Back to home page

LXR

 
 

    


0001 /*
0002  * zsmalloc memory allocator
0003  *
0004  * Copyright (C) 2011  Nitin Gupta
0005  * Copyright (C) 2012, 2013 Minchan Kim
0006  *
0007  * This code is released using a dual license strategy: BSD/GPL
0008  * You can choose the license that better fits your requirements.
0009  *
0010  * Released under the terms of 3-clause BSD License
0011  * Released under the terms of GNU General Public License Version 2.0
0012  */
0013 
0014 /*
0015  * Following is how we use various fields and flags of underlying
0016  * struct page(s) to form a zspage.
0017  *
0018  * Usage of struct page fields:
0019  *  page->private: points to zspage
0020  *  page->freelist(index): links together all component pages of a zspage
0021  *      For the huge page, this is always 0, so we use this field
0022  *      to store handle.
0023  *  page->units: first object offset in a subpage of zspage
0024  *
0025  * Usage of struct page flags:
0026  *  PG_private: identifies the first component page
0027  *  PG_private2: identifies the last component page
0028  *  PG_owner_priv_1: indentifies the huge component page
0029  *
0030  */
0031 
0032 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
0033 
0034 #include <linux/module.h>
0035 #include <linux/kernel.h>
0036 #include <linux/sched.h>
0037 #include <linux/bitops.h>
0038 #include <linux/errno.h>
0039 #include <linux/highmem.h>
0040 #include <linux/string.h>
0041 #include <linux/slab.h>
0042 #include <asm/tlbflush.h>
0043 #include <asm/pgtable.h>
0044 #include <linux/cpumask.h>
0045 #include <linux/cpu.h>
0046 #include <linux/vmalloc.h>
0047 #include <linux/preempt.h>
0048 #include <linux/spinlock.h>
0049 #include <linux/types.h>
0050 #include <linux/debugfs.h>
0051 #include <linux/zsmalloc.h>
0052 #include <linux/zpool.h>
0053 #include <linux/mount.h>
0054 #include <linux/migrate.h>
0055 #include <linux/pagemap.h>
0056 
0057 #define ZSPAGE_MAGIC    0x58
0058 
0059 /*
0060  * This must be power of 2 and greater than of equal to sizeof(link_free).
0061  * These two conditions ensure that any 'struct link_free' itself doesn't
0062  * span more than 1 page which avoids complex case of mapping 2 pages simply
0063  * to restore link_free pointer values.
0064  */
0065 #define ZS_ALIGN        8
0066 
0067 /*
0068  * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
0069  * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
0070  */
0071 #define ZS_MAX_ZSPAGE_ORDER 2
0072 #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
0073 
0074 #define ZS_HANDLE_SIZE (sizeof(unsigned long))
0075 
0076 /*
0077  * Object location (<PFN>, <obj_idx>) is encoded as
0078  * as single (unsigned long) handle value.
0079  *
0080  * Note that object index <obj_idx> starts from 0.
0081  *
0082  * This is made more complicated by various memory models and PAE.
0083  */
0084 
0085 #ifndef MAX_PHYSMEM_BITS
0086 #ifdef CONFIG_HIGHMEM64G
0087 #define MAX_PHYSMEM_BITS 36
0088 #else /* !CONFIG_HIGHMEM64G */
0089 /*
0090  * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
0091  * be PAGE_SHIFT
0092  */
0093 #define MAX_PHYSMEM_BITS BITS_PER_LONG
0094 #endif
0095 #endif
0096 #define _PFN_BITS       (MAX_PHYSMEM_BITS - PAGE_SHIFT)
0097 
0098 /*
0099  * Memory for allocating for handle keeps object position by
0100  * encoding <page, obj_idx> and the encoded value has a room
0101  * in least bit(ie, look at obj_to_location).
0102  * We use the bit to synchronize between object access by
0103  * user and migration.
0104  */
0105 #define HANDLE_PIN_BIT  0
0106 
0107 /*
0108  * Head in allocated object should have OBJ_ALLOCATED_TAG
0109  * to identify the object was allocated or not.
0110  * It's okay to add the status bit in the least bit because
0111  * header keeps handle which is 4byte-aligned address so we
0112  * have room for two bit at least.
0113  */
0114 #define OBJ_ALLOCATED_TAG 1
0115 #define OBJ_TAG_BITS 1
0116 #define OBJ_INDEX_BITS  (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
0117 #define OBJ_INDEX_MASK  ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
0118 
0119 #define MAX(a, b) ((a) >= (b) ? (a) : (b))
0120 /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
0121 #define ZS_MIN_ALLOC_SIZE \
0122     MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
0123 /* each chunk includes extra space to keep handle */
0124 #define ZS_MAX_ALLOC_SIZE   PAGE_SIZE
0125 
0126 /*
0127  * On systems with 4K page size, this gives 255 size classes! There is a
0128  * trader-off here:
0129  *  - Large number of size classes is potentially wasteful as free page are
0130  *    spread across these classes
0131  *  - Small number of size classes causes large internal fragmentation
0132  *  - Probably its better to use specific size classes (empirically
0133  *    determined). NOTE: all those class sizes must be set as multiple of
0134  *    ZS_ALIGN to make sure link_free itself never has to span 2 pages.
0135  *
0136  *  ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
0137  *  (reason above)
0138  */
0139 #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> CLASS_BITS)
0140 
0141 enum fullness_group {
0142     ZS_EMPTY,
0143     ZS_ALMOST_EMPTY,
0144     ZS_ALMOST_FULL,
0145     ZS_FULL,
0146     NR_ZS_FULLNESS,
0147 };
0148 
0149 enum zs_stat_type {
0150     CLASS_EMPTY,
0151     CLASS_ALMOST_EMPTY,
0152     CLASS_ALMOST_FULL,
0153     CLASS_FULL,
0154     OBJ_ALLOCATED,
0155     OBJ_USED,
0156     NR_ZS_STAT_TYPE,
0157 };
0158 
0159 struct zs_size_stat {
0160     unsigned long objs[NR_ZS_STAT_TYPE];
0161 };
0162 
0163 #ifdef CONFIG_ZSMALLOC_STAT
0164 static struct dentry *zs_stat_root;
0165 #endif
0166 
0167 #ifdef CONFIG_COMPACTION
0168 static struct vfsmount *zsmalloc_mnt;
0169 #endif
0170 
0171 /*
0172  * number of size_classes
0173  */
0174 static int zs_size_classes;
0175 
0176 /*
0177  * We assign a page to ZS_ALMOST_EMPTY fullness group when:
0178  *  n <= N / f, where
0179  * n = number of allocated objects
0180  * N = total number of objects zspage can store
0181  * f = fullness_threshold_frac
0182  *
0183  * Similarly, we assign zspage to:
0184  *  ZS_ALMOST_FULL  when n > N / f
0185  *  ZS_EMPTY    when n == 0
0186  *  ZS_FULL     when n == N
0187  *
0188  * (see: fix_fullness_group())
0189  */
0190 static const int fullness_threshold_frac = 4;
0191 
0192 struct size_class {
0193     spinlock_t lock;
0194     struct list_head fullness_list[NR_ZS_FULLNESS];
0195     /*
0196      * Size of objects stored in this class. Must be multiple
0197      * of ZS_ALIGN.
0198      */
0199     int size;
0200     int objs_per_zspage;
0201     /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
0202     int pages_per_zspage;
0203 
0204     unsigned int index;
0205     struct zs_size_stat stats;
0206 };
0207 
0208 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
0209 static void SetPageHugeObject(struct page *page)
0210 {
0211     SetPageOwnerPriv1(page);
0212 }
0213 
0214 static void ClearPageHugeObject(struct page *page)
0215 {
0216     ClearPageOwnerPriv1(page);
0217 }
0218 
0219 static int PageHugeObject(struct page *page)
0220 {
0221     return PageOwnerPriv1(page);
0222 }
0223 
0224 /*
0225  * Placed within free objects to form a singly linked list.
0226  * For every zspage, zspage->freeobj gives head of this list.
0227  *
0228  * This must be power of 2 and less than or equal to ZS_ALIGN
0229  */
0230 struct link_free {
0231     union {
0232         /*
0233          * Free object index;
0234          * It's valid for non-allocated object
0235          */
0236         unsigned long next;
0237         /*
0238          * Handle of allocated object.
0239          */
0240         unsigned long handle;
0241     };
0242 };
0243 
0244 struct zs_pool {
0245     const char *name;
0246 
0247     struct size_class **size_class;
0248     struct kmem_cache *handle_cachep;
0249     struct kmem_cache *zspage_cachep;
0250 
0251     atomic_long_t pages_allocated;
0252 
0253     struct zs_pool_stats stats;
0254 
0255     /* Compact classes */
0256     struct shrinker shrinker;
0257     /*
0258      * To signify that register_shrinker() was successful
0259      * and unregister_shrinker() will not Oops.
0260      */
0261     bool shrinker_enabled;
0262 #ifdef CONFIG_ZSMALLOC_STAT
0263     struct dentry *stat_dentry;
0264 #endif
0265 #ifdef CONFIG_COMPACTION
0266     struct inode *inode;
0267     struct work_struct free_work;
0268 #endif
0269 };
0270 
0271 /*
0272  * A zspage's class index and fullness group
0273  * are encoded in its (first)page->mapping
0274  */
0275 #define FULLNESS_BITS   2
0276 #define CLASS_BITS  8
0277 #define ISOLATED_BITS   3
0278 #define MAGIC_VAL_BITS  8
0279 
0280 struct zspage {
0281     struct {
0282         unsigned int fullness:FULLNESS_BITS;
0283         unsigned int class:CLASS_BITS;
0284         unsigned int isolated:ISOLATED_BITS;
0285         unsigned int magic:MAGIC_VAL_BITS;
0286     };
0287     unsigned int inuse;
0288     unsigned int freeobj;
0289     struct page *first_page;
0290     struct list_head list; /* fullness list */
0291 #ifdef CONFIG_COMPACTION
0292     rwlock_t lock;
0293 #endif
0294 };
0295 
0296 struct mapping_area {
0297 #ifdef CONFIG_PGTABLE_MAPPING
0298     struct vm_struct *vm; /* vm area for mapping object that span pages */
0299 #else
0300     char *vm_buf; /* copy buffer for objects that span pages */
0301 #endif
0302     char *vm_addr; /* address of kmap_atomic()'ed pages */
0303     enum zs_mapmode vm_mm; /* mapping mode */
0304 };
0305 
0306 #ifdef CONFIG_COMPACTION
0307 static int zs_register_migration(struct zs_pool *pool);
0308 static void zs_unregister_migration(struct zs_pool *pool);
0309 static void migrate_lock_init(struct zspage *zspage);
0310 static void migrate_read_lock(struct zspage *zspage);
0311 static void migrate_read_unlock(struct zspage *zspage);
0312 static void kick_deferred_free(struct zs_pool *pool);
0313 static void init_deferred_free(struct zs_pool *pool);
0314 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage);
0315 #else
0316 static int zsmalloc_mount(void) { return 0; }
0317 static void zsmalloc_unmount(void) {}
0318 static int zs_register_migration(struct zs_pool *pool) { return 0; }
0319 static void zs_unregister_migration(struct zs_pool *pool) {}
0320 static void migrate_lock_init(struct zspage *zspage) {}
0321 static void migrate_read_lock(struct zspage *zspage) {}
0322 static void migrate_read_unlock(struct zspage *zspage) {}
0323 static void kick_deferred_free(struct zs_pool *pool) {}
0324 static void init_deferred_free(struct zs_pool *pool) {}
0325 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage) {}
0326 #endif
0327 
0328 static int create_cache(struct zs_pool *pool)
0329 {
0330     pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE,
0331                     0, 0, NULL);
0332     if (!pool->handle_cachep)
0333         return 1;
0334 
0335     pool->zspage_cachep = kmem_cache_create("zspage", sizeof(struct zspage),
0336                     0, 0, NULL);
0337     if (!pool->zspage_cachep) {
0338         kmem_cache_destroy(pool->handle_cachep);
0339         pool->handle_cachep = NULL;
0340         return 1;
0341     }
0342 
0343     return 0;
0344 }
0345 
0346 static void destroy_cache(struct zs_pool *pool)
0347 {
0348     kmem_cache_destroy(pool->handle_cachep);
0349     kmem_cache_destroy(pool->zspage_cachep);
0350 }
0351 
0352 static unsigned long cache_alloc_handle(struct zs_pool *pool, gfp_t gfp)
0353 {
0354     return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
0355             gfp & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
0356 }
0357 
0358 static void cache_free_handle(struct zs_pool *pool, unsigned long handle)
0359 {
0360     kmem_cache_free(pool->handle_cachep, (void *)handle);
0361 }
0362 
0363 static struct zspage *cache_alloc_zspage(struct zs_pool *pool, gfp_t flags)
0364 {
0365     return kmem_cache_alloc(pool->zspage_cachep,
0366             flags & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
0367 };
0368 
0369 static void cache_free_zspage(struct zs_pool *pool, struct zspage *zspage)
0370 {
0371     kmem_cache_free(pool->zspage_cachep, zspage);
0372 }
0373 
0374 static void record_obj(unsigned long handle, unsigned long obj)
0375 {
0376     /*
0377      * lsb of @obj represents handle lock while other bits
0378      * represent object value the handle is pointing so
0379      * updating shouldn't do store tearing.
0380      */
0381     WRITE_ONCE(*(unsigned long *)handle, obj);
0382 }
0383 
0384 /* zpool driver */
0385 
0386 #ifdef CONFIG_ZPOOL
0387 
0388 static void *zs_zpool_create(const char *name, gfp_t gfp,
0389                  const struct zpool_ops *zpool_ops,
0390                  struct zpool *zpool)
0391 {
0392     /*
0393      * Ignore global gfp flags: zs_malloc() may be invoked from
0394      * different contexts and its caller must provide a valid
0395      * gfp mask.
0396      */
0397     return zs_create_pool(name);
0398 }
0399 
0400 static void zs_zpool_destroy(void *pool)
0401 {
0402     zs_destroy_pool(pool);
0403 }
0404 
0405 static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
0406             unsigned long *handle)
0407 {
0408     *handle = zs_malloc(pool, size, gfp);
0409     return *handle ? 0 : -1;
0410 }
0411 static void zs_zpool_free(void *pool, unsigned long handle)
0412 {
0413     zs_free(pool, handle);
0414 }
0415 
0416 static int zs_zpool_shrink(void *pool, unsigned int pages,
0417             unsigned int *reclaimed)
0418 {
0419     return -EINVAL;
0420 }
0421 
0422 static void *zs_zpool_map(void *pool, unsigned long handle,
0423             enum zpool_mapmode mm)
0424 {
0425     enum zs_mapmode zs_mm;
0426 
0427     switch (mm) {
0428     case ZPOOL_MM_RO:
0429         zs_mm = ZS_MM_RO;
0430         break;
0431     case ZPOOL_MM_WO:
0432         zs_mm = ZS_MM_WO;
0433         break;
0434     case ZPOOL_MM_RW: /* fallthru */
0435     default:
0436         zs_mm = ZS_MM_RW;
0437         break;
0438     }
0439 
0440     return zs_map_object(pool, handle, zs_mm);
0441 }
0442 static void zs_zpool_unmap(void *pool, unsigned long handle)
0443 {
0444     zs_unmap_object(pool, handle);
0445 }
0446 
0447 static u64 zs_zpool_total_size(void *pool)
0448 {
0449     return zs_get_total_pages(pool) << PAGE_SHIFT;
0450 }
0451 
0452 static struct zpool_driver zs_zpool_driver = {
0453     .type =     "zsmalloc",
0454     .owner =    THIS_MODULE,
0455     .create =   zs_zpool_create,
0456     .destroy =  zs_zpool_destroy,
0457     .malloc =   zs_zpool_malloc,
0458     .free =     zs_zpool_free,
0459     .shrink =   zs_zpool_shrink,
0460     .map =      zs_zpool_map,
0461     .unmap =    zs_zpool_unmap,
0462     .total_size =   zs_zpool_total_size,
0463 };
0464 
0465 MODULE_ALIAS("zpool-zsmalloc");
0466 #endif /* CONFIG_ZPOOL */
0467 
0468 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
0469 static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
0470 
0471 static bool is_zspage_isolated(struct zspage *zspage)
0472 {
0473     return zspage->isolated;
0474 }
0475 
0476 static int is_first_page(struct page *page)
0477 {
0478     return PagePrivate(page);
0479 }
0480 
0481 /* Protected by class->lock */
0482 static inline int get_zspage_inuse(struct zspage *zspage)
0483 {
0484     return zspage->inuse;
0485 }
0486 
0487 static inline void set_zspage_inuse(struct zspage *zspage, int val)
0488 {
0489     zspage->inuse = val;
0490 }
0491 
0492 static inline void mod_zspage_inuse(struct zspage *zspage, int val)
0493 {
0494     zspage->inuse += val;
0495 }
0496 
0497 static inline struct page *get_first_page(struct zspage *zspage)
0498 {
0499     struct page *first_page = zspage->first_page;
0500 
0501     VM_BUG_ON_PAGE(!is_first_page(first_page), first_page);
0502     return first_page;
0503 }
0504 
0505 static inline int get_first_obj_offset(struct page *page)
0506 {
0507     return page->units;
0508 }
0509 
0510 static inline void set_first_obj_offset(struct page *page, int offset)
0511 {
0512     page->units = offset;
0513 }
0514 
0515 static inline unsigned int get_freeobj(struct zspage *zspage)
0516 {
0517     return zspage->freeobj;
0518 }
0519 
0520 static inline void set_freeobj(struct zspage *zspage, unsigned int obj)
0521 {
0522     zspage->freeobj = obj;
0523 }
0524 
0525 static void get_zspage_mapping(struct zspage *zspage,
0526                 unsigned int *class_idx,
0527                 enum fullness_group *fullness)
0528 {
0529     BUG_ON(zspage->magic != ZSPAGE_MAGIC);
0530 
0531     *fullness = zspage->fullness;
0532     *class_idx = zspage->class;
0533 }
0534 
0535 static void set_zspage_mapping(struct zspage *zspage,
0536                 unsigned int class_idx,
0537                 enum fullness_group fullness)
0538 {
0539     zspage->class = class_idx;
0540     zspage->fullness = fullness;
0541 }
0542 
0543 /*
0544  * zsmalloc divides the pool into various size classes where each
0545  * class maintains a list of zspages where each zspage is divided
0546  * into equal sized chunks. Each allocation falls into one of these
0547  * classes depending on its size. This function returns index of the
0548  * size class which has chunk size big enough to hold the give size.
0549  */
0550 static int get_size_class_index(int size)
0551 {
0552     int idx = 0;
0553 
0554     if (likely(size > ZS_MIN_ALLOC_SIZE))
0555         idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
0556                 ZS_SIZE_CLASS_DELTA);
0557 
0558     return min(zs_size_classes - 1, idx);
0559 }
0560 
0561 static inline void zs_stat_inc(struct size_class *class,
0562                 enum zs_stat_type type, unsigned long cnt)
0563 {
0564     class->stats.objs[type] += cnt;
0565 }
0566 
0567 static inline void zs_stat_dec(struct size_class *class,
0568                 enum zs_stat_type type, unsigned long cnt)
0569 {
0570     class->stats.objs[type] -= cnt;
0571 }
0572 
0573 static inline unsigned long zs_stat_get(struct size_class *class,
0574                 enum zs_stat_type type)
0575 {
0576     return class->stats.objs[type];
0577 }
0578 
0579 #ifdef CONFIG_ZSMALLOC_STAT
0580 
0581 static void __init zs_stat_init(void)
0582 {
0583     if (!debugfs_initialized()) {
0584         pr_warn("debugfs not available, stat dir not created\n");
0585         return;
0586     }
0587 
0588     zs_stat_root = debugfs_create_dir("zsmalloc", NULL);
0589     if (!zs_stat_root)
0590         pr_warn("debugfs 'zsmalloc' stat dir creation failed\n");
0591 }
0592 
0593 static void __exit zs_stat_exit(void)
0594 {
0595     debugfs_remove_recursive(zs_stat_root);
0596 }
0597 
0598 static unsigned long zs_can_compact(struct size_class *class);
0599 
0600 static int zs_stats_size_show(struct seq_file *s, void *v)
0601 {
0602     int i;
0603     struct zs_pool *pool = s->private;
0604     struct size_class *class;
0605     int objs_per_zspage;
0606     unsigned long class_almost_full, class_almost_empty;
0607     unsigned long obj_allocated, obj_used, pages_used, freeable;
0608     unsigned long total_class_almost_full = 0, total_class_almost_empty = 0;
0609     unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0;
0610     unsigned long total_freeable = 0;
0611 
0612     seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n",
0613             "class", "size", "almost_full", "almost_empty",
0614             "obj_allocated", "obj_used", "pages_used",
0615             "pages_per_zspage", "freeable");
0616 
0617     for (i = 0; i < zs_size_classes; i++) {
0618         class = pool->size_class[i];
0619 
0620         if (class->index != i)
0621             continue;
0622 
0623         spin_lock(&class->lock);
0624         class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL);
0625         class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY);
0626         obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
0627         obj_used = zs_stat_get(class, OBJ_USED);
0628         freeable = zs_can_compact(class);
0629         spin_unlock(&class->lock);
0630 
0631         objs_per_zspage = class->objs_per_zspage;
0632         pages_used = obj_allocated / objs_per_zspage *
0633                 class->pages_per_zspage;
0634 
0635         seq_printf(s, " %5u %5u %11lu %12lu %13lu"
0636                 " %10lu %10lu %16d %8lu\n",
0637             i, class->size, class_almost_full, class_almost_empty,
0638             obj_allocated, obj_used, pages_used,
0639             class->pages_per_zspage, freeable);
0640 
0641         total_class_almost_full += class_almost_full;
0642         total_class_almost_empty += class_almost_empty;
0643         total_objs += obj_allocated;
0644         total_used_objs += obj_used;
0645         total_pages += pages_used;
0646         total_freeable += freeable;
0647     }
0648 
0649     seq_puts(s, "\n");
0650     seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n",
0651             "Total", "", total_class_almost_full,
0652             total_class_almost_empty, total_objs,
0653             total_used_objs, total_pages, "", total_freeable);
0654 
0655     return 0;
0656 }
0657 
0658 static int zs_stats_size_open(struct inode *inode, struct file *file)
0659 {
0660     return single_open(file, zs_stats_size_show, inode->i_private);
0661 }
0662 
0663 static const struct file_operations zs_stat_size_ops = {
0664     .open           = zs_stats_size_open,
0665     .read           = seq_read,
0666     .llseek         = seq_lseek,
0667     .release        = single_release,
0668 };
0669 
0670 static void zs_pool_stat_create(struct zs_pool *pool, const char *name)
0671 {
0672     struct dentry *entry;
0673 
0674     if (!zs_stat_root) {
0675         pr_warn("no root stat dir, not creating <%s> stat dir\n", name);
0676         return;
0677     }
0678 
0679     entry = debugfs_create_dir(name, zs_stat_root);
0680     if (!entry) {
0681         pr_warn("debugfs dir <%s> creation failed\n", name);
0682         return;
0683     }
0684     pool->stat_dentry = entry;
0685 
0686     entry = debugfs_create_file("classes", S_IFREG | S_IRUGO,
0687             pool->stat_dentry, pool, &zs_stat_size_ops);
0688     if (!entry) {
0689         pr_warn("%s: debugfs file entry <%s> creation failed\n",
0690                 name, "classes");
0691         debugfs_remove_recursive(pool->stat_dentry);
0692         pool->stat_dentry = NULL;
0693     }
0694 }
0695 
0696 static void zs_pool_stat_destroy(struct zs_pool *pool)
0697 {
0698     debugfs_remove_recursive(pool->stat_dentry);
0699 }
0700 
0701 #else /* CONFIG_ZSMALLOC_STAT */
0702 static void __init zs_stat_init(void)
0703 {
0704 }
0705 
0706 static void __exit zs_stat_exit(void)
0707 {
0708 }
0709 
0710 static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name)
0711 {
0712 }
0713 
0714 static inline void zs_pool_stat_destroy(struct zs_pool *pool)
0715 {
0716 }
0717 #endif
0718 
0719 
0720 /*
0721  * For each size class, zspages are divided into different groups
0722  * depending on how "full" they are. This was done so that we could
0723  * easily find empty or nearly empty zspages when we try to shrink
0724  * the pool (not yet implemented). This function returns fullness
0725  * status of the given page.
0726  */
0727 static enum fullness_group get_fullness_group(struct size_class *class,
0728                         struct zspage *zspage)
0729 {
0730     int inuse, objs_per_zspage;
0731     enum fullness_group fg;
0732 
0733     inuse = get_zspage_inuse(zspage);
0734     objs_per_zspage = class->objs_per_zspage;
0735 
0736     if (inuse == 0)
0737         fg = ZS_EMPTY;
0738     else if (inuse == objs_per_zspage)
0739         fg = ZS_FULL;
0740     else if (inuse <= 3 * objs_per_zspage / fullness_threshold_frac)
0741         fg = ZS_ALMOST_EMPTY;
0742     else
0743         fg = ZS_ALMOST_FULL;
0744 
0745     return fg;
0746 }
0747 
0748 /*
0749  * Each size class maintains various freelists and zspages are assigned
0750  * to one of these freelists based on the number of live objects they
0751  * have. This functions inserts the given zspage into the freelist
0752  * identified by <class, fullness_group>.
0753  */
0754 static void insert_zspage(struct size_class *class,
0755                 struct zspage *zspage,
0756                 enum fullness_group fullness)
0757 {
0758     struct zspage *head;
0759 
0760     zs_stat_inc(class, fullness, 1);
0761     head = list_first_entry_or_null(&class->fullness_list[fullness],
0762                     struct zspage, list);
0763     /*
0764      * We want to see more ZS_FULL pages and less almost empty/full.
0765      * Put pages with higher ->inuse first.
0766      */
0767     if (head) {
0768         if (get_zspage_inuse(zspage) < get_zspage_inuse(head)) {
0769             list_add(&zspage->list, &head->list);
0770             return;
0771         }
0772     }
0773     list_add(&zspage->list, &class->fullness_list[fullness]);
0774 }
0775 
0776 /*
0777  * This function removes the given zspage from the freelist identified
0778  * by <class, fullness_group>.
0779  */
0780 static void remove_zspage(struct size_class *class,
0781                 struct zspage *zspage,
0782                 enum fullness_group fullness)
0783 {
0784     VM_BUG_ON(list_empty(&class->fullness_list[fullness]));
0785     VM_BUG_ON(is_zspage_isolated(zspage));
0786 
0787     list_del_init(&zspage->list);
0788     zs_stat_dec(class, fullness, 1);
0789 }
0790 
0791 /*
0792  * Each size class maintains zspages in different fullness groups depending
0793  * on the number of live objects they contain. When allocating or freeing
0794  * objects, the fullness status of the page can change, say, from ALMOST_FULL
0795  * to ALMOST_EMPTY when freeing an object. This function checks if such
0796  * a status change has occurred for the given page and accordingly moves the
0797  * page from the freelist of the old fullness group to that of the new
0798  * fullness group.
0799  */
0800 static enum fullness_group fix_fullness_group(struct size_class *class,
0801                         struct zspage *zspage)
0802 {
0803     int class_idx;
0804     enum fullness_group currfg, newfg;
0805 
0806     get_zspage_mapping(zspage, &class_idx, &currfg);
0807     newfg = get_fullness_group(class, zspage);
0808     if (newfg == currfg)
0809         goto out;
0810 
0811     if (!is_zspage_isolated(zspage)) {
0812         remove_zspage(class, zspage, currfg);
0813         insert_zspage(class, zspage, newfg);
0814     }
0815 
0816     set_zspage_mapping(zspage, class_idx, newfg);
0817 
0818 out:
0819     return newfg;
0820 }
0821 
0822 /*
0823  * We have to decide on how many pages to link together
0824  * to form a zspage for each size class. This is important
0825  * to reduce wastage due to unusable space left at end of
0826  * each zspage which is given as:
0827  *     wastage = Zp % class_size
0828  *     usage = Zp - wastage
0829  * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
0830  *
0831  * For example, for size class of 3/8 * PAGE_SIZE, we should
0832  * link together 3 PAGE_SIZE sized pages to form a zspage
0833  * since then we can perfectly fit in 8 such objects.
0834  */
0835 static int get_pages_per_zspage(int class_size)
0836 {
0837     int i, max_usedpc = 0;
0838     /* zspage order which gives maximum used size per KB */
0839     int max_usedpc_order = 1;
0840 
0841     for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
0842         int zspage_size;
0843         int waste, usedpc;
0844 
0845         zspage_size = i * PAGE_SIZE;
0846         waste = zspage_size % class_size;
0847         usedpc = (zspage_size - waste) * 100 / zspage_size;
0848 
0849         if (usedpc > max_usedpc) {
0850             max_usedpc = usedpc;
0851             max_usedpc_order = i;
0852         }
0853     }
0854 
0855     return max_usedpc_order;
0856 }
0857 
0858 static struct zspage *get_zspage(struct page *page)
0859 {
0860     struct zspage *zspage = (struct zspage *)page->private;
0861 
0862     BUG_ON(zspage->magic != ZSPAGE_MAGIC);
0863     return zspage;
0864 }
0865 
0866 static struct page *get_next_page(struct page *page)
0867 {
0868     if (unlikely(PageHugeObject(page)))
0869         return NULL;
0870 
0871     return page->freelist;
0872 }
0873 
0874 /**
0875  * obj_to_location - get (<page>, <obj_idx>) from encoded object value
0876  * @page: page object resides in zspage
0877  * @obj_idx: object index
0878  */
0879 static void obj_to_location(unsigned long obj, struct page **page,
0880                 unsigned int *obj_idx)
0881 {
0882     obj >>= OBJ_TAG_BITS;
0883     *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
0884     *obj_idx = (obj & OBJ_INDEX_MASK);
0885 }
0886 
0887 /**
0888  * location_to_obj - get obj value encoded from (<page>, <obj_idx>)
0889  * @page: page object resides in zspage
0890  * @obj_idx: object index
0891  */
0892 static unsigned long location_to_obj(struct page *page, unsigned int obj_idx)
0893 {
0894     unsigned long obj;
0895 
0896     obj = page_to_pfn(page) << OBJ_INDEX_BITS;
0897     obj |= obj_idx & OBJ_INDEX_MASK;
0898     obj <<= OBJ_TAG_BITS;
0899 
0900     return obj;
0901 }
0902 
0903 static unsigned long handle_to_obj(unsigned long handle)
0904 {
0905     return *(unsigned long *)handle;
0906 }
0907 
0908 static unsigned long obj_to_head(struct page *page, void *obj)
0909 {
0910     if (unlikely(PageHugeObject(page))) {
0911         VM_BUG_ON_PAGE(!is_first_page(page), page);
0912         return page->index;
0913     } else
0914         return *(unsigned long *)obj;
0915 }
0916 
0917 static inline int testpin_tag(unsigned long handle)
0918 {
0919     return bit_spin_is_locked(HANDLE_PIN_BIT, (unsigned long *)handle);
0920 }
0921 
0922 static inline int trypin_tag(unsigned long handle)
0923 {
0924     return bit_spin_trylock(HANDLE_PIN_BIT, (unsigned long *)handle);
0925 }
0926 
0927 static void pin_tag(unsigned long handle)
0928 {
0929     bit_spin_lock(HANDLE_PIN_BIT, (unsigned long *)handle);
0930 }
0931 
0932 static void unpin_tag(unsigned long handle)
0933 {
0934     bit_spin_unlock(HANDLE_PIN_BIT, (unsigned long *)handle);
0935 }
0936 
0937 static void reset_page(struct page *page)
0938 {
0939     __ClearPageMovable(page);
0940     ClearPagePrivate(page);
0941     ClearPagePrivate2(page);
0942     set_page_private(page, 0);
0943     page_mapcount_reset(page);
0944     ClearPageHugeObject(page);
0945     page->freelist = NULL;
0946 }
0947 
0948 /*
0949  * To prevent zspage destroy during migration, zspage freeing should
0950  * hold locks of all pages in the zspage.
0951  */
0952 void lock_zspage(struct zspage *zspage)
0953 {
0954     struct page *page = get_first_page(zspage);
0955 
0956     do {
0957         lock_page(page);
0958     } while ((page = get_next_page(page)) != NULL);
0959 }
0960 
0961 int trylock_zspage(struct zspage *zspage)
0962 {
0963     struct page *cursor, *fail;
0964 
0965     for (cursor = get_first_page(zspage); cursor != NULL; cursor =
0966                     get_next_page(cursor)) {
0967         if (!trylock_page(cursor)) {
0968             fail = cursor;
0969             goto unlock;
0970         }
0971     }
0972 
0973     return 1;
0974 unlock:
0975     for (cursor = get_first_page(zspage); cursor != fail; cursor =
0976                     get_next_page(cursor))
0977         unlock_page(cursor);
0978 
0979     return 0;
0980 }
0981 
0982 static void __free_zspage(struct zs_pool *pool, struct size_class *class,
0983                 struct zspage *zspage)
0984 {
0985     struct page *page, *next;
0986     enum fullness_group fg;
0987     unsigned int class_idx;
0988 
0989     get_zspage_mapping(zspage, &class_idx, &fg);
0990 
0991     assert_spin_locked(&class->lock);
0992 
0993     VM_BUG_ON(get_zspage_inuse(zspage));
0994     VM_BUG_ON(fg != ZS_EMPTY);
0995 
0996     next = page = get_first_page(zspage);
0997     do {
0998         VM_BUG_ON_PAGE(!PageLocked(page), page);
0999         next = get_next_page(page);
1000         reset_page(page);
1001         unlock_page(page);
1002         dec_zone_page_state(page, NR_ZSPAGES);
1003         put_page(page);
1004         page = next;
1005     } while (page != NULL);
1006 
1007     cache_free_zspage(pool, zspage);
1008 
1009     zs_stat_dec(class, OBJ_ALLOCATED, class->objs_per_zspage);
1010     atomic_long_sub(class->pages_per_zspage,
1011                     &pool->pages_allocated);
1012 }
1013 
1014 static void free_zspage(struct zs_pool *pool, struct size_class *class,
1015                 struct zspage *zspage)
1016 {
1017     VM_BUG_ON(get_zspage_inuse(zspage));
1018     VM_BUG_ON(list_empty(&zspage->list));
1019 
1020     if (!trylock_zspage(zspage)) {
1021         kick_deferred_free(pool);
1022         return;
1023     }
1024 
1025     remove_zspage(class, zspage, ZS_EMPTY);
1026     __free_zspage(pool, class, zspage);
1027 }
1028 
1029 /* Initialize a newly allocated zspage */
1030 static void init_zspage(struct size_class *class, struct zspage *zspage)
1031 {
1032     unsigned int freeobj = 1;
1033     unsigned long off = 0;
1034     struct page *page = get_first_page(zspage);
1035 
1036     while (page) {
1037         struct page *next_page;
1038         struct link_free *link;
1039         void *vaddr;
1040 
1041         set_first_obj_offset(page, off);
1042 
1043         vaddr = kmap_atomic(page);
1044         link = (struct link_free *)vaddr + off / sizeof(*link);
1045 
1046         while ((off += class->size) < PAGE_SIZE) {
1047             link->next = freeobj++ << OBJ_TAG_BITS;
1048             link += class->size / sizeof(*link);
1049         }
1050 
1051         /*
1052          * We now come to the last (full or partial) object on this
1053          * page, which must point to the first object on the next
1054          * page (if present)
1055          */
1056         next_page = get_next_page(page);
1057         if (next_page) {
1058             link->next = freeobj++ << OBJ_TAG_BITS;
1059         } else {
1060             /*
1061              * Reset OBJ_TAG_BITS bit to last link to tell
1062              * whether it's allocated object or not.
1063              */
1064             link->next = -1 << OBJ_TAG_BITS;
1065         }
1066         kunmap_atomic(vaddr);
1067         page = next_page;
1068         off %= PAGE_SIZE;
1069     }
1070 
1071     set_freeobj(zspage, 0);
1072 }
1073 
1074 static void create_page_chain(struct size_class *class, struct zspage *zspage,
1075                 struct page *pages[])
1076 {
1077     int i;
1078     struct page *page;
1079     struct page *prev_page = NULL;
1080     int nr_pages = class->pages_per_zspage;
1081 
1082     /*
1083      * Allocate individual pages and link them together as:
1084      * 1. all pages are linked together using page->freelist
1085      * 2. each sub-page point to zspage using page->private
1086      *
1087      * we set PG_private to identify the first page (i.e. no other sub-page
1088      * has this flag set) and PG_private_2 to identify the last page.
1089      */
1090     for (i = 0; i < nr_pages; i++) {
1091         page = pages[i];
1092         set_page_private(page, (unsigned long)zspage);
1093         page->freelist = NULL;
1094         if (i == 0) {
1095             zspage->first_page = page;
1096             SetPagePrivate(page);
1097             if (unlikely(class->objs_per_zspage == 1 &&
1098                     class->pages_per_zspage == 1))
1099                 SetPageHugeObject(page);
1100         } else {
1101             prev_page->freelist = page;
1102         }
1103         if (i == nr_pages - 1)
1104             SetPagePrivate2(page);
1105         prev_page = page;
1106     }
1107 }
1108 
1109 /*
1110  * Allocate a zspage for the given size class
1111  */
1112 static struct zspage *alloc_zspage(struct zs_pool *pool,
1113                     struct size_class *class,
1114                     gfp_t gfp)
1115 {
1116     int i;
1117     struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE];
1118     struct zspage *zspage = cache_alloc_zspage(pool, gfp);
1119 
1120     if (!zspage)
1121         return NULL;
1122 
1123     memset(zspage, 0, sizeof(struct zspage));
1124     zspage->magic = ZSPAGE_MAGIC;
1125     migrate_lock_init(zspage);
1126 
1127     for (i = 0; i < class->pages_per_zspage; i++) {
1128         struct page *page;
1129 
1130         page = alloc_page(gfp);
1131         if (!page) {
1132             while (--i >= 0) {
1133                 dec_zone_page_state(pages[i], NR_ZSPAGES);
1134                 __free_page(pages[i]);
1135             }
1136             cache_free_zspage(pool, zspage);
1137             return NULL;
1138         }
1139 
1140         inc_zone_page_state(page, NR_ZSPAGES);
1141         pages[i] = page;
1142     }
1143 
1144     create_page_chain(class, zspage, pages);
1145     init_zspage(class, zspage);
1146 
1147     return zspage;
1148 }
1149 
1150 static struct zspage *find_get_zspage(struct size_class *class)
1151 {
1152     int i;
1153     struct zspage *zspage;
1154 
1155     for (i = ZS_ALMOST_FULL; i >= ZS_EMPTY; i--) {
1156         zspage = list_first_entry_or_null(&class->fullness_list[i],
1157                 struct zspage, list);
1158         if (zspage)
1159             break;
1160     }
1161 
1162     return zspage;
1163 }
1164 
1165 #ifdef CONFIG_PGTABLE_MAPPING
1166 static inline int __zs_cpu_up(struct mapping_area *area)
1167 {
1168     /*
1169      * Make sure we don't leak memory if a cpu UP notification
1170      * and zs_init() race and both call zs_cpu_up() on the same cpu
1171      */
1172     if (area->vm)
1173         return 0;
1174     area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL);
1175     if (!area->vm)
1176         return -ENOMEM;
1177     return 0;
1178 }
1179 
1180 static inline void __zs_cpu_down(struct mapping_area *area)
1181 {
1182     if (area->vm)
1183         free_vm_area(area->vm);
1184     area->vm = NULL;
1185 }
1186 
1187 static inline void *__zs_map_object(struct mapping_area *area,
1188                 struct page *pages[2], int off, int size)
1189 {
1190     BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages));
1191     area->vm_addr = area->vm->addr;
1192     return area->vm_addr + off;
1193 }
1194 
1195 static inline void __zs_unmap_object(struct mapping_area *area,
1196                 struct page *pages[2], int off, int size)
1197 {
1198     unsigned long addr = (unsigned long)area->vm_addr;
1199 
1200     unmap_kernel_range(addr, PAGE_SIZE * 2);
1201 }
1202 
1203 #else /* CONFIG_PGTABLE_MAPPING */
1204 
1205 static inline int __zs_cpu_up(struct mapping_area *area)
1206 {
1207     /*
1208      * Make sure we don't leak memory if a cpu UP notification
1209      * and zs_init() race and both call zs_cpu_up() on the same cpu
1210      */
1211     if (area->vm_buf)
1212         return 0;
1213     area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
1214     if (!area->vm_buf)
1215         return -ENOMEM;
1216     return 0;
1217 }
1218 
1219 static inline void __zs_cpu_down(struct mapping_area *area)
1220 {
1221     kfree(area->vm_buf);
1222     area->vm_buf = NULL;
1223 }
1224 
1225 static void *__zs_map_object(struct mapping_area *area,
1226             struct page *pages[2], int off, int size)
1227 {
1228     int sizes[2];
1229     void *addr;
1230     char *buf = area->vm_buf;
1231 
1232     /* disable page faults to match kmap_atomic() return conditions */
1233     pagefault_disable();
1234 
1235     /* no read fastpath */
1236     if (area->vm_mm == ZS_MM_WO)
1237         goto out;
1238 
1239     sizes[0] = PAGE_SIZE - off;
1240     sizes[1] = size - sizes[0];
1241 
1242     /* copy object to per-cpu buffer */
1243     addr = kmap_atomic(pages[0]);
1244     memcpy(buf, addr + off, sizes[0]);
1245     kunmap_atomic(addr);
1246     addr = kmap_atomic(pages[1]);
1247     memcpy(buf + sizes[0], addr, sizes[1]);
1248     kunmap_atomic(addr);
1249 out:
1250     return area->vm_buf;
1251 }
1252 
1253 static void __zs_unmap_object(struct mapping_area *area,
1254             struct page *pages[2], int off, int size)
1255 {
1256     int sizes[2];
1257     void *addr;
1258     char *buf;
1259 
1260     /* no write fastpath */
1261     if (area->vm_mm == ZS_MM_RO)
1262         goto out;
1263 
1264     buf = area->vm_buf;
1265     buf = buf + ZS_HANDLE_SIZE;
1266     size -= ZS_HANDLE_SIZE;
1267     off += ZS_HANDLE_SIZE;
1268 
1269     sizes[0] = PAGE_SIZE - off;
1270     sizes[1] = size - sizes[0];
1271 
1272     /* copy per-cpu buffer to object */
1273     addr = kmap_atomic(pages[0]);
1274     memcpy(addr + off, buf, sizes[0]);
1275     kunmap_atomic(addr);
1276     addr = kmap_atomic(pages[1]);
1277     memcpy(addr, buf + sizes[0], sizes[1]);
1278     kunmap_atomic(addr);
1279 
1280 out:
1281     /* enable page faults to match kunmap_atomic() return conditions */
1282     pagefault_enable();
1283 }
1284 
1285 #endif /* CONFIG_PGTABLE_MAPPING */
1286 
1287 static int zs_cpu_prepare(unsigned int cpu)
1288 {
1289     struct mapping_area *area;
1290 
1291     area = &per_cpu(zs_map_area, cpu);
1292     return __zs_cpu_up(area);
1293 }
1294 
1295 static int zs_cpu_dead(unsigned int cpu)
1296 {
1297     struct mapping_area *area;
1298 
1299     area = &per_cpu(zs_map_area, cpu);
1300     __zs_cpu_down(area);
1301     return 0;
1302 }
1303 
1304 static void __init init_zs_size_classes(void)
1305 {
1306     int nr;
1307 
1308     nr = (ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / ZS_SIZE_CLASS_DELTA + 1;
1309     if ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) % ZS_SIZE_CLASS_DELTA)
1310         nr += 1;
1311 
1312     zs_size_classes = nr;
1313 }
1314 
1315 static bool can_merge(struct size_class *prev, int pages_per_zspage,
1316                     int objs_per_zspage)
1317 {
1318     if (prev->pages_per_zspage == pages_per_zspage &&
1319         prev->objs_per_zspage == objs_per_zspage)
1320         return true;
1321 
1322     return false;
1323 }
1324 
1325 static bool zspage_full(struct size_class *class, struct zspage *zspage)
1326 {
1327     return get_zspage_inuse(zspage) == class->objs_per_zspage;
1328 }
1329 
1330 unsigned long zs_get_total_pages(struct zs_pool *pool)
1331 {
1332     return atomic_long_read(&pool->pages_allocated);
1333 }
1334 EXPORT_SYMBOL_GPL(zs_get_total_pages);
1335 
1336 /**
1337  * zs_map_object - get address of allocated object from handle.
1338  * @pool: pool from which the object was allocated
1339  * @handle: handle returned from zs_malloc
1340  *
1341  * Before using an object allocated from zs_malloc, it must be mapped using
1342  * this function. When done with the object, it must be unmapped using
1343  * zs_unmap_object.
1344  *
1345  * Only one object can be mapped per cpu at a time. There is no protection
1346  * against nested mappings.
1347  *
1348  * This function returns with preemption and page faults disabled.
1349  */
1350 void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1351             enum zs_mapmode mm)
1352 {
1353     struct zspage *zspage;
1354     struct page *page;
1355     unsigned long obj, off;
1356     unsigned int obj_idx;
1357 
1358     unsigned int class_idx;
1359     enum fullness_group fg;
1360     struct size_class *class;
1361     struct mapping_area *area;
1362     struct page *pages[2];
1363     void *ret;
1364 
1365     /*
1366      * Because we use per-cpu mapping areas shared among the
1367      * pools/users, we can't allow mapping in interrupt context
1368      * because it can corrupt another users mappings.
1369      */
1370     WARN_ON_ONCE(in_interrupt());
1371 
1372     /* From now on, migration cannot move the object */
1373     pin_tag(handle);
1374 
1375     obj = handle_to_obj(handle);
1376     obj_to_location(obj, &page, &obj_idx);
1377     zspage = get_zspage(page);
1378 
1379     /* migration cannot move any subpage in this zspage */
1380     migrate_read_lock(zspage);
1381 
1382     get_zspage_mapping(zspage, &class_idx, &fg);
1383     class = pool->size_class[class_idx];
1384     off = (class->size * obj_idx) & ~PAGE_MASK;
1385 
1386     area = &get_cpu_var(zs_map_area);
1387     area->vm_mm = mm;
1388     if (off + class->size <= PAGE_SIZE) {
1389         /* this object is contained entirely within a page */
1390         area->vm_addr = kmap_atomic(page);
1391         ret = area->vm_addr + off;
1392         goto out;
1393     }
1394 
1395     /* this object spans two pages */
1396     pages[0] = page;
1397     pages[1] = get_next_page(page);
1398     BUG_ON(!pages[1]);
1399 
1400     ret = __zs_map_object(area, pages, off, class->size);
1401 out:
1402     if (likely(!PageHugeObject(page)))
1403         ret += ZS_HANDLE_SIZE;
1404 
1405     return ret;
1406 }
1407 EXPORT_SYMBOL_GPL(zs_map_object);
1408 
1409 void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1410 {
1411     struct zspage *zspage;
1412     struct page *page;
1413     unsigned long obj, off;
1414     unsigned int obj_idx;
1415 
1416     unsigned int class_idx;
1417     enum fullness_group fg;
1418     struct size_class *class;
1419     struct mapping_area *area;
1420 
1421     obj = handle_to_obj(handle);
1422     obj_to_location(obj, &page, &obj_idx);
1423     zspage = get_zspage(page);
1424     get_zspage_mapping(zspage, &class_idx, &fg);
1425     class = pool->size_class[class_idx];
1426     off = (class->size * obj_idx) & ~PAGE_MASK;
1427 
1428     area = this_cpu_ptr(&zs_map_area);
1429     if (off + class->size <= PAGE_SIZE)
1430         kunmap_atomic(area->vm_addr);
1431     else {
1432         struct page *pages[2];
1433 
1434         pages[0] = page;
1435         pages[1] = get_next_page(page);
1436         BUG_ON(!pages[1]);
1437 
1438         __zs_unmap_object(area, pages, off, class->size);
1439     }
1440     put_cpu_var(zs_map_area);
1441 
1442     migrate_read_unlock(zspage);
1443     unpin_tag(handle);
1444 }
1445 EXPORT_SYMBOL_GPL(zs_unmap_object);
1446 
1447 static unsigned long obj_malloc(struct size_class *class,
1448                 struct zspage *zspage, unsigned long handle)
1449 {
1450     int i, nr_page, offset;
1451     unsigned long obj;
1452     struct link_free *link;
1453 
1454     struct page *m_page;
1455     unsigned long m_offset;
1456     void *vaddr;
1457 
1458     handle |= OBJ_ALLOCATED_TAG;
1459     obj = get_freeobj(zspage);
1460 
1461     offset = obj * class->size;
1462     nr_page = offset >> PAGE_SHIFT;
1463     m_offset = offset & ~PAGE_MASK;
1464     m_page = get_first_page(zspage);
1465 
1466     for (i = 0; i < nr_page; i++)
1467         m_page = get_next_page(m_page);
1468 
1469     vaddr = kmap_atomic(m_page);
1470     link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1471     set_freeobj(zspage, link->next >> OBJ_TAG_BITS);
1472     if (likely(!PageHugeObject(m_page)))
1473         /* record handle in the header of allocated chunk */
1474         link->handle = handle;
1475     else
1476         /* record handle to page->index */
1477         zspage->first_page->index = handle;
1478 
1479     kunmap_atomic(vaddr);
1480     mod_zspage_inuse(zspage, 1);
1481     zs_stat_inc(class, OBJ_USED, 1);
1482 
1483     obj = location_to_obj(m_page, obj);
1484 
1485     return obj;
1486 }
1487 
1488 
1489 /**
1490  * zs_malloc - Allocate block of given size from pool.
1491  * @pool: pool to allocate from
1492  * @size: size of block to allocate
1493  * @gfp: gfp flags when allocating object
1494  *
1495  * On success, handle to the allocated object is returned,
1496  * otherwise 0.
1497  * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1498  */
1499 unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp)
1500 {
1501     unsigned long handle, obj;
1502     struct size_class *class;
1503     enum fullness_group newfg;
1504     struct zspage *zspage;
1505 
1506     if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
1507         return 0;
1508 
1509     handle = cache_alloc_handle(pool, gfp);
1510     if (!handle)
1511         return 0;
1512 
1513     /* extra space in chunk to keep the handle */
1514     size += ZS_HANDLE_SIZE;
1515     class = pool->size_class[get_size_class_index(size)];
1516 
1517     spin_lock(&class->lock);
1518     zspage = find_get_zspage(class);
1519     if (likely(zspage)) {
1520         obj = obj_malloc(class, zspage, handle);
1521         /* Now move the zspage to another fullness group, if required */
1522         fix_fullness_group(class, zspage);
1523         record_obj(handle, obj);
1524         spin_unlock(&class->lock);
1525 
1526         return handle;
1527     }
1528 
1529     spin_unlock(&class->lock);
1530 
1531     zspage = alloc_zspage(pool, class, gfp);
1532     if (!zspage) {
1533         cache_free_handle(pool, handle);
1534         return 0;
1535     }
1536 
1537     spin_lock(&class->lock);
1538     obj = obj_malloc(class, zspage, handle);
1539     newfg = get_fullness_group(class, zspage);
1540     insert_zspage(class, zspage, newfg);
1541     set_zspage_mapping(zspage, class->index, newfg);
1542     record_obj(handle, obj);
1543     atomic_long_add(class->pages_per_zspage,
1544                 &pool->pages_allocated);
1545     zs_stat_inc(class, OBJ_ALLOCATED, class->objs_per_zspage);
1546 
1547     /* We completely set up zspage so mark them as movable */
1548     SetZsPageMovable(pool, zspage);
1549     spin_unlock(&class->lock);
1550 
1551     return handle;
1552 }
1553 EXPORT_SYMBOL_GPL(zs_malloc);
1554 
1555 static void obj_free(struct size_class *class, unsigned long obj)
1556 {
1557     struct link_free *link;
1558     struct zspage *zspage;
1559     struct page *f_page;
1560     unsigned long f_offset;
1561     unsigned int f_objidx;
1562     void *vaddr;
1563 
1564     obj &= ~OBJ_ALLOCATED_TAG;
1565     obj_to_location(obj, &f_page, &f_objidx);
1566     f_offset = (class->size * f_objidx) & ~PAGE_MASK;
1567     zspage = get_zspage(f_page);
1568 
1569     vaddr = kmap_atomic(f_page);
1570 
1571     /* Insert this object in containing zspage's freelist */
1572     link = (struct link_free *)(vaddr + f_offset);
1573     link->next = get_freeobj(zspage) << OBJ_TAG_BITS;
1574     kunmap_atomic(vaddr);
1575     set_freeobj(zspage, f_objidx);
1576     mod_zspage_inuse(zspage, -1);
1577     zs_stat_dec(class, OBJ_USED, 1);
1578 }
1579 
1580 void zs_free(struct zs_pool *pool, unsigned long handle)
1581 {
1582     struct zspage *zspage;
1583     struct page *f_page;
1584     unsigned long obj;
1585     unsigned int f_objidx;
1586     int class_idx;
1587     struct size_class *class;
1588     enum fullness_group fullness;
1589     bool isolated;
1590 
1591     if (unlikely(!handle))
1592         return;
1593 
1594     pin_tag(handle);
1595     obj = handle_to_obj(handle);
1596     obj_to_location(obj, &f_page, &f_objidx);
1597     zspage = get_zspage(f_page);
1598 
1599     migrate_read_lock(zspage);
1600 
1601     get_zspage_mapping(zspage, &class_idx, &fullness);
1602     class = pool->size_class[class_idx];
1603 
1604     spin_lock(&class->lock);
1605     obj_free(class, obj);
1606     fullness = fix_fullness_group(class, zspage);
1607     if (fullness != ZS_EMPTY) {
1608         migrate_read_unlock(zspage);
1609         goto out;
1610     }
1611 
1612     isolated = is_zspage_isolated(zspage);
1613     migrate_read_unlock(zspage);
1614     /* If zspage is isolated, zs_page_putback will free the zspage */
1615     if (likely(!isolated))
1616         free_zspage(pool, class, zspage);
1617 out:
1618 
1619     spin_unlock(&class->lock);
1620     unpin_tag(handle);
1621     cache_free_handle(pool, handle);
1622 }
1623 EXPORT_SYMBOL_GPL(zs_free);
1624 
1625 static void zs_object_copy(struct size_class *class, unsigned long dst,
1626                 unsigned long src)
1627 {
1628     struct page *s_page, *d_page;
1629     unsigned int s_objidx, d_objidx;
1630     unsigned long s_off, d_off;
1631     void *s_addr, *d_addr;
1632     int s_size, d_size, size;
1633     int written = 0;
1634 
1635     s_size = d_size = class->size;
1636 
1637     obj_to_location(src, &s_page, &s_objidx);
1638     obj_to_location(dst, &d_page, &d_objidx);
1639 
1640     s_off = (class->size * s_objidx) & ~PAGE_MASK;
1641     d_off = (class->size * d_objidx) & ~PAGE_MASK;
1642 
1643     if (s_off + class->size > PAGE_SIZE)
1644         s_size = PAGE_SIZE - s_off;
1645 
1646     if (d_off + class->size > PAGE_SIZE)
1647         d_size = PAGE_SIZE - d_off;
1648 
1649     s_addr = kmap_atomic(s_page);
1650     d_addr = kmap_atomic(d_page);
1651 
1652     while (1) {
1653         size = min(s_size, d_size);
1654         memcpy(d_addr + d_off, s_addr + s_off, size);
1655         written += size;
1656 
1657         if (written == class->size)
1658             break;
1659 
1660         s_off += size;
1661         s_size -= size;
1662         d_off += size;
1663         d_size -= size;
1664 
1665         if (s_off >= PAGE_SIZE) {
1666             kunmap_atomic(d_addr);
1667             kunmap_atomic(s_addr);
1668             s_page = get_next_page(s_page);
1669             s_addr = kmap_atomic(s_page);
1670             d_addr = kmap_atomic(d_page);
1671             s_size = class->size - written;
1672             s_off = 0;
1673         }
1674 
1675         if (d_off >= PAGE_SIZE) {
1676             kunmap_atomic(d_addr);
1677             d_page = get_next_page(d_page);
1678             d_addr = kmap_atomic(d_page);
1679             d_size = class->size - written;
1680             d_off = 0;
1681         }
1682     }
1683 
1684     kunmap_atomic(d_addr);
1685     kunmap_atomic(s_addr);
1686 }
1687 
1688 /*
1689  * Find alloced object in zspage from index object and
1690  * return handle.
1691  */
1692 static unsigned long find_alloced_obj(struct size_class *class,
1693                     struct page *page, int *obj_idx)
1694 {
1695     unsigned long head;
1696     int offset = 0;
1697     int index = *obj_idx;
1698     unsigned long handle = 0;
1699     void *addr = kmap_atomic(page);
1700 
1701     offset = get_first_obj_offset(page);
1702     offset += class->size * index;
1703 
1704     while (offset < PAGE_SIZE) {
1705         head = obj_to_head(page, addr + offset);
1706         if (head & OBJ_ALLOCATED_TAG) {
1707             handle = head & ~OBJ_ALLOCATED_TAG;
1708             if (trypin_tag(handle))
1709                 break;
1710             handle = 0;
1711         }
1712 
1713         offset += class->size;
1714         index++;
1715     }
1716 
1717     kunmap_atomic(addr);
1718 
1719     *obj_idx = index;
1720 
1721     return handle;
1722 }
1723 
1724 struct zs_compact_control {
1725     /* Source spage for migration which could be a subpage of zspage */
1726     struct page *s_page;
1727     /* Destination page for migration which should be a first page
1728      * of zspage. */
1729     struct page *d_page;
1730      /* Starting object index within @s_page which used for live object
1731       * in the subpage. */
1732     int obj_idx;
1733 };
1734 
1735 static int migrate_zspage(struct zs_pool *pool, struct size_class *class,
1736                 struct zs_compact_control *cc)
1737 {
1738     unsigned long used_obj, free_obj;
1739     unsigned long handle;
1740     struct page *s_page = cc->s_page;
1741     struct page *d_page = cc->d_page;
1742     int obj_idx = cc->obj_idx;
1743     int ret = 0;
1744 
1745     while (1) {
1746         handle = find_alloced_obj(class, s_page, &obj_idx);
1747         if (!handle) {
1748             s_page = get_next_page(s_page);
1749             if (!s_page)
1750                 break;
1751             obj_idx = 0;
1752             continue;
1753         }
1754 
1755         /* Stop if there is no more space */
1756         if (zspage_full(class, get_zspage(d_page))) {
1757             unpin_tag(handle);
1758             ret = -ENOMEM;
1759             break;
1760         }
1761 
1762         used_obj = handle_to_obj(handle);
1763         free_obj = obj_malloc(class, get_zspage(d_page), handle);
1764         zs_object_copy(class, free_obj, used_obj);
1765         obj_idx++;
1766         /*
1767          * record_obj updates handle's value to free_obj and it will
1768          * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which
1769          * breaks synchronization using pin_tag(e,g, zs_free) so
1770          * let's keep the lock bit.
1771          */
1772         free_obj |= BIT(HANDLE_PIN_BIT);
1773         record_obj(handle, free_obj);
1774         unpin_tag(handle);
1775         obj_free(class, used_obj);
1776     }
1777 
1778     /* Remember last position in this iteration */
1779     cc->s_page = s_page;
1780     cc->obj_idx = obj_idx;
1781 
1782     return ret;
1783 }
1784 
1785 static struct zspage *isolate_zspage(struct size_class *class, bool source)
1786 {
1787     int i;
1788     struct zspage *zspage;
1789     enum fullness_group fg[2] = {ZS_ALMOST_EMPTY, ZS_ALMOST_FULL};
1790 
1791     if (!source) {
1792         fg[0] = ZS_ALMOST_FULL;
1793         fg[1] = ZS_ALMOST_EMPTY;
1794     }
1795 
1796     for (i = 0; i < 2; i++) {
1797         zspage = list_first_entry_or_null(&class->fullness_list[fg[i]],
1798                             struct zspage, list);
1799         if (zspage) {
1800             VM_BUG_ON(is_zspage_isolated(zspage));
1801             remove_zspage(class, zspage, fg[i]);
1802             return zspage;
1803         }
1804     }
1805 
1806     return zspage;
1807 }
1808 
1809 /*
1810  * putback_zspage - add @zspage into right class's fullness list
1811  * @class: destination class
1812  * @zspage: target page
1813  *
1814  * Return @zspage's fullness_group
1815  */
1816 static enum fullness_group putback_zspage(struct size_class *class,
1817             struct zspage *zspage)
1818 {
1819     enum fullness_group fullness;
1820 
1821     VM_BUG_ON(is_zspage_isolated(zspage));
1822 
1823     fullness = get_fullness_group(class, zspage);
1824     insert_zspage(class, zspage, fullness);
1825     set_zspage_mapping(zspage, class->index, fullness);
1826 
1827     return fullness;
1828 }
1829 
1830 #ifdef CONFIG_COMPACTION
1831 static struct dentry *zs_mount(struct file_system_type *fs_type,
1832                 int flags, const char *dev_name, void *data)
1833 {
1834     static const struct dentry_operations ops = {
1835         .d_dname = simple_dname,
1836     };
1837 
1838     return mount_pseudo(fs_type, "zsmalloc:", NULL, &ops, ZSMALLOC_MAGIC);
1839 }
1840 
1841 static struct file_system_type zsmalloc_fs = {
1842     .name       = "zsmalloc",
1843     .mount      = zs_mount,
1844     .kill_sb    = kill_anon_super,
1845 };
1846 
1847 static int zsmalloc_mount(void)
1848 {
1849     int ret = 0;
1850 
1851     zsmalloc_mnt = kern_mount(&zsmalloc_fs);
1852     if (IS_ERR(zsmalloc_mnt))
1853         ret = PTR_ERR(zsmalloc_mnt);
1854 
1855     return ret;
1856 }
1857 
1858 static void zsmalloc_unmount(void)
1859 {
1860     kern_unmount(zsmalloc_mnt);
1861 }
1862 
1863 static void migrate_lock_init(struct zspage *zspage)
1864 {
1865     rwlock_init(&zspage->lock);
1866 }
1867 
1868 static void migrate_read_lock(struct zspage *zspage)
1869 {
1870     read_lock(&zspage->lock);
1871 }
1872 
1873 static void migrate_read_unlock(struct zspage *zspage)
1874 {
1875     read_unlock(&zspage->lock);
1876 }
1877 
1878 static void migrate_write_lock(struct zspage *zspage)
1879 {
1880     write_lock(&zspage->lock);
1881 }
1882 
1883 static void migrate_write_unlock(struct zspage *zspage)
1884 {
1885     write_unlock(&zspage->lock);
1886 }
1887 
1888 /* Number of isolated subpage for *page migration* in this zspage */
1889 static void inc_zspage_isolation(struct zspage *zspage)
1890 {
1891     zspage->isolated++;
1892 }
1893 
1894 static void dec_zspage_isolation(struct zspage *zspage)
1895 {
1896     zspage->isolated--;
1897 }
1898 
1899 static void replace_sub_page(struct size_class *class, struct zspage *zspage,
1900                 struct page *newpage, struct page *oldpage)
1901 {
1902     struct page *page;
1903     struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE] = {NULL, };
1904     int idx = 0;
1905 
1906     page = get_first_page(zspage);
1907     do {
1908         if (page == oldpage)
1909             pages[idx] = newpage;
1910         else
1911             pages[idx] = page;
1912         idx++;
1913     } while ((page = get_next_page(page)) != NULL);
1914 
1915     create_page_chain(class, zspage, pages);
1916     set_first_obj_offset(newpage, get_first_obj_offset(oldpage));
1917     if (unlikely(PageHugeObject(oldpage)))
1918         newpage->index = oldpage->index;
1919     __SetPageMovable(newpage, page_mapping(oldpage));
1920 }
1921 
1922 bool zs_page_isolate(struct page *page, isolate_mode_t mode)
1923 {
1924     struct zs_pool *pool;
1925     struct size_class *class;
1926     int class_idx;
1927     enum fullness_group fullness;
1928     struct zspage *zspage;
1929     struct address_space *mapping;
1930 
1931     /*
1932      * Page is locked so zspage couldn't be destroyed. For detail, look at
1933      * lock_zspage in free_zspage.
1934      */
1935     VM_BUG_ON_PAGE(!PageMovable(page), page);
1936     VM_BUG_ON_PAGE(PageIsolated(page), page);
1937 
1938     zspage = get_zspage(page);
1939 
1940     /*
1941      * Without class lock, fullness could be stale while class_idx is okay
1942      * because class_idx is constant unless page is freed so we should get
1943      * fullness again under class lock.
1944      */
1945     get_zspage_mapping(zspage, &class_idx, &fullness);
1946     mapping = page_mapping(page);
1947     pool = mapping->private_data;
1948     class = pool->size_class[class_idx];
1949 
1950     spin_lock(&class->lock);
1951     if (get_zspage_inuse(zspage) == 0) {
1952         spin_unlock(&class->lock);
1953         return false;
1954     }
1955 
1956     /* zspage is isolated for object migration */
1957     if (list_empty(&zspage->list) && !is_zspage_isolated(zspage)) {
1958         spin_unlock(&class->lock);
1959         return false;
1960     }
1961 
1962     /*
1963      * If this is first time isolation for the zspage, isolate zspage from
1964      * size_class to prevent further object allocation from the zspage.
1965      */
1966     if (!list_empty(&zspage->list) && !is_zspage_isolated(zspage)) {
1967         get_zspage_mapping(zspage, &class_idx, &fullness);
1968         remove_zspage(class, zspage, fullness);
1969     }
1970 
1971     inc_zspage_isolation(zspage);
1972     spin_unlock(&class->lock);
1973 
1974     return true;
1975 }
1976 
1977 int zs_page_migrate(struct address_space *mapping, struct page *newpage,
1978         struct page *page, enum migrate_mode mode)
1979 {
1980     struct zs_pool *pool;
1981     struct size_class *class;
1982     int class_idx;
1983     enum fullness_group fullness;
1984     struct zspage *zspage;
1985     struct page *dummy;
1986     void *s_addr, *d_addr, *addr;
1987     int offset, pos;
1988     unsigned long handle, head;
1989     unsigned long old_obj, new_obj;
1990     unsigned int obj_idx;
1991     int ret = -EAGAIN;
1992 
1993     VM_BUG_ON_PAGE(!PageMovable(page), page);
1994     VM_BUG_ON_PAGE(!PageIsolated(page), page);
1995 
1996     zspage = get_zspage(page);
1997 
1998     /* Concurrent compactor cannot migrate any subpage in zspage */
1999     migrate_write_lock(zspage);
2000     get_zspage_mapping(zspage, &class_idx, &fullness);
2001     pool = mapping->private_data;
2002     class = pool->size_class[class_idx];
2003     offset = get_first_obj_offset(page);
2004 
2005     spin_lock(&class->lock);
2006     if (!get_zspage_inuse(zspage)) {
2007         ret = -EBUSY;
2008         goto unlock_class;
2009     }
2010 
2011     pos = offset;
2012     s_addr = kmap_atomic(page);
2013     while (pos < PAGE_SIZE) {
2014         head = obj_to_head(page, s_addr + pos);
2015         if (head & OBJ_ALLOCATED_TAG) {
2016             handle = head & ~OBJ_ALLOCATED_TAG;
2017             if (!trypin_tag(handle))
2018                 goto unpin_objects;
2019         }
2020         pos += class->size;
2021     }
2022 
2023     /*
2024      * Here, any user cannot access all objects in the zspage so let's move.
2025      */
2026     d_addr = kmap_atomic(newpage);
2027     memcpy(d_addr, s_addr, PAGE_SIZE);
2028     kunmap_atomic(d_addr);
2029 
2030     for (addr = s_addr + offset; addr < s_addr + pos;
2031                     addr += class->size) {
2032         head = obj_to_head(page, addr);
2033         if (head & OBJ_ALLOCATED_TAG) {
2034             handle = head & ~OBJ_ALLOCATED_TAG;
2035             if (!testpin_tag(handle))
2036                 BUG();
2037 
2038             old_obj = handle_to_obj(handle);
2039             obj_to_location(old_obj, &dummy, &obj_idx);
2040             new_obj = (unsigned long)location_to_obj(newpage,
2041                                 obj_idx);
2042             new_obj |= BIT(HANDLE_PIN_BIT);
2043             record_obj(handle, new_obj);
2044         }
2045     }
2046 
2047     replace_sub_page(class, zspage, newpage, page);
2048     get_page(newpage);
2049 
2050     dec_zspage_isolation(zspage);
2051 
2052     /*
2053      * Page migration is done so let's putback isolated zspage to
2054      * the list if @page is final isolated subpage in the zspage.
2055      */
2056     if (!is_zspage_isolated(zspage))
2057         putback_zspage(class, zspage);
2058 
2059     reset_page(page);
2060     put_page(page);
2061     page = newpage;
2062 
2063     ret = MIGRATEPAGE_SUCCESS;
2064 unpin_objects:
2065     for (addr = s_addr + offset; addr < s_addr + pos;
2066                         addr += class->size) {
2067         head = obj_to_head(page, addr);
2068         if (head & OBJ_ALLOCATED_TAG) {
2069             handle = head & ~OBJ_ALLOCATED_TAG;
2070             if (!testpin_tag(handle))
2071                 BUG();
2072             unpin_tag(handle);
2073         }
2074     }
2075     kunmap_atomic(s_addr);
2076 unlock_class:
2077     spin_unlock(&class->lock);
2078     migrate_write_unlock(zspage);
2079 
2080     return ret;
2081 }
2082 
2083 void zs_page_putback(struct page *page)
2084 {
2085     struct zs_pool *pool;
2086     struct size_class *class;
2087     int class_idx;
2088     enum fullness_group fg;
2089     struct address_space *mapping;
2090     struct zspage *zspage;
2091 
2092     VM_BUG_ON_PAGE(!PageMovable(page), page);
2093     VM_BUG_ON_PAGE(!PageIsolated(page), page);
2094 
2095     zspage = get_zspage(page);
2096     get_zspage_mapping(zspage, &class_idx, &fg);
2097     mapping = page_mapping(page);
2098     pool = mapping->private_data;
2099     class = pool->size_class[class_idx];
2100 
2101     spin_lock(&class->lock);
2102     dec_zspage_isolation(zspage);
2103     if (!is_zspage_isolated(zspage)) {
2104         fg = putback_zspage(class, zspage);
2105         /*
2106          * Due to page_lock, we cannot free zspage immediately
2107          * so let's defer.
2108          */
2109         if (fg == ZS_EMPTY)
2110             schedule_work(&pool->free_work);
2111     }
2112     spin_unlock(&class->lock);
2113 }
2114 
2115 const struct address_space_operations zsmalloc_aops = {
2116     .isolate_page = zs_page_isolate,
2117     .migratepage = zs_page_migrate,
2118     .putback_page = zs_page_putback,
2119 };
2120 
2121 static int zs_register_migration(struct zs_pool *pool)
2122 {
2123     pool->inode = alloc_anon_inode(zsmalloc_mnt->mnt_sb);
2124     if (IS_ERR(pool->inode)) {
2125         pool->inode = NULL;
2126         return 1;
2127     }
2128 
2129     pool->inode->i_mapping->private_data = pool;
2130     pool->inode->i_mapping->a_ops = &zsmalloc_aops;
2131     return 0;
2132 }
2133 
2134 static void zs_unregister_migration(struct zs_pool *pool)
2135 {
2136     flush_work(&pool->free_work);
2137     iput(pool->inode);
2138 }
2139 
2140 /*
2141  * Caller should hold page_lock of all pages in the zspage
2142  * In here, we cannot use zspage meta data.
2143  */
2144 static void async_free_zspage(struct work_struct *work)
2145 {
2146     int i;
2147     struct size_class *class;
2148     unsigned int class_idx;
2149     enum fullness_group fullness;
2150     struct zspage *zspage, *tmp;
2151     LIST_HEAD(free_pages);
2152     struct zs_pool *pool = container_of(work, struct zs_pool,
2153                     free_work);
2154 
2155     for (i = 0; i < zs_size_classes; i++) {
2156         class = pool->size_class[i];
2157         if (class->index != i)
2158             continue;
2159 
2160         spin_lock(&class->lock);
2161         list_splice_init(&class->fullness_list[ZS_EMPTY], &free_pages);
2162         spin_unlock(&class->lock);
2163     }
2164 
2165 
2166     list_for_each_entry_safe(zspage, tmp, &free_pages, list) {
2167         list_del(&zspage->list);
2168         lock_zspage(zspage);
2169 
2170         get_zspage_mapping(zspage, &class_idx, &fullness);
2171         VM_BUG_ON(fullness != ZS_EMPTY);
2172         class = pool->size_class[class_idx];
2173         spin_lock(&class->lock);
2174         __free_zspage(pool, pool->size_class[class_idx], zspage);
2175         spin_unlock(&class->lock);
2176     }
2177 };
2178 
2179 static void kick_deferred_free(struct zs_pool *pool)
2180 {
2181     schedule_work(&pool->free_work);
2182 }
2183 
2184 static void init_deferred_free(struct zs_pool *pool)
2185 {
2186     INIT_WORK(&pool->free_work, async_free_zspage);
2187 }
2188 
2189 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage)
2190 {
2191     struct page *page = get_first_page(zspage);
2192 
2193     do {
2194         WARN_ON(!trylock_page(page));
2195         __SetPageMovable(page, pool->inode->i_mapping);
2196         unlock_page(page);
2197     } while ((page = get_next_page(page)) != NULL);
2198 }
2199 #endif
2200 
2201 /*
2202  *
2203  * Based on the number of unused allocated objects calculate
2204  * and return the number of pages that we can free.
2205  */
2206 static unsigned long zs_can_compact(struct size_class *class)
2207 {
2208     unsigned long obj_wasted;
2209     unsigned long obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
2210     unsigned long obj_used = zs_stat_get(class, OBJ_USED);
2211 
2212     if (obj_allocated <= obj_used)
2213         return 0;
2214 
2215     obj_wasted = obj_allocated - obj_used;
2216     obj_wasted /= class->objs_per_zspage;
2217 
2218     return obj_wasted * class->pages_per_zspage;
2219 }
2220 
2221 static void __zs_compact(struct zs_pool *pool, struct size_class *class)
2222 {
2223     struct zs_compact_control cc;
2224     struct zspage *src_zspage;
2225     struct zspage *dst_zspage = NULL;
2226 
2227     spin_lock(&class->lock);
2228     while ((src_zspage = isolate_zspage(class, true))) {
2229 
2230         if (!zs_can_compact(class))
2231             break;
2232 
2233         cc.obj_idx = 0;
2234         cc.s_page = get_first_page(src_zspage);
2235 
2236         while ((dst_zspage = isolate_zspage(class, false))) {
2237             cc.d_page = get_first_page(dst_zspage);
2238             /*
2239              * If there is no more space in dst_page, resched
2240              * and see if anyone had allocated another zspage.
2241              */
2242             if (!migrate_zspage(pool, class, &cc))
2243                 break;
2244 
2245             putback_zspage(class, dst_zspage);
2246         }
2247 
2248         /* Stop if we couldn't find slot */
2249         if (dst_zspage == NULL)
2250             break;
2251 
2252         putback_zspage(class, dst_zspage);
2253         if (putback_zspage(class, src_zspage) == ZS_EMPTY) {
2254             free_zspage(pool, class, src_zspage);
2255             pool->stats.pages_compacted += class->pages_per_zspage;
2256         }
2257         spin_unlock(&class->lock);
2258         cond_resched();
2259         spin_lock(&class->lock);
2260     }
2261 
2262     if (src_zspage)
2263         putback_zspage(class, src_zspage);
2264 
2265     spin_unlock(&class->lock);
2266 }
2267 
2268 unsigned long zs_compact(struct zs_pool *pool)
2269 {
2270     int i;
2271     struct size_class *class;
2272 
2273     for (i = zs_size_classes - 1; i >= 0; i--) {
2274         class = pool->size_class[i];
2275         if (!class)
2276             continue;
2277         if (class->index != i)
2278             continue;
2279         __zs_compact(pool, class);
2280     }
2281 
2282     return pool->stats.pages_compacted;
2283 }
2284 EXPORT_SYMBOL_GPL(zs_compact);
2285 
2286 void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats)
2287 {
2288     memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats));
2289 }
2290 EXPORT_SYMBOL_GPL(zs_pool_stats);
2291 
2292 static unsigned long zs_shrinker_scan(struct shrinker *shrinker,
2293         struct shrink_control *sc)
2294 {
2295     unsigned long pages_freed;
2296     struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2297             shrinker);
2298 
2299     pages_freed = pool->stats.pages_compacted;
2300     /*
2301      * Compact classes and calculate compaction delta.
2302      * Can run concurrently with a manually triggered
2303      * (by user) compaction.
2304      */
2305     pages_freed = zs_compact(pool) - pages_freed;
2306 
2307     return pages_freed ? pages_freed : SHRINK_STOP;
2308 }
2309 
2310 static unsigned long zs_shrinker_count(struct shrinker *shrinker,
2311         struct shrink_control *sc)
2312 {
2313     int i;
2314     struct size_class *class;
2315     unsigned long pages_to_free = 0;
2316     struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2317             shrinker);
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 
2326         pages_to_free += zs_can_compact(class);
2327     }
2328 
2329     return pages_to_free;
2330 }
2331 
2332 static void zs_unregister_shrinker(struct zs_pool *pool)
2333 {
2334     if (pool->shrinker_enabled) {
2335         unregister_shrinker(&pool->shrinker);
2336         pool->shrinker_enabled = false;
2337     }
2338 }
2339 
2340 static int zs_register_shrinker(struct zs_pool *pool)
2341 {
2342     pool->shrinker.scan_objects = zs_shrinker_scan;
2343     pool->shrinker.count_objects = zs_shrinker_count;
2344     pool->shrinker.batch = 0;
2345     pool->shrinker.seeks = DEFAULT_SEEKS;
2346 
2347     return register_shrinker(&pool->shrinker);
2348 }
2349 
2350 /**
2351  * zs_create_pool - Creates an allocation pool to work from.
2352  * @name: pool name to be created
2353  *
2354  * This function must be called before anything when using
2355  * the zsmalloc allocator.
2356  *
2357  * On success, a pointer to the newly created pool is returned,
2358  * otherwise NULL.
2359  */
2360 struct zs_pool *zs_create_pool(const char *name)
2361 {
2362     int i;
2363     struct zs_pool *pool;
2364     struct size_class *prev_class = NULL;
2365 
2366     pool = kzalloc(sizeof(*pool), GFP_KERNEL);
2367     if (!pool)
2368         return NULL;
2369 
2370     init_deferred_free(pool);
2371     pool->size_class = kcalloc(zs_size_classes, sizeof(struct size_class *),
2372             GFP_KERNEL);
2373     if (!pool->size_class) {
2374         kfree(pool);
2375         return NULL;
2376     }
2377 
2378     pool->name = kstrdup(name, GFP_KERNEL);
2379     if (!pool->name)
2380         goto err;
2381 
2382     if (create_cache(pool))
2383         goto err;
2384 
2385     /*
2386      * Iterate reversly, because, size of size_class that we want to use
2387      * for merging should be larger or equal to current size.
2388      */
2389     for (i = zs_size_classes - 1; i >= 0; i--) {
2390         int size;
2391         int pages_per_zspage;
2392         int objs_per_zspage;
2393         struct size_class *class;
2394         int fullness = 0;
2395 
2396         size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
2397         if (size > ZS_MAX_ALLOC_SIZE)
2398             size = ZS_MAX_ALLOC_SIZE;
2399         pages_per_zspage = get_pages_per_zspage(size);
2400         objs_per_zspage = pages_per_zspage * PAGE_SIZE / size;
2401 
2402         /*
2403          * size_class is used for normal zsmalloc operation such
2404          * as alloc/free for that size. Although it is natural that we
2405          * have one size_class for each size, there is a chance that we
2406          * can get more memory utilization if we use one size_class for
2407          * many different sizes whose size_class have same
2408          * characteristics. So, we makes size_class point to
2409          * previous size_class if possible.
2410          */
2411         if (prev_class) {
2412             if (can_merge(prev_class, pages_per_zspage, objs_per_zspage)) {
2413                 pool->size_class[i] = prev_class;
2414                 continue;
2415             }
2416         }
2417 
2418         class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
2419         if (!class)
2420             goto err;
2421 
2422         class->size = size;
2423         class->index = i;
2424         class->pages_per_zspage = pages_per_zspage;
2425         class->objs_per_zspage = objs_per_zspage;
2426         spin_lock_init(&class->lock);
2427         pool->size_class[i] = class;
2428         for (fullness = ZS_EMPTY; fullness < NR_ZS_FULLNESS;
2429                             fullness++)
2430             INIT_LIST_HEAD(&class->fullness_list[fullness]);
2431 
2432         prev_class = class;
2433     }
2434 
2435     /* debug only, don't abort if it fails */
2436     zs_pool_stat_create(pool, name);
2437 
2438     if (zs_register_migration(pool))
2439         goto err;
2440 
2441     /*
2442      * Not critical, we still can use the pool
2443      * and user can trigger compaction manually.
2444      */
2445     if (zs_register_shrinker(pool) == 0)
2446         pool->shrinker_enabled = true;
2447     return pool;
2448 
2449 err:
2450     zs_destroy_pool(pool);
2451     return NULL;
2452 }
2453 EXPORT_SYMBOL_GPL(zs_create_pool);
2454 
2455 void zs_destroy_pool(struct zs_pool *pool)
2456 {
2457     int i;
2458 
2459     zs_unregister_shrinker(pool);
2460     zs_unregister_migration(pool);
2461     zs_pool_stat_destroy(pool);
2462 
2463     for (i = 0; i < zs_size_classes; i++) {
2464         int fg;
2465         struct size_class *class = pool->size_class[i];
2466 
2467         if (!class)
2468             continue;
2469 
2470         if (class->index != i)
2471             continue;
2472 
2473         for (fg = ZS_EMPTY; fg < NR_ZS_FULLNESS; fg++) {
2474             if (!list_empty(&class->fullness_list[fg])) {
2475                 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
2476                     class->size, fg);
2477             }
2478         }
2479         kfree(class);
2480     }
2481 
2482     destroy_cache(pool);
2483     kfree(pool->size_class);
2484     kfree(pool->name);
2485     kfree(pool);
2486 }
2487 EXPORT_SYMBOL_GPL(zs_destroy_pool);
2488 
2489 static int __init zs_init(void)
2490 {
2491     int ret;
2492 
2493     ret = zsmalloc_mount();
2494     if (ret)
2495         goto out;
2496 
2497     ret = cpuhp_setup_state(CPUHP_MM_ZS_PREPARE, "mm/zsmalloc:prepare",
2498                 zs_cpu_prepare, zs_cpu_dead);
2499     if (ret)
2500         goto hp_setup_fail;
2501 
2502     init_zs_size_classes();
2503 
2504 #ifdef CONFIG_ZPOOL
2505     zpool_register_driver(&zs_zpool_driver);
2506 #endif
2507 
2508     zs_stat_init();
2509 
2510     return 0;
2511 
2512 hp_setup_fail:
2513     zsmalloc_unmount();
2514 out:
2515     return ret;
2516 }
2517 
2518 static void __exit zs_exit(void)
2519 {
2520 #ifdef CONFIG_ZPOOL
2521     zpool_unregister_driver(&zs_zpool_driver);
2522 #endif
2523     zsmalloc_unmount();
2524     cpuhp_remove_state(CPUHP_MM_ZS_PREPARE);
2525 
2526     zs_stat_exit();
2527 }
2528 
2529 module_init(zs_init);
2530 module_exit(zs_exit);
2531 
2532 MODULE_LICENSE("Dual BSD/GPL");
2533 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");