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