mm/zsmalloc.c

   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 <asm/tlbflush.h>
  43 #include <asm/pgtable.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/migrate.h>
  56 #include <linux/pagemap.h>
  57 #include <linux/fs.h>
  58
  59 #define ZSPAGE_MAGIC    0x58
  60
  61 /*
  62  * This must be power of 2 and greater than of equal to sizeof(link_free).
  63  * These two conditions ensure that any 'struct link_free' itself doesn't
  64  * span more than 1 page which avoids complex case of mapping 2 pages simply
  65  * to restore link_free pointer values.
  66  */
  67 #define ZS_ALIGN                8
  68
  69 /*
  70  * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
  71  * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
  72  */
  73 #define ZS_MAX_ZSPAGE_ORDER 2
  74 #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
  75
  76 #define ZS_HANDLE_SIZE (sizeof(unsigned long))
  77
  78 /*
  79  * Object location (<PFN>, <obj_idx>) is encoded as
  80  * as single (unsigned long) handle value.
  81  *
  82  * Note that object index <obj_idx> starts from 0.
  83  *
  84  * This is made more complicated by various memory models and PAE.
  85  */
  86
  87 #ifndef MAX_PHYSMEM_BITS
  88 #ifdef CONFIG_HIGHMEM64G
  89 #define MAX_PHYSMEM_BITS 36
  90 #else /* !CONFIG_HIGHMEM64G */
  91 /*
  92  * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
  93  * be PAGE_SHIFT
  94  */
  95 #define MAX_PHYSMEM_BITS BITS_PER_LONG
  96 #endif
  97 #endif
  98 #define _PFN_BITS               (MAX_PHYSMEM_BITS - PAGE_SHIFT)
  99
 100 /*
 101  * Memory for allocating for handle keeps object position by
 102  * encoding <page, obj_idx> and the encoded value has a room
 103  * in least bit(ie, look at obj_to_location).
 104  * We use the bit to synchronize between object access by
 105  * user and migration.
 106  */
 107 #define HANDLE_PIN_BIT  0
 108
 109 /*
 110  * Head in allocated object should have OBJ_ALLOCATED_TAG
 111  * to identify the object was allocated or not.
 112  * It's okay to add the status bit in the least bit because
 113  * header keeps handle which is 4byte-aligned address so we
 114  * have room for two bit at least.
 115  */
 116 #define OBJ_ALLOCATED_TAG 1
 117 #define OBJ_TAG_BITS 1
 118 #define OBJ_INDEX_BITS  (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
 119 #define OBJ_INDEX_MASK  ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
 120
 121 #define FULLNESS_BITS   2
 122 #define CLASS_BITS      8
 123 #define ISOLATED_BITS   3
 124 #define MAGIC_VAL_BITS  8
 125
 126 #define MAX(a, b) ((a) >= (b) ? (a) : (b))
 127 /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
 128 #define ZS_MIN_ALLOC_SIZE \
 129         MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
 130 /* each chunk includes extra space to keep handle */
 131 #define ZS_MAX_ALLOC_SIZE       PAGE_SIZE
 132
 133 /*
 134  * On systems with 4K page size, this gives 255 size classes! There is a
 135  * trader-off here:
 136  *  - Large number of size classes is potentially wasteful as free page are
 137  *    spread across these classes
 138  *  - Small number of size classes causes large internal fragmentation
 139  *  - Probably its better to use specific size classes (empirically
 140  *    determined). NOTE: all those class sizes must be set as multiple of
 141  *    ZS_ALIGN to make sure link_free itself never has to span 2 pages.
 142  *
 143  *  ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
 144  *  (reason above)
 145  */
 146 #define ZS_SIZE_CLASS_DELTA     (PAGE_SIZE >> CLASS_BITS)
 147 #define ZS_SIZE_CLASSES (DIV_ROUND_UP(ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE, \
 148                                       ZS_SIZE_CLASS_DELTA) + 1)
 149
 150 enum fullness_group {
 151         ZS_EMPTY,
 152         ZS_ALMOST_EMPTY,
 153         ZS_ALMOST_FULL,
 154         ZS_FULL,
 155         NR_ZS_FULLNESS,
 156 };
 157
 158 enum zs_stat_type {
 159         CLASS_EMPTY,
 160         CLASS_ALMOST_EMPTY,
 161         CLASS_ALMOST_FULL,
 162         CLASS_FULL,
 163         OBJ_ALLOCATED,
 164         OBJ_USED,
 165         NR_ZS_STAT_TYPE,
 166 };
 167
 168 struct zs_size_stat {
 169         unsigned long objs[NR_ZS_STAT_TYPE];
 170 };
 171
 172 #ifdef CONFIG_ZSMALLOC_STAT
 173 static struct dentry *zs_stat_root;
 174 #endif
 175
 176 #ifdef CONFIG_COMPACTION
 177 static struct vfsmount *zsmalloc_mnt;
 178 #endif
 179
 180 /*
 181  * We assign a page to ZS_ALMOST_EMPTY fullness group when:
 182  *      n <= N / f, where
 183  * n = number of allocated objects
 184  * N = total number of objects zspage can store
 185  * f = fullness_threshold_frac
 186  *
 187  * Similarly, we assign zspage to:
 188  *      ZS_ALMOST_FULL  when n > N / f
 189  *      ZS_EMPTY        when n == 0
 190  *      ZS_FULL         when n == N
 191  *
 192  * (see: fix_fullness_group())
 193  */
 194 static const int fullness_threshold_frac = 4;
 195
 196 struct size_class {
 197         spinlock_t lock;
 198         struct list_head fullness_list[NR_ZS_FULLNESS];
 199         /*
 200          * Size of objects stored in this class. Must be multiple
 201          * of ZS_ALIGN.
 202          */
 203         int size;
 204         int objs_per_zspage;
 205         /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
 206         int pages_per_zspage;
 207
 208         unsigned int index;
 209         struct zs_size_stat stats;
 210 };
 211
 212 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
 213 static void SetPageHugeObject(struct page *page)
 214 {
 215         SetPageOwnerPriv1(page);
 216 }
 217
 218 static void ClearPageHugeObject(struct page *page)
 219 {
 220         ClearPageOwnerPriv1(page);
 221 }
 222
 223 static int PageHugeObject(struct page *page)
 224 {
 225         return PageOwnerPriv1(page);
 226 }
 227
 228 /*
 229  * Placed within free objects to form a singly linked list.
 230  * For every zspage, zspage->freeobj gives head of this list.
 231  *
 232  * This must be power of 2 and less than or equal to ZS_ALIGN
 233  */
 234 struct link_free {
 235         union {
 236                 /*
 237                  * Free object index;
 238                  * It's valid for non-allocated object
 239                  */
 240                 unsigned long next;
 241                 /*
 242                  * Handle of allocated object.
 243                  */
 244                 unsigned long handle;
 245         };
 246 };
 247
 248 struct zs_pool {
 249         const char *name;
 250
 251         struct size_class *size_class[ZS_SIZE_CLASSES];
 252         struct kmem_cache *handle_cachep;
 253         struct kmem_cache *zspage_cachep;
 254
 255         atomic_long_t pages_allocated;
 256
 257         struct zs_pool_stats stats;
 258
 259         /* Compact classes */
 260         struct shrinker shrinker;
 261
 262 #ifdef CONFIG_ZSMALLOC_STAT
 263         struct dentry *stat_dentry;
 264 #endif
 265 #ifdef CONFIG_COMPACTION
 266         struct inode *inode;
 267         struct work_struct free_work;
 268 #endif
 269 };
 270
 271 struct zspage {
 272         struct {
 273                 unsigned int fullness:FULLNESS_BITS;
 274                 unsigned int class:CLASS_BITS + 1;
 275                 unsigned int isolated:ISOLATED_BITS;
 276                 unsigned int magic:MAGIC_VAL_BITS;
 277         };
 278         unsigned int inuse;
 279         unsigned int freeobj;
 280         struct page *first_page;
 281         struct list_head list; /* fullness list */
 282 #ifdef CONFIG_COMPACTION
 283         rwlock_t lock;
 284 #endif
 285 };
 286
 287 struct mapping_area {
 288 #ifdef CONFIG_PGTABLE_MAPPING
 289         struct vm_struct *vm; /* vm area for mapping object that span pages */
 290 #else
 291         char *vm_buf; /* copy buffer for objects that span pages */
 292 #endif
 293         char *vm_addr; /* address of kmap_atomic()'ed pages */
 294         enum zs_mapmode vm_mm; /* mapping mode */
 295 };
 296
 297 #ifdef CONFIG_COMPACTION
 298 static int zs_register_migration(struct zs_pool *pool);
 299 static void zs_unregister_migration(struct zs_pool *pool);
 300 static void migrate_lock_init(struct zspage *zspage);
 301 static void migrate_read_lock(struct zspage *zspage);
 302 static void migrate_read_unlock(struct zspage *zspage);
 303 static void kick_deferred_free(struct zs_pool *pool);
 304 static void init_deferred_free(struct zs_pool *pool);
 305 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage);
 306 #else
 307 static int zsmalloc_mount(void) { return 0; }
 308 static void zsmalloc_unmount(void) {}
 309 static int zs_register_migration(struct zs_pool *pool) { return 0; }
 310 static void zs_unregister_migration(struct zs_pool *pool) {}
 311 static void migrate_lock_init(struct zspage *zspage) {}
 312 static void migrate_read_lock(struct zspage *zspage) {}
 313 static void migrate_read_unlock(struct zspage *zspage) {}
 314 static void kick_deferred_free(struct zs_pool *pool) {}
 315 static void init_deferred_free(struct zs_pool *pool) {}
 316 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage) {}
 317 #endif
 318
 319 static int create_cache(struct zs_pool *pool)
 320 {
 321         pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE,
 322                                         0, 0, NULL);
 323         if (!pool->handle_cachep)
 324                 return 1;
 325
 326         pool->zspage_cachep = kmem_cache_create("zspage", sizeof(struct zspage),
 327                                         0, 0, NULL);
 328         if (!pool->zspage_cachep) {
 329                 kmem_cache_destroy(pool->handle_cachep);
 330                 pool->handle_cachep = NULL;
 331                 return 1;
 332         }
 333
 334         return 0;
 335 }
 336
 337 static void destroy_cache(struct zs_pool *pool)
 338 {
 339         kmem_cache_destroy(pool->handle_cachep);
 340         kmem_cache_destroy(pool->zspage_cachep);
 341 }
 342
 343 static unsigned long cache_alloc_handle(struct zs_pool *pool, gfp_t gfp)
 344 {
 345         return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
 346                         gfp & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
 347 }
 348
 349 static void cache_free_handle(struct zs_pool *pool, unsigned long handle)
 350 {
 351         kmem_cache_free(pool->handle_cachep, (void *)handle);
 352 }
 353
 354 static struct zspage *cache_alloc_zspage(struct zs_pool *pool, gfp_t flags)
 355 {
 356         return kmem_cache_alloc(pool->zspage_cachep,
 357                         flags & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
 358 }
 359
 360 static void cache_free_zspage(struct zs_pool *pool, struct zspage *zspage)
 361 {
 362         kmem_cache_free(pool->zspage_cachep, zspage);
 363 }
 364
 365 static void record_obj(unsigned long handle, unsigned long obj)
 366 {
 367         /*
 368          * lsb of @obj represents handle lock while other bits
 369          * represent object value the handle is pointing so
 370          * updating shouldn't do store tearing.
 371          */
 372         WRITE_ONCE(*(unsigned long *)handle, obj);
 373 }
 374
 375 /* zpool driver */
 376
 377 #ifdef CONFIG_ZPOOL
 378
 379 static void *zs_zpool_create(const char *name, gfp_t gfp,
 380                              const struct zpool_ops *zpool_ops,
 381                              struct zpool *zpool)
 382 {
 383         /*
 384          * Ignore global gfp flags: zs_malloc() may be invoked from
 385          * different contexts and its caller must provide a valid
 386          * gfp mask.
 387          */
 388         return zs_create_pool(name);
 389 }
 390
 391 static void zs_zpool_destroy(void *pool)
 392 {
 393         zs_destroy_pool(pool);
 394 }
 395
 396 static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
 397                         unsigned long *handle)
 398 {
 399         *handle = zs_malloc(pool, size, gfp);
 400         return *handle ? 0 : -1;
 401 }
 402 static void zs_zpool_free(void *pool, unsigned long handle)
 403 {
 404         zs_free(pool, handle);
 405 }
 406
 407 static void *zs_zpool_map(void *pool, unsigned long handle,
 408                         enum zpool_mapmode mm)
 409 {
 410         enum zs_mapmode zs_mm;
 411
 412         switch (mm) {
 413         case ZPOOL_MM_RO:
 414                 zs_mm = ZS_MM_RO;
 415                 break;
 416         case ZPOOL_MM_WO:
 417                 zs_mm = ZS_MM_WO;
 418                 break;
 419         case ZPOOL_MM_RW: /* fallthru */
 420         default:
 421                 zs_mm = ZS_MM_RW;
 422                 break;
 423         }
 424
 425         return zs_map_object(pool, handle, zs_mm);
 426 }
 427 static void zs_zpool_unmap(void *pool, unsigned long handle)
 428 {
 429         zs_unmap_object(pool, handle);
 430 }
 431
 432 static u64 zs_zpool_total_size(void *pool)
 433 {
 434         return zs_get_total_pages(pool) << PAGE_SHIFT;
 435 }
 436
 437 static struct zpool_driver zs_zpool_driver = {
 438         .type =         "zsmalloc",
 439         .owner =        THIS_MODULE,
 440         .create =       zs_zpool_create,
 441         .destroy =      zs_zpool_destroy,
 442         .malloc =       zs_zpool_malloc,
 443         .free =         zs_zpool_free,
 444         .map =          zs_zpool_map,
 445         .unmap =        zs_zpool_unmap,
 446         .total_size =   zs_zpool_total_size,
 447 };
 448
 449 MODULE_ALIAS("zpool-zsmalloc");
 450 #endif /* CONFIG_ZPOOL */
 451
 452 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
 453 static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
 454
 455 static bool is_zspage_isolated(struct zspage *zspage)
 456 {
 457         return zspage->isolated;
 458 }
 459
 460 static __maybe_unused int is_first_page(struct page *page)
 461 {
 462         return PagePrivate(page);
 463 }
 464
 465 /* Protected by class->lock */
 466 static inline int get_zspage_inuse(struct zspage *zspage)
 467 {
 468         return zspage->inuse;
 469 }
 470
 471 static inline void set_zspage_inuse(struct zspage *zspage, int val)
 472 {
 473         zspage->inuse = val;
 474 }
 475
 476 static inline void mod_zspage_inuse(struct zspage *zspage, int val)
 477 {
 478         zspage->inuse += val;
 479 }
 480
 481 static inline struct page *get_first_page(struct zspage *zspage)
 482 {
 483         struct page *first_page = zspage->first_page;
 484
 485         VM_BUG_ON_PAGE(!is_first_page(first_page), first_page);
 486         return first_page;
 487 }
 488
 489 static inline int get_first_obj_offset(struct page *page)
 490 {
 491         return page->units;
 492 }
 493
 494 static inline void set_first_obj_offset(struct page *page, int offset)
 495 {
 496         page->units = offset;
 497 }
 498
 499 static inline unsigned int get_freeobj(struct zspage *zspage)
 500 {
 501         return zspage->freeobj;
 502 }
 503
 504 static inline void set_freeobj(struct zspage *zspage, unsigned int obj)
 505 {
 506         zspage->freeobj = obj;
 507 }
 508
 509 static void get_zspage_mapping(struct zspage *zspage,
 510                                 unsigned int *class_idx,
 511                                 enum fullness_group *fullness)
 512 {
 513         BUG_ON(zspage->magic != ZSPAGE_MAGIC);
 514
 515         *fullness = zspage->fullness;
 516         *class_idx = zspage->class;
 517 }
 518
 519 static void set_zspage_mapping(struct zspage *zspage,
 520                                 unsigned int class_idx,
 521                                 enum fullness_group fullness)
 522 {
 523         zspage->class = class_idx;
 524         zspage->fullness = fullness;
 525 }
 526
 527 /*
 528  * zsmalloc divides the pool into various size classes where each
 529  * class maintains a list of zspages where each zspage is divided
 530  * into equal sized chunks. Each allocation falls into one of these
 531  * classes depending on its size. This function returns index of the
 532  * size class which has chunk size big enough to hold the give size.
 533  */
 534 static int get_size_class_index(int size)
 535 {
 536         int idx = 0;
 537
 538         if (likely(size > ZS_MIN_ALLOC_SIZE))
 539                 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
 540                                 ZS_SIZE_CLASS_DELTA);
 541
 542         return min_t(int, ZS_SIZE_CLASSES - 1, idx);
 543 }
 544
 545 /* type can be of enum type zs_stat_type or fullness_group */
 546 static inline void zs_stat_inc(struct size_class *class,
 547                                 int type, unsigned long cnt)
 548 {
 549         class->stats.objs[type] += cnt;
 550 }
 551
 552 /* type can be of enum type zs_stat_type or fullness_group */
 553 static inline void zs_stat_dec(struct size_class *class,
 554                                 int type, unsigned long cnt)
 555 {
 556         class->stats.objs[type] -= cnt;
 557 }
 558
 559 /* type can be of enum type zs_stat_type or fullness_group */
 560 static inline unsigned long zs_stat_get(struct size_class *class,
 561                                 int type)
 562 {
 563         return class->stats.objs[type];
 564 }
 565
 566 #ifdef CONFIG_ZSMALLOC_STAT
 567
 568 static void __init zs_stat_init(void)
 569 {
 570         if (!debugfs_initialized()) {
 571                 pr_warn("debugfs not available, stat dir not created\n");
 572                 return;
 573         }
 574
 575         zs_stat_root = debugfs_create_dir("zsmalloc", NULL);
 576         if (!zs_stat_root)
 577                 pr_warn("debugfs 'zsmalloc' stat dir creation failed\n");
 578 }
 579
 580 static void __exit zs_stat_exit(void)
 581 {
 582         debugfs_remove_recursive(zs_stat_root);
 583 }
 584
 585 static unsigned long zs_can_compact(struct size_class *class);
 586
 587 static int zs_stats_size_show(struct seq_file *s, void *v)
 588 {
 589         int i;
 590         struct zs_pool *pool = s->private;
 591         struct size_class *class;
 592         int objs_per_zspage;
 593         unsigned long class_almost_full, class_almost_empty;
 594         unsigned long obj_allocated, obj_used, pages_used, freeable;
 595         unsigned long total_class_almost_full = 0, total_class_almost_empty = 0;
 596         unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0;
 597         unsigned long total_freeable = 0;
 598
 599         seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n",
 600                         "class", "size", "almost_full", "almost_empty",
 601                         "obj_allocated", "obj_used", "pages_used",
 602                         "pages_per_zspage", "freeable");
 603
 604         for (i = 0; i < ZS_SIZE_CLASSES; i++) {
 605                 class = pool->size_class[i];
 606
 607                 if (class->index != i)
 608                         continue;
 609
 610                 spin_lock(&class->lock);
 611                 class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL);
 612                 class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY);
 613                 obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
 614                 obj_used = zs_stat_get(class, OBJ_USED);
 615                 freeable = zs_can_compact(class);
 616                 spin_unlock(&class->lock);
 617
 618                 objs_per_zspage = class->objs_per_zspage;
 619                 pages_used = obj_allocated / objs_per_zspage *
 620                                 class->pages_per_zspage;
 621
 622                 seq_printf(s, " %5u %5u %11lu %12lu %13lu"
 623                                 " %10lu %10lu %16d %8lu\n",
 624                         i, class->size, class_almost_full, class_almost_empty,
 625                         obj_allocated, obj_used, pages_used,
 626                         class->pages_per_zspage, freeable);
 627
 628                 total_class_almost_full += class_almost_full;
 629                 total_class_almost_empty += class_almost_empty;
 630                 total_objs += obj_allocated;
 631                 total_used_objs += obj_used;
 632                 total_pages += pages_used;
 633                 total_freeable += freeable;
 634         }
 635
 636         seq_puts(s, "\n");
 637         seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n",
 638                         "Total", "", total_class_almost_full,
 639                         total_class_almost_empty, total_objs,
 640                         total_used_objs, total_pages, "", total_freeable);
 641
 642         return 0;
 643 }
 644
 645 static int zs_stats_size_open(struct inode *inode, struct file *file)
 646 {
 647         return single_open(file, zs_stats_size_show, inode->i_private);
 648 }
 649
 650 static const struct file_operations zs_stat_size_ops = {
 651         .open           = zs_stats_size_open,
 652         .read           = seq_read,
 653         .llseek         = seq_lseek,
 654         .release        = single_release,
 655 };
 656
 657 static void zs_pool_stat_create(struct zs_pool *pool, const char *name)
 658 {
 659         struct dentry *entry;
 660
 661         if (!zs_stat_root) {
 662                 pr_warn("no root stat dir, not creating <%s> stat dir\n", name);
 663                 return;
 664         }
 665
 666         entry = debugfs_create_dir(name, zs_stat_root);
 667         if (!entry) {
 668                 pr_warn("debugfs dir <%s> creation failed\n", name);
 669                 return;
 670         }
 671         pool->stat_dentry = entry;
 672
 673         entry = debugfs_create_file("classes", S_IFREG | S_IRUGO,
 674                         pool->stat_dentry, pool, &zs_stat_size_ops);
 675         if (!entry) {
 676                 pr_warn("%s: debugfs file entry <%s> creation failed\n",
 677                                 name, "classes");
 678                 debugfs_remove_recursive(pool->stat_dentry);
 679                 pool->stat_dentry = NULL;
 680         }
 681 }
 682
 683 static void zs_pool_stat_destroy(struct zs_pool *pool)
 684 {
 685         debugfs_remove_recursive(pool->stat_dentry);
 686 }
 687
 688 #else /* CONFIG_ZSMALLOC_STAT */
 689 static void __init zs_stat_init(void)
 690 {
 691 }
 692
 693 static void __exit zs_stat_exit(void)
 694 {
 695 }
 696
 697 static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name)
 698 {
 699 }
 700
 701 static inline void zs_pool_stat_destroy(struct zs_pool *pool)
 702 {
 703 }
 704 #endif
 705
 706
 707 /*
 708  * For each size class, zspages are divided into different groups
 709  * depending on how "full" they are. This was done so that we could
 710  * easily find empty or nearly empty zspages when we try to shrink
 711  * the pool (not yet implemented). This function returns fullness
 712  * status of the given page.
 713  */
 714 static enum fullness_group get_fullness_group(struct size_class *class,
 715                                                 struct zspage *zspage)
 716 {
 717         int inuse, objs_per_zspage;
 718         enum fullness_group fg;
 719
 720         inuse = get_zspage_inuse(zspage);
 721         objs_per_zspage = class->objs_per_zspage;
 722
 723         if (inuse == 0)
 724                 fg = ZS_EMPTY;
 725         else if (inuse == objs_per_zspage)
 726                 fg = ZS_FULL;
 727         else if (inuse <= 3 * objs_per_zspage / fullness_threshold_frac)
 728                 fg = ZS_ALMOST_EMPTY;
 729         else
 730                 fg = ZS_ALMOST_FULL;
 731
 732         return fg;
 733 }
 734
 735 /*
 736  * Each size class maintains various freelists and zspages are assigned
 737  * to one of these freelists based on the number of live objects they
 738  * have. This functions inserts the given zspage into the freelist
 739  * identified by <class, fullness_group>.
 740  */
 741 static void insert_zspage(struct size_class *class,
 742                                 struct zspage *zspage,
 743                                 enum fullness_group fullness)
 744 {
 745         struct zspage *head;
 746
 747         zs_stat_inc(class, fullness, 1);
 748         head = list_first_entry_or_null(&class->fullness_list[fullness],
 749                                         struct zspage, list);
 750         /*
 751          * We want to see more ZS_FULL pages and less almost empty/full.
 752          * Put pages with higher ->inuse first.
 753          */
 754         if (head) {
 755                 if (get_zspage_inuse(zspage) < get_zspage_inuse(head)) {
 756                         list_add(&zspage->list, &head->list);
 757                         return;
 758                 }
 759         }
 760         list_add(&zspage->list, &class->fullness_list[fullness]);
 761 }
 762
 763 /*
 764  * This function removes the given zspage from the freelist identified
 765  * by <class, fullness_group>.
 766  */
 767 static void remove_zspage(struct size_class *class,
 768                                 struct zspage *zspage,
 769                                 enum fullness_group fullness)
 770 {
 771         VM_BUG_ON(list_empty(&class->fullness_list[fullness]));
 772         VM_BUG_ON(is_zspage_isolated(zspage));
 773
 774         list_del_init(&zspage->list);
 775         zs_stat_dec(class, fullness, 1);
 776 }
 777
 778 /*
 779  * Each size class maintains zspages in different fullness groups depending
 780  * on the number of live objects they contain. When allocating or freeing
 781  * objects, the fullness status of the page can change, say, from ALMOST_FULL
 782  * to ALMOST_EMPTY when freeing an object. This function checks if such
 783  * a status change has occurred for the given page and accordingly moves the
 784  * page from the freelist of the old fullness group to that of the new
 785  * fullness group.
 786  */
 787 static enum fullness_group fix_fullness_group(struct size_class *class,
 788                                                 struct zspage *zspage)
 789 {
 790         int class_idx;
 791         enum fullness_group currfg, newfg;
 792
 793         get_zspage_mapping(zspage, &class_idx, &currfg);
 794         newfg = get_fullness_group(class, zspage);
 795         if (newfg == currfg)
 796                 goto out;
 797
 798         if (!is_zspage_isolated(zspage)) {
 799                 remove_zspage(class, zspage, currfg);
 800                 insert_zspage(class, zspage, newfg);
 801         }
 802
 803         set_zspage_mapping(zspage, class_idx, newfg);
 804
 805 out:
 806         return newfg;
 807 }
 808
 809 /*
 810  * We have to decide on how many pages to link together
 811  * to form a zspage for each size class. This is important
 812  * to reduce wastage due to unusable space left at end of
 813  * each zspage which is given as:
 814  *     wastage = Zp % class_size
 815  *     usage = Zp - wastage
 816  * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
 817  *
 818  * For example, for size class of 3/8 * PAGE_SIZE, we should
 819  * link together 3 PAGE_SIZE sized pages to form a zspage
 820  * since then we can perfectly fit in 8 such objects.
 821  */
 822 static int get_pages_per_zspage(int class_size)
 823 {
 824         int i, max_usedpc = 0;
 825         /* zspage order which gives maximum used size per KB */
 826         int max_usedpc_order = 1;
 827
 828         for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
 829                 int zspage_size;
 830                 int waste, usedpc;
 831
 832                 zspage_size = i * PAGE_SIZE;
 833                 waste = zspage_size % class_size;
 834                 usedpc = (zspage_size - waste) * 100 / zspage_size;
 835
 836                 if (usedpc > max_usedpc) {
 837                         max_usedpc = usedpc;
 838                         max_usedpc_order = i;
 839                 }
 840         }
 841
 842         return max_usedpc_order;
 843 }
 844
 845 static struct zspage *get_zspage(struct page *page)
 846 {
 847         struct zspage *zspage = (struct zspage *)page->private;
 848
 849         BUG_ON(zspage->magic != ZSPAGE_MAGIC);
 850         return zspage;
 851 }
 852
 853 static struct page *get_next_page(struct page *page)
 854 {
 855         if (unlikely(PageHugeObject(page)))
 856                 return NULL;
 857
 858         return page->freelist;
 859 }
 860
 861 /**
 862  * obj_to_location - get (<page>, <obj_idx>) from encoded object value
 863  * @page: page object resides in zspage
 864  * @obj_idx: object index
 865  */
 866 static void obj_to_location(unsigned long obj, struct page **page,
 867                                 unsigned int *obj_idx)
 868 {
 869         obj >>= OBJ_TAG_BITS;
 870         *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
 871         *obj_idx = (obj & OBJ_INDEX_MASK);
 872 }
 873
 874 /**
 875  * location_to_obj - get obj value encoded from (<page>, <obj_idx>)
 876  * @page: page object resides in zspage
 877  * @obj_idx: object index
 878  */
 879 static unsigned long location_to_obj(struct page *page, unsigned int obj_idx)
 880 {
 881         unsigned long obj;
 882
 883         obj = page_to_pfn(page) << OBJ_INDEX_BITS;
 884         obj |= obj_idx & OBJ_INDEX_MASK;
 885         obj <<= OBJ_TAG_BITS;
 886
 887         return obj;
 888 }
 889
 890 static unsigned long handle_to_obj(unsigned long handle)
 891 {
 892         return *(unsigned long *)handle;
 893 }
 894
 895 static unsigned long obj_to_head(struct page *page, void *obj)
 896 {
 897         if (unlikely(PageHugeObject(page))) {
 898                 VM_BUG_ON_PAGE(!is_first_page(page), page);
 899                 return page->index;
 900         } else
 901                 return *(unsigned long *)obj;
 902 }
 903
 904 static inline int testpin_tag(unsigned long handle)
 905 {
 906         return bit_spin_is_locked(HANDLE_PIN_BIT, (unsigned long *)handle);
 907 }
 908
 909 static inline int trypin_tag(unsigned long handle)
 910 {
 911         return bit_spin_trylock(HANDLE_PIN_BIT, (unsigned long *)handle);
 912 }
 913
 914 static void pin_tag(unsigned long handle)
 915 {
 916         bit_spin_lock(HANDLE_PIN_BIT, (unsigned long *)handle);
 917 }
 918
 919 static void unpin_tag(unsigned long handle)
 920 {
 921         bit_spin_unlock(HANDLE_PIN_BIT, (unsigned long *)handle);
 922 }
 923
 924 static void reset_page(struct page *page)
 925 {
 926         __ClearPageMovable(page);
 927         ClearPagePrivate(page);
 928         set_page_private(page, 0);
 929         page_mapcount_reset(page);
 930         ClearPageHugeObject(page);
 931         page->freelist = NULL;
 932 }
 933
 934 /*
 935  * To prevent zspage destroy during migration, zspage freeing should
 936  * hold locks of all pages in the zspage.
 937  */
 938 void lock_zspage(struct zspage *zspage)
 939 {
 940         struct page *page = get_first_page(zspage);
 941
 942         do {
 943                 lock_page(page);
 944         } while ((page = get_next_page(page)) != NULL);
 945 }
 946
 947 int trylock_zspage(struct zspage *zspage)
 948 {
 949         struct page *cursor, *fail;
 950
 951         for (cursor = get_first_page(zspage); cursor != NULL; cursor =
 952                                         get_next_page(cursor)) {
 953                 if (!trylock_page(cursor)) {
 954                         fail = cursor;
 955                         goto unlock;
 956                 }
 957         }
 958
 959         return 1;
 960 unlock:
 961         for (cursor = get_first_page(zspage); cursor != fail; cursor =
 962                                         get_next_page(cursor))
 963                 unlock_page(cursor);
 964
 965         return 0;
 966 }
 967
 968 static void __free_zspage(struct zs_pool *pool, struct size_class *class,
 969                                 struct zspage *zspage)
 970 {
 971         struct page *page, *next;
 972         enum fullness_group fg;
 973         unsigned int class_idx;
 974
 975         get_zspage_mapping(zspage, &class_idx, &fg);
 976
 977         assert_spin_locked(&class->lock);
 978
 979         VM_BUG_ON(get_zspage_inuse(zspage));
 980         VM_BUG_ON(fg != ZS_EMPTY);
 981
 982         next = page = get_first_page(zspage);
 983         do {
 984                 VM_BUG_ON_PAGE(!PageLocked(page), page);
 985                 next = get_next_page(page);
 986                 reset_page(page);
 987                 unlock_page(page);
 988                 dec_zone_page_state(page, NR_ZSPAGES);
 989                 put_page(page);
 990                 page = next;
 991         } while (page != NULL);
 992
 993         cache_free_zspage(pool, zspage);
 994
 995         zs_stat_dec(class, OBJ_ALLOCATED, class->objs_per_zspage);
 996         atomic_long_sub(class->pages_per_zspage,
 997                                         &pool->pages_allocated);
 998 }
 999
1000 static void free_zspage(struct zs_pool *pool, struct size_class *class,
1001                                 struct zspage *zspage)
1002 {
1003         VM_BUG_ON(get_zspage_inuse(zspage));
1004         VM_BUG_ON(list_empty(&zspage->list));
1005
1006         if (!trylock_zspage(zspage)) {
1007                 kick_deferred_free(pool);
1008                 return;
1009         }
1010
1011         remove_zspage(class, zspage, ZS_EMPTY);
1012         __free_zspage(pool, class, zspage);
1013 }
1014
1015 /* Initialize a newly allocated zspage */
1016 static void init_zspage(struct size_class *class, struct zspage *zspage)
1017 {
1018         unsigned int freeobj = 1;
1019         unsigned long off = 0;
1020         struct page *page = get_first_page(zspage);
1021
1022         while (page) {
1023                 struct page *next_page;
1024                 struct link_free *link;
1025                 void *vaddr;
1026
1027                 set_first_obj_offset(page, off);
1028
1029                 vaddr = kmap_atomic(page);
1030                 link = (struct link_free *)vaddr + off / sizeof(*link);
1031
1032                 while ((off += class->size) < PAGE_SIZE) {
1033                         link->next = freeobj++ << OBJ_TAG_BITS;
1034                         link += class->size / sizeof(*link);
1035                 }
1036
1037                 /*
1038                  * We now come to the last (full or partial) object on this
1039                  * page, which must point to the first object on the next
1040                  * page (if present)
1041                  */
1042                 next_page = get_next_page(page);
1043                 if (next_page) {
1044                         link->next = freeobj++ << OBJ_TAG_BITS;
1045                 } else {
1046                         /*
1047                          * Reset OBJ_TAG_BITS bit to last link to tell
1048                          * whether it's allocated object or not.
1049                          */
1050                         link->next = -1UL << OBJ_TAG_BITS;
1051                 }
1052                 kunmap_atomic(vaddr);
1053                 page = next_page;
1054                 off %= PAGE_SIZE;
1055         }
1056
1057         set_freeobj(zspage, 0);
1058 }
1059
1060 static void create_page_chain(struct size_class *class, struct zspage *zspage,
1061                                 struct page *pages[])
1062 {
1063         int i;
1064         struct page *page;
1065         struct page *prev_page = NULL;
1066         int nr_pages = class->pages_per_zspage;
1067
1068         /*
1069          * Allocate individual pages and link them together as:
1070          * 1. all pages are linked together using page->freelist
1071          * 2. each sub-page point to zspage using page->private
1072          *
1073          * we set PG_private to identify the first page (i.e. no other sub-page
1074          * has this flag set).
1075          */
1076         for (i = 0; i < nr_pages; i++) {
1077                 page = pages[i];
1078                 set_page_private(page, (unsigned long)zspage);
1079                 page->freelist = NULL;
1080                 if (i == 0) {
1081                         zspage->first_page = page;
1082                         SetPagePrivate(page);
1083                         if (unlikely(class->objs_per_zspage == 1 &&
1084                                         class->pages_per_zspage == 1))
1085                                 SetPageHugeObject(page);
1086                 } else {
1087                         prev_page->freelist = page;
1088                 }
1089                 prev_page = page;
1090         }
1091 }
1092
1093 /*
1094  * Allocate a zspage for the given size class
1095  */
1096 static struct zspage *alloc_zspage(struct zs_pool *pool,
1097                                         struct size_class *class,
1098                                         gfp_t gfp)
1099 {
1100         int i;
1101         struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE];
1102         struct zspage *zspage = cache_alloc_zspage(pool, gfp);
1103
1104         if (!zspage)
1105                 return NULL;
1106
1107         memset(zspage, 0, sizeof(struct zspage));
1108         zspage->magic = ZSPAGE_MAGIC;
1109         migrate_lock_init(zspage);
1110
1111         for (i = 0; i < class->pages_per_zspage; i++) {
1112                 struct page *page;
1113
1114                 page = alloc_page(gfp);
1115                 if (!page) {
1116                         while (--i >= 0) {
1117                                 dec_zone_page_state(pages[i], NR_ZSPAGES);
1118                                 __free_page(pages[i]);
1119                         }
1120                         cache_free_zspage(pool, zspage);
1121                         return NULL;
1122                 }
1123
1124                 inc_zone_page_state(page, NR_ZSPAGES);
1125                 pages[i] = page;
1126         }
1127
1128         create_page_chain(class, zspage, pages);
1129         init_zspage(class, zspage);
1130
1131         return zspage;
1132 }
1133
1134 static struct zspage *find_get_zspage(struct size_class *class)
1135 {
1136         int i;
1137         struct zspage *zspage;
1138
1139         for (i = ZS_ALMOST_FULL; i >= ZS_EMPTY; i--) {
1140                 zspage = list_first_entry_or_null(&class->fullness_list[i],
1141                                 struct zspage, list);
1142                 if (zspage)
1143                         break;
1144         }
1145
1146         return zspage;
1147 }
1148
1149 #ifdef CONFIG_PGTABLE_MAPPING
1150 static inline int __zs_cpu_up(struct mapping_area *area)
1151 {
1152         /*
1153          * Make sure we don't leak memory if a cpu UP notification
1154          * and zs_init() race and both call zs_cpu_up() on the same cpu
1155          */
1156         if (area->vm)
1157                 return 0;
1158         area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL);
1159         if (!area->vm)
1160                 return -ENOMEM;
1161         return 0;
1162 }
1163
1164 static inline void __zs_cpu_down(struct mapping_area *area)
1165 {
1166         if (area->vm)
1167                 free_vm_area(area->vm);
1168         area->vm = NULL;
1169 }
1170
1171 static inline void *__zs_map_object(struct mapping_area *area,
1172                                 struct page *pages[2], int off, int size)
1173 {
1174         BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages));
1175         area->vm_addr = area->vm->addr;
1176         return area->vm_addr + off;
1177 }
1178
1179 static inline void __zs_unmap_object(struct mapping_area *area,
1180                                 struct page *pages[2], int off, int size)
1181 {
1182         unsigned long addr = (unsigned long)area->vm_addr;
1183
1184         unmap_kernel_range(addr, PAGE_SIZE * 2);
1185 }
1186
1187 #else /* CONFIG_PGTABLE_MAPPING */
1188
1189 static inline int __zs_cpu_up(struct mapping_area *area)
1190 {
1191         /*
1192          * Make sure we don't leak memory if a cpu UP notification
1193          * and zs_init() race and both call zs_cpu_up() on the same cpu
1194          */
1195         if (area->vm_buf)
1196                 return 0;
1197         area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
1198         if (!area->vm_buf)
1199                 return -ENOMEM;
1200         return 0;
1201 }
1202
1203 static inline void __zs_cpu_down(struct mapping_area *area)
1204 {
1205         kfree(area->vm_buf);
1206         area->vm_buf = NULL;
1207 }
1208
1209 static void *__zs_map_object(struct mapping_area *area,
1210                         struct page *pages[2], int off, int size)
1211 {
1212         int sizes[2];
1213         void *addr;
1214         char *buf = area->vm_buf;
1215
1216         /* disable page faults to match kmap_atomic() return conditions */
1217         pagefault_disable();
1218
1219         /* no read fastpath */
1220         if (area->vm_mm == ZS_MM_WO)
1221                 goto out;
1222
1223         sizes[0] = PAGE_SIZE - off;
1224         sizes[1] = size - sizes[0];
1225
1226         /* copy object to per-cpu buffer */
1227         addr = kmap_atomic(pages[0]);
1228         memcpy(buf, addr + off, sizes[0]);
1229         kunmap_atomic(addr);
1230         addr = kmap_atomic(pages[1]);
1231         memcpy(buf + sizes[0], addr, sizes[1]);
1232         kunmap_atomic(addr);
1233 out:
1234         return area->vm_buf;
1235 }
1236
1237 static void __zs_unmap_object(struct mapping_area *area,
1238                         struct page *pages[2], int off, int size)
1239 {
1240         int sizes[2];
1241         void *addr;
1242         char *buf;
1243
1244         /* no write fastpath */
1245         if (area->vm_mm == ZS_MM_RO)
1246                 goto out;
1247
1248         buf = area->vm_buf;
1249         buf = buf + ZS_HANDLE_SIZE;
1250         size -= ZS_HANDLE_SIZE;
1251         off += ZS_HANDLE_SIZE;
1252
1253         sizes[0] = PAGE_SIZE - off;
1254         sizes[1] = size - sizes[0];
1255
1256         /* copy per-cpu buffer to object */
1257         addr = kmap_atomic(pages[0]);
1258         memcpy(addr + off, buf, sizes[0]);
1259         kunmap_atomic(addr);
1260         addr = kmap_atomic(pages[1]);
1261         memcpy(addr, buf + sizes[0], sizes[1]);
1262         kunmap_atomic(addr);
1263
1264 out:
1265         /* enable page faults to match kunmap_atomic() return conditions */
1266         pagefault_enable();
1267 }
1268
1269 #endif /* CONFIG_PGTABLE_MAPPING */
1270
1271 static int zs_cpu_prepare(unsigned int cpu)
1272 {
1273         struct mapping_area *area;
1274
1275         area = &per_cpu(zs_map_area, cpu);
1276         return __zs_cpu_up(area);
1277 }
1278
1279 static int zs_cpu_dead(unsigned int cpu)
1280 {
1281         struct mapping_area *area;
1282
1283         area = &per_cpu(zs_map_area, cpu);
1284         __zs_cpu_down(area);
1285         return 0;
1286 }
1287
1288 static bool can_merge(struct size_class *prev, int pages_per_zspage,
1289                                         int objs_per_zspage)
1290 {
1291         if (prev->pages_per_zspage == pages_per_zspage &&
1292                 prev->objs_per_zspage == objs_per_zspage)
1293                 return true;
1294
1295         return false;
1296 }
1297
1298 static bool zspage_full(struct size_class *class, struct zspage *zspage)
1299 {
1300         return get_zspage_inuse(zspage) == class->objs_per_zspage;
1301 }
1302
1303 unsigned long zs_get_total_pages(struct zs_pool *pool)
1304 {
1305         return atomic_long_read(&pool->pages_allocated);
1306 }
1307 EXPORT_SYMBOL_GPL(zs_get_total_pages);
1308
1309 /**
1310  * zs_map_object - get address of allocated object from handle.
1311  * @pool: pool from which the object was allocated
1312  * @handle: handle returned from zs_malloc
1313  *
1314  * Before using an object allocated from zs_malloc, it must be mapped using
1315  * this function. When done with the object, it must be unmapped using
1316  * zs_unmap_object.
1317  *
1318  * Only one object can be mapped per cpu at a time. There is no protection
1319  * against nested mappings.
1320  *
1321  * This function returns with preemption and page faults disabled.
1322  */
1323 void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1324                         enum zs_mapmode mm)
1325 {
1326         struct zspage *zspage;
1327         struct page *page;
1328         unsigned long obj, off;
1329         unsigned int obj_idx;
1330
1331         unsigned int class_idx;
1332         enum fullness_group fg;
1333         struct size_class *class;
1334         struct mapping_area *area;
1335         struct page *pages[2];
1336         void *ret;
1337
1338         /*
1339          * Because we use per-cpu mapping areas shared among the
1340          * pools/users, we can't allow mapping in interrupt context
1341          * because it can corrupt another users mappings.
1342          */
1343         BUG_ON(in_interrupt());
1344
1345         /* From now on, migration cannot move the object */
1346         pin_tag(handle);
1347
1348         obj = handle_to_obj(handle);
1349         obj_to_location(obj, &page, &obj_idx);
1350         zspage = get_zspage(page);
1351
1352         /* migration cannot move any subpage in this zspage */
1353         migrate_read_lock(zspage);
1354
1355         get_zspage_mapping(zspage, &class_idx, &fg);
1356         class = pool->size_class[class_idx];
1357         off = (class->size * obj_idx) & ~PAGE_MASK;
1358
1359         area = &get_cpu_var(zs_map_area);
1360         area->vm_mm = mm;
1361         if (off + class->size <= PAGE_SIZE) {
1362                 /* this object is contained entirely within a page */
1363                 area->vm_addr = kmap_atomic(page);
1364                 ret = area->vm_addr + off;
1365                 goto out;
1366         }
1367
1368         /* this object spans two pages */
1369         pages[0] = page;
1370         pages[1] = get_next_page(page);
1371         BUG_ON(!pages[1]);
1372
1373         ret = __zs_map_object(area, pages, off, class->size);
1374 out:
1375         if (likely(!PageHugeObject(page)))
1376                 ret += ZS_HANDLE_SIZE;
1377
1378         return ret;
1379 }
1380 EXPORT_SYMBOL_GPL(zs_map_object);
1381
1382 void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1383 {
1384         struct zspage *zspage;
1385         struct page *page;
1386         unsigned long obj, off;
1387         unsigned int obj_idx;
1388
1389         unsigned int class_idx;
1390         enum fullness_group fg;
1391         struct size_class *class;
1392         struct mapping_area *area;
1393
1394         obj = handle_to_obj(handle);
1395         obj_to_location(obj, &page, &obj_idx);
1396         zspage = get_zspage(page);
1397         get_zspage_mapping(zspage, &class_idx, &fg);
1398         class = pool->size_class[class_idx];
1399         off = (class->size * obj_idx) & ~PAGE_MASK;
1400
1401         area = this_cpu_ptr(&zs_map_area);
1402         if (off + class->size <= PAGE_SIZE)
1403                 kunmap_atomic(area->vm_addr);
1404         else {
1405                 struct page *pages[2];
1406
1407                 pages[0] = page;
1408                 pages[1] = get_next_page(page);
1409                 BUG_ON(!pages[1]);
1410
1411                 __zs_unmap_object(area, pages, off, class->size);
1412         }
1413         put_cpu_var(zs_map_area);
1414
1415         migrate_read_unlock(zspage);
1416         unpin_tag(handle);
1417 }
1418 EXPORT_SYMBOL_GPL(zs_unmap_object);
1419
1420 static unsigned long obj_malloc(struct size_class *class,
1421                                 struct zspage *zspage, unsigned long handle)
1422 {
1423         int i, nr_page, offset;
1424         unsigned long obj;
1425         struct link_free *link;
1426
1427         struct page *m_page;
1428         unsigned long m_offset;
1429         void *vaddr;
1430
1431         handle |= OBJ_ALLOCATED_TAG;
1432         obj = get_freeobj(zspage);
1433
1434         offset = obj * class->size;
1435         nr_page = offset >> PAGE_SHIFT;
1436         m_offset = offset & ~PAGE_MASK;
1437         m_page = get_first_page(zspage);
1438
1439         for (i = 0; i < nr_page; i++)
1440                 m_page = get_next_page(m_page);
1441
1442         vaddr = kmap_atomic(m_page);
1443         link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1444         set_freeobj(zspage, link->next >> OBJ_TAG_BITS);
1445         if (likely(!PageHugeObject(m_page)))
1446                 /* record handle in the header of allocated chunk */
1447                 link->handle = handle;
1448         else
1449                 /* record handle to page->index */
1450                 zspage->first_page->index = handle;
1451
1452         kunmap_atomic(vaddr);
1453         mod_zspage_inuse(zspage, 1);
1454         zs_stat_inc(class, OBJ_USED, 1);
1455
1456         obj = location_to_obj(m_page, obj);
1457
1458         return obj;
1459 }
1460
1461
1462 /**
1463  * zs_malloc - Allocate block of given size from pool.
1464  * @pool: pool to allocate from
1465  * @size: size of block to allocate
1466  * @gfp: gfp flags when allocating object
1467  *
1468  * On success, handle to the allocated object is returned,
1469  * otherwise 0.
1470  * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1471  */
1472 unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp)
1473 {
1474         unsigned long handle, obj;
1475         struct size_class *class;
1476         enum fullness_group newfg;
1477         struct zspage *zspage;
1478
1479         if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
1480                 return 0;
1481
1482         handle = cache_alloc_handle(pool, gfp);
1483         if (!handle)
1484                 return 0;
1485
1486         /* extra space in chunk to keep the handle */
1487         size += ZS_HANDLE_SIZE;
1488         class = pool->size_class[get_size_class_index(size)];
1489
1490         spin_lock(&class->lock);
1491         zspage = find_get_zspage(class);
1492         if (likely(zspage)) {
1493                 obj = obj_malloc(class, zspage, handle);
1494                 /* Now move the zspage to another fullness group, if required */
1495                 fix_fullness_group(class, zspage);
1496                 record_obj(handle, obj);
1497                 spin_unlock(&class->lock);
1498
1499                 return handle;
1500         }
1501
1502         spin_unlock(&class->lock);
1503
1504         zspage = alloc_zspage(pool, class, gfp);
1505         if (!zspage) {
1506                 cache_free_handle(pool, handle);
1507                 return 0;
1508         }
1509
1510         spin_lock(&class->lock);
1511         obj = obj_malloc(class, zspage, handle);
1512         newfg = get_fullness_group(class, zspage);
1513         insert_zspage(class, zspage, newfg);
1514         set_zspage_mapping(zspage, class->index, newfg);
1515         record_obj(handle, obj);
1516         atomic_long_add(class->pages_per_zspage,
1517                                 &pool->pages_allocated);
1518         zs_stat_inc(class, OBJ_ALLOCATED, class->objs_per_zspage);
1519
1520         /* We completely set up zspage so mark them as movable */
1521         SetZsPageMovable(pool, zspage);
1522         spin_unlock(&class->lock);
1523
1524         return handle;
1525 }
1526 EXPORT_SYMBOL_GPL(zs_malloc);
1527
1528 static void obj_free(struct size_class *class, unsigned long obj)
1529 {
1530         struct link_free *link;
1531         struct zspage *zspage;
1532         struct page *f_page;
1533         unsigned long f_offset;
1534         unsigned int f_objidx;
1535         void *vaddr;
1536
1537         obj &= ~OBJ_ALLOCATED_TAG;
1538         obj_to_location(obj, &f_page, &f_objidx);
1539         f_offset = (class->size * f_objidx) & ~PAGE_MASK;
1540         zspage = get_zspage(f_page);
1541
1542         vaddr = kmap_atomic(f_page);
1543
1544         /* Insert this object in containing zspage's freelist */
1545         link = (struct link_free *)(vaddr + f_offset);
1546         link->next = get_freeobj(zspage) << OBJ_TAG_BITS;
1547         kunmap_atomic(vaddr);
1548         set_freeobj(zspage, f_objidx);
1549         mod_zspage_inuse(zspage, -1);
1550         zs_stat_dec(class, OBJ_USED, 1);
1551 }
1552
1553 void zs_free(struct zs_pool *pool, unsigned long handle)
1554 {
1555         struct zspage *zspage;
1556         struct page *f_page;
1557         unsigned long obj;
1558         unsigned int f_objidx;
1559         int class_idx;
1560         struct size_class *class;
1561         enum fullness_group fullness;
1562         bool isolated;
1563
1564         if (unlikely(!handle))
1565                 return;
1566
1567         pin_tag(handle);
1568         obj = handle_to_obj(handle);
1569         obj_to_location(obj, &f_page, &f_objidx);
1570         zspage = get_zspage(f_page);
1571
1572         migrate_read_lock(zspage);
1573
1574         get_zspage_mapping(zspage, &class_idx, &fullness);
1575         class = pool->size_class[class_idx];
1576
1577         spin_lock(&class->lock);
1578         obj_free(class, obj);
1579         fullness = fix_fullness_group(class, zspage);
1580         if (fullness != ZS_EMPTY) {
1581                 migrate_read_unlock(zspage);
1582                 goto out;
1583         }
1584
1585         isolated = is_zspage_isolated(zspage);
1586         migrate_read_unlock(zspage);
1587         /* If zspage is isolated, zs_page_putback will free the zspage */
1588         if (likely(!isolated))
1589                 free_zspage(pool, class, zspage);
1590 out:
1591
1592         spin_unlock(&class->lock);
1593         unpin_tag(handle);
1594         cache_free_handle(pool, handle);
1595 }
1596 EXPORT_SYMBOL_GPL(zs_free);
1597
1598 static void zs_object_copy(struct size_class *class, unsigned long dst,
1599                                 unsigned long src)
1600 {
1601         struct page *s_page, *d_page;
1602         unsigned int s_objidx, d_objidx;
1603         unsigned long s_off, d_off;
1604         void *s_addr, *d_addr;
1605         int s_size, d_size, size;
1606         int written = 0;
1607
1608         s_size = d_size = class->size;
1609
1610         obj_to_location(src, &s_page, &s_objidx);
1611         obj_to_location(dst, &d_page, &d_objidx);
1612
1613         s_off = (class->size * s_objidx) & ~PAGE_MASK;
1614         d_off = (class->size * d_objidx) & ~PAGE_MASK;
1615
1616         if (s_off + class->size > PAGE_SIZE)
1617                 s_size = PAGE_SIZE - s_off;
1618
1619         if (d_off + class->size > PAGE_SIZE)
1620                 d_size = PAGE_SIZE - d_off;
1621
1622         s_addr = kmap_atomic(s_page);
1623         d_addr = kmap_atomic(d_page);
1624
1625         while (1) {
1626                 size = min(s_size, d_size);
1627                 memcpy(d_addr + d_off, s_addr + s_off, size);
1628                 written += size;
1629
1630                 if (written == class->size)
1631                         break;
1632
1633                 s_off += size;
1634                 s_size -= size;
1635                 d_off += size;
1636                 d_size -= size;
1637
1638                 if (s_off >= PAGE_SIZE) {
1639                         kunmap_atomic(d_addr);
1640                         kunmap_atomic(s_addr);
1641                         s_page = get_next_page(s_page);
1642                         s_addr = kmap_atomic(s_page);
1643                         d_addr = kmap_atomic(d_page);
1644                         s_size = class->size - written;
1645                         s_off = 0;
1646                 }
1647
1648                 if (d_off >= PAGE_SIZE) {
1649                         kunmap_atomic(d_addr);
1650                         d_page = get_next_page(d_page);
1651                         d_addr = kmap_atomic(d_page);
1652                         d_size = class->size - written;
1653                         d_off = 0;
1654                 }
1655         }
1656
1657         kunmap_atomic(d_addr);
1658         kunmap_atomic(s_addr);
1659 }
1660
1661 /*
1662  * Find alloced object in zspage from index object and
1663  * return handle.
1664  */
1665 static unsigned long find_alloced_obj(struct size_class *class,
1666                                         struct page *page, int *obj_idx)
1667 {
1668         unsigned long head;
1669         int offset = 0;
1670         int index = *obj_idx;
1671         unsigned long handle = 0;
1672         void *addr = kmap_atomic(page);
1673
1674         offset = get_first_obj_offset(page);
1675         offset += class->size * index;
1676
1677         while (offset < PAGE_SIZE) {
1678                 head = obj_to_head(page, addr + offset);
1679                 if (head & OBJ_ALLOCATED_TAG) {
1680                         handle = head & ~OBJ_ALLOCATED_TAG;
1681                         if (trypin_tag(handle))
1682                                 break;
1683                         handle = 0;
1684                 }
1685
1686                 offset += class->size;
1687                 index++;
1688         }
1689
1690         kunmap_atomic(addr);
1691
1692         *obj_idx = index;
1693
1694         return handle;
1695 }
1696
1697 struct zs_compact_control {
1698         /* Source spage for migration which could be a subpage of zspage */
1699         struct page *s_page;
1700         /* Destination page for migration which should be a first page
1701          * of zspage. */
1702         struct page *d_page;
1703          /* Starting object index within @s_page which used for live object
1704           * in the subpage. */
1705         int obj_idx;
1706 };
1707
1708 static int migrate_zspage(struct zs_pool *pool, struct size_class *class,
1709                                 struct zs_compact_control *cc)
1710 {
1711         unsigned long used_obj, free_obj;
1712         unsigned long handle;
1713         struct page *s_page = cc->s_page;
1714         struct page *d_page = cc->d_page;
1715         int obj_idx = cc->obj_idx;
1716         int ret = 0;
1717
1718         while (1) {
1719                 handle = find_alloced_obj(class, s_page, &obj_idx);
1720                 if (!handle) {
1721                         s_page = get_next_page(s_page);
1722                         if (!s_page)
1723                                 break;
1724                         obj_idx = 0;
1725                         continue;
1726                 }
1727
1728                 /* Stop if there is no more space */
1729                 if (zspage_full(class, get_zspage(d_page))) {
1730                         unpin_tag(handle);
1731                         ret = -ENOMEM;
1732                         break;
1733                 }
1734
1735                 used_obj = handle_to_obj(handle);
1736                 free_obj = obj_malloc(class, get_zspage(d_page), handle);
1737                 zs_object_copy(class, free_obj, used_obj);
1738                 obj_idx++;
1739                 /*
1740                  * record_obj updates handle's value to free_obj and it will
1741                  * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which
1742                  * breaks synchronization using pin_tag(e,g, zs_free) so
1743                  * let's keep the lock bit.
1744                  */
1745                 free_obj |= BIT(HANDLE_PIN_BIT);
1746                 record_obj(handle, free_obj);
1747                 unpin_tag(handle);
1748                 obj_free(class, used_obj);
1749         }
1750
1751         /* Remember last position in this iteration */
1752         cc->s_page = s_page;
1753         cc->obj_idx = obj_idx;
1754
1755         return ret;
1756 }
1757
1758 static struct zspage *isolate_zspage(struct size_class *class, bool source)
1759 {
1760         int i;
1761         struct zspage *zspage;
1762         enum fullness_group fg[2] = {ZS_ALMOST_EMPTY, ZS_ALMOST_FULL};
1763
1764         if (!source) {
1765                 fg[0] = ZS_ALMOST_FULL;
1766                 fg[1] = ZS_ALMOST_EMPTY;
1767         }
1768
1769         for (i = 0; i < 2; i++) {
1770                 zspage = list_first_entry_or_null(&class->fullness_list[fg[i]],
1771                                                         struct zspage, list);
1772                 if (zspage) {
1773                         VM_BUG_ON(is_zspage_isolated(zspage));
1774                         remove_zspage(class, zspage, fg[i]);
1775                         return zspage;
1776                 }
1777         }
1778
1779         return zspage;
1780 }
1781
1782 /*
1783  * putback_zspage - add @zspage into right class's fullness list
1784  * @class: destination class
1785  * @zspage: target page
1786  *
1787  * Return @zspage's fullness_group
1788  */
1789 static enum fullness_group putback_zspage(struct size_class *class,
1790                         struct zspage *zspage)
1791 {
1792         enum fullness_group fullness;
1793
1794         VM_BUG_ON(is_zspage_isolated(zspage));
1795
1796         fullness = get_fullness_group(class, zspage);
1797         insert_zspage(class, zspage, fullness);
1798         set_zspage_mapping(zspage, class->index, fullness);
1799
1800         return fullness;
1801 }
1802
1803 #ifdef CONFIG_COMPACTION
1804 static struct dentry *zs_mount(struct file_system_type *fs_type,
1805                                 int flags, const char *dev_name, void *data)
1806 {
1807         static const struct dentry_operations ops = {
1808                 .d_dname = simple_dname,
1809         };
1810
1811         return mount_pseudo(fs_type, "zsmalloc:", NULL, &ops, ZSMALLOC_MAGIC);
1812 }
1813
1814 static struct file_system_type zsmalloc_fs = {
1815         .name           = "zsmalloc",
1816         .mount          = zs_mount,
1817         .kill_sb        = kill_anon_super,
1818 };
1819
1820 static int zsmalloc_mount(void)
1821 {
1822         int ret = 0;
1823
1824         zsmalloc_mnt = kern_mount(&zsmalloc_fs);
1825         if (IS_ERR(zsmalloc_mnt))
1826                 ret = PTR_ERR(zsmalloc_mnt);
1827
1828         return ret;
1829 }
1830
1831 static void zsmalloc_unmount(void)
1832 {
1833         kern_unmount(zsmalloc_mnt);
1834 }
1835
1836 static void migrate_lock_init(struct zspage *zspage)
1837 {
1838         rwlock_init(&zspage->lock);
1839 }
1840
1841 static void migrate_read_lock(struct zspage *zspage)
1842 {
1843         read_lock(&zspage->lock);
1844 }
1845
1846 static void migrate_read_unlock(struct zspage *zspage)
1847 {
1848         read_unlock(&zspage->lock);
1849 }
1850
1851 static void migrate_write_lock(struct zspage *zspage)
1852 {
1853         write_lock(&zspage->lock);
1854 }
1855
1856 static void migrate_write_unlock(struct zspage *zspage)
1857 {
1858         write_unlock(&zspage->lock);
1859 }
1860
1861 /* Number of isolated subpage for *page migration* in this zspage */
1862 static void inc_zspage_isolation(struct zspage *zspage)
1863 {
1864         zspage->isolated++;
1865 }
1866
1867 static void dec_zspage_isolation(struct zspage *zspage)
1868 {
1869         zspage->isolated--;
1870 }
1871
1872 static void replace_sub_page(struct size_class *class, struct zspage *zspage,
1873                                 struct page *newpage, struct page *oldpage)
1874 {
1875         struct page *page;
1876         struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE] = {NULL, };
1877         int idx = 0;
1878
1879         page = get_first_page(zspage);
1880         do {
1881                 if (page == oldpage)
1882                         pages[idx] = newpage;
1883                 else
1884                         pages[idx] = page;
1885                 idx++;
1886         } while ((page = get_next_page(page)) != NULL);
1887
1888         create_page_chain(class, zspage, pages);
1889         set_first_obj_offset(newpage, get_first_obj_offset(oldpage));
1890         if (unlikely(PageHugeObject(oldpage)))
1891                 newpage->index = oldpage->index;
1892         __SetPageMovable(newpage, page_mapping(oldpage));
1893 }
1894
1895 bool zs_page_isolate(struct page *page, isolate_mode_t mode)
1896 {
1897         struct zs_pool *pool;
1898         struct size_class *class;
1899         int class_idx;
1900         enum fullness_group fullness;
1901         struct zspage *zspage;
1902         struct address_space *mapping;
1903
1904         /*
1905          * Page is locked so zspage couldn't be destroyed. For detail, look at
1906          * lock_zspage in free_zspage.
1907          */
1908         VM_BUG_ON_PAGE(!PageMovable(page), page);
1909         VM_BUG_ON_PAGE(PageIsolated(page), page);
1910
1911         zspage = get_zspage(page);
1912
1913         /*
1914          * Without class lock, fullness could be stale while class_idx is okay
1915          * because class_idx is constant unless page is freed so we should get
1916          * fullness again under class lock.
1917          */
1918         get_zspage_mapping(zspage, &class_idx, &fullness);
1919         mapping = page_mapping(page);
1920         pool = mapping->private_data;
1921         class = pool->size_class[class_idx];
1922
1923         spin_lock(&class->lock);
1924         if (get_zspage_inuse(zspage) == 0) {
1925                 spin_unlock(&class->lock);
1926                 return false;
1927         }
1928
1929         /* zspage is isolated for object migration */
1930         if (list_empty(&zspage->list) && !is_zspage_isolated(zspage)) {
1931                 spin_unlock(&class->lock);
1932                 return false;
1933         }
1934
1935         /*
1936          * If this is first time isolation for the zspage, isolate zspage from
1937          * size_class to prevent further object allocation from the zspage.
1938          */
1939         if (!list_empty(&zspage->list) && !is_zspage_isolated(zspage)) {
1940                 get_zspage_mapping(zspage, &class_idx, &fullness);
1941                 remove_zspage(class, zspage, fullness);
1942         }
1943
1944         inc_zspage_isolation(zspage);
1945         spin_unlock(&class->lock);
1946
1947         return true;
1948 }
1949
1950 int zs_page_migrate(struct address_space *mapping, struct page *newpage,
1951                 struct page *page, enum migrate_mode mode)
1952 {
1953         struct zs_pool *pool;
1954         struct size_class *class;
1955         int class_idx;
1956         enum fullness_group fullness;
1957         struct zspage *zspage;
1958         struct page *dummy;
1959         void *s_addr, *d_addr, *addr;
1960         int offset, pos;
1961         unsigned long handle, head;
1962         unsigned long old_obj, new_obj;
1963         unsigned int obj_idx;
1964         int ret = -EAGAIN;
1965
1966         /*
1967          * We cannot support the _NO_COPY case here, because copy needs to
1968          * happen under the zs lock, which does not work with
1969          * MIGRATE_SYNC_NO_COPY workflow.
1970          */
1971         if (mode == MIGRATE_SYNC_NO_COPY)
1972                 return -EINVAL;
1973
1974         VM_BUG_ON_PAGE(!PageMovable(page), page);
1975         VM_BUG_ON_PAGE(!PageIsolated(page), page);
1976
1977         zspage = get_zspage(page);
1978
1979         /* Concurrent compactor cannot migrate any subpage in zspage */
1980         migrate_write_lock(zspage);
1981         get_zspage_mapping(zspage, &class_idx, &fullness);
1982         pool = mapping->private_data;
1983         class = pool->size_class[class_idx];
1984         offset = get_first_obj_offset(page);
1985
1986         spin_lock(&class->lock);
1987         if (!get_zspage_inuse(zspage)) {
1988                 /*
1989                  * Set "offset" to end of the page so that every loops
1990                  * skips unnecessary object scanning.
1991                  */
1992                 offset = PAGE_SIZE;
1993         }
1994
1995         pos = offset;
1996         s_addr = kmap_atomic(page);
1997         while (pos < PAGE_SIZE) {
1998                 head = obj_to_head(page, s_addr + pos);
1999                 if (head & OBJ_ALLOCATED_TAG) {
2000                         handle = head & ~OBJ_ALLOCATED_TAG;
2001                         if (!trypin_tag(handle))
2002                                 goto unpin_objects;
2003                 }
2004                 pos += class->size;
2005         }
2006
2007         /*
2008          * Here, any user cannot access all objects in the zspage so let's move.
2009          */
2010         d_addr = kmap_atomic(newpage);
2011         memcpy(d_addr, s_addr, PAGE_SIZE);
2012         kunmap_atomic(d_addr);
2013
2014         for (addr = s_addr + offset; addr < s_addr + pos;
2015                                         addr += class->size) {
2016                 head = obj_to_head(page, addr);
2017                 if (head & OBJ_ALLOCATED_TAG) {
2018                         handle = head & ~OBJ_ALLOCATED_TAG;
2019                         if (!testpin_tag(handle))
2020                                 BUG();
2021
2022                         old_obj = handle_to_obj(handle);
2023                         obj_to_location(old_obj, &dummy, &obj_idx);
2024                         new_obj = (unsigned long)location_to_obj(newpage,
2025                                                                 obj_idx);
2026                         new_obj |= BIT(HANDLE_PIN_BIT);
2027                         record_obj(handle, new_obj);
2028                 }
2029         }
2030
2031         replace_sub_page(class, zspage, newpage, page);
2032         get_page(newpage);
2033
2034         dec_zspage_isolation(zspage);
2035
2036         /*
2037          * Page migration is done so let's putback isolated zspage to
2038          * the list if @page is final isolated subpage in the zspage.
2039          */
2040         if (!is_zspage_isolated(zspage))
2041                 putback_zspage(class, zspage);
2042
2043         reset_page(page);
2044         put_page(page);
2045         page = newpage;
2046
2047         ret = MIGRATEPAGE_SUCCESS;
2048 unpin_objects:
2049         for (addr = s_addr + offset; addr < s_addr + pos;
2050                                                 addr += class->size) {
2051                 head = obj_to_head(page, addr);
2052                 if (head & OBJ_ALLOCATED_TAG) {
2053                         handle = head & ~OBJ_ALLOCATED_TAG;
2054                         if (!testpin_tag(handle))
2055                                 BUG();
2056                         unpin_tag(handle);
2057                 }
2058         }
2059         kunmap_atomic(s_addr);
2060         spin_unlock(&class->lock);
2061         migrate_write_unlock(zspage);
2062
2063         return ret;
2064 }
2065
2066 void zs_page_putback(struct page *page)
2067 {
2068         struct zs_pool *pool;
2069         struct size_class *class;
2070         int class_idx;
2071         enum fullness_group fg;
2072         struct address_space *mapping;
2073         struct zspage *zspage;
2074
2075         VM_BUG_ON_PAGE(!PageMovable(page), page);
2076         VM_BUG_ON_PAGE(!PageIsolated(page), page);
2077
2078         zspage = get_zspage(page);
2079         get_zspage_mapping(zspage, &class_idx, &fg);
2080         mapping = page_mapping(page);
2081         pool = mapping->private_data;
2082         class = pool->size_class[class_idx];
2083
2084         spin_lock(&class->lock);
2085         dec_zspage_isolation(zspage);
2086         if (!is_zspage_isolated(zspage)) {
2087                 fg = putback_zspage(class, zspage);
2088                 /*
2089                  * Due to page_lock, we cannot free zspage immediately
2090                  * so let's defer.
2091                  */
2092                 if (fg == ZS_EMPTY)
2093                         schedule_work(&pool->free_work);
2094         }
2095         spin_unlock(&class->lock);
2096 }
2097
2098 const struct address_space_operations zsmalloc_aops = {
2099         .isolate_page = zs_page_isolate,
2100         .migratepage = zs_page_migrate,
2101         .putback_page = zs_page_putback,
2102 };
2103
2104 static int zs_register_migration(struct zs_pool *pool)
2105 {
2106         pool->inode = alloc_anon_inode(zsmalloc_mnt->mnt_sb);
2107         if (IS_ERR(pool->inode)) {
2108                 pool->inode = NULL;
2109                 return 1;
2110         }
2111
2112         pool->inode->i_mapping->private_data = pool;
2113         pool->inode->i_mapping->a_ops = &zsmalloc_aops;
2114         return 0;
2115 }
2116
2117 static void zs_unregister_migration(struct zs_pool *pool)
2118 {
2119         flush_work(&pool->free_work);
2120         iput(pool->inode);
2121 }
2122
2123 /*
2124  * Caller should hold page_lock of all pages in the zspage
2125  * In here, we cannot use zspage meta data.
2126  */
2127 static void async_free_zspage(struct work_struct *work)
2128 {
2129         int i;
2130         struct size_class *class;
2131         unsigned int class_idx;
2132         enum fullness_group fullness;
2133         struct zspage *zspage, *tmp;
2134         LIST_HEAD(free_pages);
2135         struct zs_pool *pool = container_of(work, struct zs_pool,
2136                                         free_work);
2137
2138         for (i = 0; i < ZS_SIZE_CLASSES; i++) {
2139                 class = pool->size_class[i];
2140                 if (class->index != i)
2141                         continue;
2142
2143                 spin_lock(&class->lock);
2144                 list_splice_init(&class->fullness_list[ZS_EMPTY], &free_pages);
2145                 spin_unlock(&class->lock);
2146         }
2147
2148
2149         list_for_each_entry_safe(zspage, tmp, &free_pages, list) {
2150                 list_del(&zspage->list);
2151                 lock_zspage(zspage);
2152
2153                 get_zspage_mapping(zspage, &class_idx, &fullness);
2154                 VM_BUG_ON(fullness != ZS_EMPTY);
2155                 class = pool->size_class[class_idx];
2156                 spin_lock(&class->lock);
2157                 __free_zspage(pool, pool->size_class[class_idx], zspage);
2158                 spin_unlock(&class->lock);
2159         }
2160 };
2161
2162 static void kick_deferred_free(struct zs_pool *pool)
2163 {
2164         schedule_work(&pool->free_work);
2165 }
2166
2167 static void init_deferred_free(struct zs_pool *pool)
2168 {
2169         INIT_WORK(&pool->free_work, async_free_zspage);
2170 }
2171
2172 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage)
2173 {
2174         struct page *page = get_first_page(zspage);
2175
2176         do {
2177                 WARN_ON(!trylock_page(page));
2178                 __SetPageMovable(page, pool->inode->i_mapping);
2179                 unlock_page(page);
2180         } while ((page = get_next_page(page)) != NULL);
2181 }
2182 #endif
2183
2184 /*
2185  *
2186  * Based on the number of unused allocated objects calculate
2187  * and return the number of pages that we can free.
2188  */
2189 static unsigned long zs_can_compact(struct size_class *class)
2190 {
2191         unsigned long obj_wasted;
2192         unsigned long obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
2193         unsigned long obj_used = zs_stat_get(class, OBJ_USED);
2194
2195         if (obj_allocated <= obj_used)
2196                 return 0;
2197
2198         obj_wasted = obj_allocated - obj_used;
2199         obj_wasted /= class->objs_per_zspage;
2200
2201         return obj_wasted * class->pages_per_zspage;
2202 }
2203
2204 static void __zs_compact(struct zs_pool *pool, struct size_class *class)
2205 {
2206         struct zs_compact_control cc;
2207         struct zspage *src_zspage;
2208         struct zspage *dst_zspage = NULL;
2209
2210         spin_lock(&class->lock);
2211         while ((src_zspage = isolate_zspage(class, true))) {
2212
2213                 if (!zs_can_compact(class))
2214                         break;
2215
2216                 cc.obj_idx = 0;
2217                 cc.s_page = get_first_page(src_zspage);
2218
2219                 while ((dst_zspage = isolate_zspage(class, false))) {
2220                         cc.d_page = get_first_page(dst_zspage);
2221                         /*
2222                          * If there is no more space in dst_page, resched
2223                          * and see if anyone had allocated another zspage.
2224                          */
2225                         if (!migrate_zspage(pool, class, &cc))
2226                                 break;
2227
2228                         putback_zspage(class, dst_zspage);
2229                 }
2230
2231                 /* Stop if we couldn't find slot */
2232                 if (dst_zspage == NULL)
2233                         break;
2234
2235                 putback_zspage(class, dst_zspage);
2236                 if (putback_zspage(class, src_zspage) == ZS_EMPTY) {
2237                         free_zspage(pool, class, src_zspage);
2238                         pool->stats.pages_compacted += class->pages_per_zspage;
2239                 }
2240                 spin_unlock(&class->lock);
2241                 cond_resched();
2242                 spin_lock(&class->lock);
2243         }
2244
2245         if (src_zspage)
2246                 putback_zspage(class, src_zspage);
2247
2248         spin_unlock(&class->lock);
2249 }
2250
2251 unsigned long zs_compact(struct zs_pool *pool)
2252 {
2253         int i;
2254         struct size_class *class;
2255
2256         for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2257                 class = pool->size_class[i];
2258                 if (!class)
2259                         continue;
2260                 if (class->index != i)
2261                         continue;
2262                 __zs_compact(pool, class);
2263         }
2264
2265         return pool->stats.pages_compacted;
2266 }
2267 EXPORT_SYMBOL_GPL(zs_compact);
2268
2269 void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats)
2270 {
2271         memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats));
2272 }
2273 EXPORT_SYMBOL_GPL(zs_pool_stats);
2274
2275 static unsigned long zs_shrinker_scan(struct shrinker *shrinker,
2276                 struct shrink_control *sc)
2277 {
2278         unsigned long pages_freed;
2279         struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2280                         shrinker);
2281
2282         pages_freed = pool->stats.pages_compacted;
2283         /*
2284          * Compact classes and calculate compaction delta.
2285          * Can run concurrently with a manually triggered
2286          * (by user) compaction.
2287          */
2288         pages_freed = zs_compact(pool) - pages_freed;
2289
2290         return pages_freed ? pages_freed : SHRINK_STOP;
2291 }
2292
2293 static unsigned long zs_shrinker_count(struct shrinker *shrinker,
2294                 struct shrink_control *sc)
2295 {
2296         int i;
2297         struct size_class *class;
2298         unsigned long pages_to_free = 0;
2299         struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2300                         shrinker);
2301
2302         for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2303                 class = pool->size_class[i];
2304                 if (!class)
2305                         continue;
2306                 if (class->index != i)
2307                         continue;
2308
2309                 pages_to_free += zs_can_compact(class);
2310         }
2311
2312         return pages_to_free;
2313 }
2314
2315 static void zs_unregister_shrinker(struct zs_pool *pool)
2316 {
2317         unregister_shrinker(&pool->shrinker);
2318 }
2319
2320 static int zs_register_shrinker(struct zs_pool *pool)
2321 {
2322         pool->shrinker.scan_objects = zs_shrinker_scan;
2323         pool->shrinker.count_objects = zs_shrinker_count;
2324         pool->shrinker.batch = 0;
2325         pool->shrinker.seeks = DEFAULT_SEEKS;
2326
2327         return register_shrinker(&pool->shrinker);
2328 }
2329
2330 /**
2331  * zs_create_pool - Creates an allocation pool to work from.
2332  * @name: pool name to be created
2333  *
2334  * This function must be called before anything when using
2335  * the zsmalloc allocator.
2336  *
2337  * On success, a pointer to the newly created pool is returned,
2338  * otherwise NULL.
2339  */
2340 struct zs_pool *zs_create_pool(const char *name)
2341 {
2342         int i;
2343         struct zs_pool *pool;
2344         struct size_class *prev_class = NULL;
2345
2346         pool = kzalloc(sizeof(*pool), GFP_KERNEL);
2347         if (!pool)
2348                 return NULL;
2349
2350         init_deferred_free(pool);
2351
2352         pool->name = kstrdup(name, GFP_KERNEL);
2353         if (!pool->name)
2354                 goto err;
2355
2356         if (create_cache(pool))
2357                 goto err;
2358
2359         /*
2360          * Iterate reversely, because, size of size_class that we want to use
2361          * for merging should be larger or equal to current size.
2362          */
2363         for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2364                 int size;
2365                 int pages_per_zspage;
2366                 int objs_per_zspage;
2367                 struct size_class *class;
2368                 int fullness = 0;
2369
2370                 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
2371                 if (size > ZS_MAX_ALLOC_SIZE)
2372                         size = ZS_MAX_ALLOC_SIZE;
2373                 pages_per_zspage = get_pages_per_zspage(size);
2374                 objs_per_zspage = pages_per_zspage * PAGE_SIZE / size;
2375
2376                 /*
2377                  * size_class is used for normal zsmalloc operation such
2378                  * as alloc/free for that size. Although it is natural that we
2379                  * have one size_class for each size, there is a chance that we
2380                  * can get more memory utilization if we use one size_class for
2381                  * many different sizes whose size_class have same
2382                  * characteristics. So, we makes size_class point to
2383                  * previous size_class if possible.
2384                  */
2385                 if (prev_class) {
2386                         if (can_merge(prev_class, pages_per_zspage, objs_per_zspage)) {
2387                                 pool->size_class[i] = prev_class;
2388                                 continue;
2389                         }
2390                 }
2391
2392                 class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
2393                 if (!class)
2394                         goto err;
2395
2396                 class->size = size;
2397                 class->index = i;
2398                 class->pages_per_zspage = pages_per_zspage;
2399                 class->objs_per_zspage = objs_per_zspage;
2400                 spin_lock_init(&class->lock);
2401                 pool->size_class[i] = class;
2402                 for (fullness = ZS_EMPTY; fullness < NR_ZS_FULLNESS;
2403                                                         fullness++)
2404                         INIT_LIST_HEAD(&class->fullness_list[fullness]);
2405
2406                 prev_class = class;
2407         }
2408
2409         /* debug only, don't abort if it fails */
2410         zs_pool_stat_create(pool, name);
2411
2412         if (zs_register_migration(pool))
2413                 goto err;
2414
2415         /*
2416          * Not critical since shrinker is only used to trigger internal
2417          * defragmentation of the pool which is pretty optional thing.  If
2418          * registration fails we still can use the pool normally and user can
2419          * trigger compaction manually. Thus, ignore return code.
2420          */
2421         zs_register_shrinker(pool);
2422
2423         return pool;
2424
2425 err:
2426         zs_destroy_pool(pool);
2427         return NULL;
2428 }
2429 EXPORT_SYMBOL_GPL(zs_create_pool);
2430
2431 void zs_destroy_pool(struct zs_pool *pool)
2432 {
2433         int i;
2434
2435         zs_unregister_shrinker(pool);
2436         zs_unregister_migration(pool);
2437         zs_pool_stat_destroy(pool);
2438
2439         for (i = 0; i < ZS_SIZE_CLASSES; i++) {
2440                 int fg;
2441                 struct size_class *class = pool->size_class[i];
2442
2443                 if (!class)
2444                         continue;
2445
2446                 if (class->index != i)
2447                         continue;
2448
2449                 for (fg = ZS_EMPTY; fg < NR_ZS_FULLNESS; fg++) {
2450                         if (!list_empty(&class->fullness_list[fg])) {
2451                                 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
2452                                         class->size, fg);
2453                         }
2454                 }
2455                 kfree(class);
2456         }
2457
2458         destroy_cache(pool);
2459         kfree(pool->name);
2460         kfree(pool);
2461 }
2462 EXPORT_SYMBOL_GPL(zs_destroy_pool);
2463
2464 static int __init zs_init(void)
2465 {
2466         int ret;
2467
2468         ret = zsmalloc_mount();
2469         if (ret)
2470                 goto out;
2471
2472         ret = cpuhp_setup_state(CPUHP_MM_ZS_PREPARE, "mm/zsmalloc:prepare",
2473                                 zs_cpu_prepare, zs_cpu_dead);
2474         if (ret)
2475                 goto hp_setup_fail;
2476
2477 #ifdef CONFIG_ZPOOL
2478         zpool_register_driver(&zs_zpool_driver);
2479 #endif
2480
2481         zs_stat_init();
2482
2483         return 0;
2484
2485 hp_setup_fail:
2486         zsmalloc_unmount();
2487 out:
2488         return ret;
2489 }
2490
2491 static void __exit zs_exit(void)
2492 {
2493 #ifdef CONFIG_ZPOOL
2494         zpool_unregister_driver(&zs_zpool_driver);
2495 #endif
2496         zsmalloc_unmount();
2497         cpuhp_remove_state(CPUHP_MM_ZS_PREPARE);
2498
2499         zs_stat_exit();
2500 }
2501
2502 module_init(zs_init);
2503 module_exit(zs_exit);
2504
2505 MODULE_LICENSE("Dual BSD/GPL");
2506 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");