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