size holds the size of the first huge
2446		 * class. Any object bigger than or equal to that will
2447		 * endup in the huge class.
2448		 */
2449		if (pages_per_zspage != 1 && objs_per_zspage != 1 &&
2450				!huge_class_size) {
2451			huge_class_size = size;
2452			/*
2453			 * The object uses ZS_HANDLE_SIZE bytes to store the
2454			 * handle. We need to subtract it, because zs_malloc()
2455			 * unconditionally adds handle size before it performs
2456			 * size class search - so object may be smaller than
2457			 * huge class size, yet it still can end up in the huge
2458			 * class because it grows by ZS_HANDLE_SIZE extra bytes
2459			 * right before class lookup.
2460			 */
2461			huge_class_size -= (ZS_HANDLE_SIZE - 1);
2462		}
2463
2464		/*
2465		 * size_class is used for normal zsmalloc operation such
2466		 * as alloc/free for that size. Although it is natural that we
2467		 * have one size_class for each size, there is a chance that we
2468		 * can get more memory utilization if we use one size_class for
2469		 * many different sizes whose size_class have same
2470		 * characteristics. So, we makes size_class point to
2471		 * previous size_class if possible.
2472		 */
2473		if (prev_class) {
2474			if (can_merge(prev_class, pages_per_zspage, objs_per_zspage)) {
2475				pool->size_class[i] = prev_class;
2476				continue;
2477			}
2478		}
2479
2480		class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
2481		if (!class)
2482			goto err;
2483
2484		class->size = size;
2485		class->index = i;
2486		class->pages_per_zspage = pages_per_zspage;
2487		class->objs_per_zspage = objs_per_zspage;
2488		spin_lock_init(&class->lock);
2489		pool->size_class[i] = class;
2490		for (fullness = ZS_EMPTY; fullness < NR_ZS_FULLNESS;
2491							fullness++)
2492			INIT_LIST_HEAD(&class->fullness_list[fullness]);
2493
2494		prev_class = class;
2495	}
2496
2497	/* debug only, don't abort if it fails */
2498	zs_pool_stat_create(pool, name);
2499
2500	if (zs_register_migration(pool))
2501		goto err;
2502
2503	/*
2504	 * Not critical since shrinker is only used to trigger internal
2505	 * defragmentation of the pool which is pretty optional thing.  If
2506	 * registration fails we still can use the pool normally and user can
2507	 * trigger compaction manually. Thus, ignore return code.
2508	 */
2509	zs_register_shrinker(pool);
2510
2511	return pool;
2512
2513err:
2514	zs_destroy_pool(pool);
2515	return NULL;
2516}
2517EXPORT_SYMBOL_GPL(zs_create_pool);
2518
2519void zs_destroy_pool(struct zs_pool *pool)
2520{
2521	int i;
2522
2523	zs_unregister_shrinker(pool);
2524	zs_unregister_migration(pool);
2525	zs_pool_stat_destroy(pool);
2526
2527	for (i = 0; i < ZS_SIZE_CLASSES; i++) {
2528		int fg;
2529		struct size_class *class = pool->size_class[i];
2530
2531		if (!class)
2532			continue;
2533
2534		if (class->index != i)
2535			continue;
2536
2537		for (fg = ZS_EMPTY; fg < NR_ZS_FULLNESS; fg++) {
2538			if (!list_empty(&class->fullness_list[fg])) {
2539				pr_info("Freeing non-empty class with size %db, fullness group %d\n",
2540					class->size, fg);
2541			}
2542		}
2543		kfree(class);
2544	}
2545
2546	destroy_cache(pool);
2547	kfree(pool->name);
2548	kfree(pool);
2549}
2550EXPORT_SYMBOL_GPL(zs_destroy_pool);
2551
2552static int __init zs_init(void)
2553{
2554	int ret;
2555
2556	ret = zsmalloc_mount();
2557	if (ret)
2558		goto out;
2559
2560	ret = cpuhp_setup_state(CPUHP_MM_ZS_PREPARE, "mm/zsmalloc:prepare",
2561				zs_cpu_prepare, zs_cpu_dead);
2562	if (ret)
2563		goto hp_setup_fail;
2564
2565#ifdef CONFIG_ZPOOL
2566	zpool_register_driver(&zs_zpool_driver);
2567#endif
2568
2569	zs_stat_init();
2570
2571	return 0;
2572
2573hp_setup_fail:
2574	zsmalloc_unmount();
2575out:
2576	return ret;
2577}
2578
2579static void __exit zs_exit(void)
2580{
2581#ifdef CONFIG_ZPOOL
2582	zpool_unregister_driver(&zs_zpool_driver);
2583#endif
2584	zsmalloc_unmount();
2585	cpuhp_remove_state(CPUHP_MM_ZS_PREPARE);
2586
2587	zs_stat_exit();
2588}
2589
2590module_init(zs_init);
2591module_exit(zs_exit);
2592
2593MODULE_LICENSE("Dual BSD/GPL");
2594MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");