mm/z3fold.c

   1 /*
   2  * z3fold.c
   3  *
   4  * Author: Vitaly Wool <vitaly.wool@konsulko.com>
   5  * Copyright (C) 2016, Sony Mobile Communications Inc.
   6  *
   7  * This implementation is based on zbud written by Seth Jennings.
   8  *
   9  * z3fold is an special purpose allocator for storing compressed pages. It
  10  * can store up to three compressed pages per page which improves the
  11  * compression ratio of zbud while retaining its main concepts (e. g. always
  12  * storing an integral number of objects per page) and simplicity.
  13  * It still has simple and deterministic reclaim properties that make it
  14  * preferable to a higher density approach (with no requirement on integral
  15  * number of object per page) when reclaim is used.
  16  *
  17  * As in zbud, pages are divided into "chunks".  The size of the chunks is
  18  * fixed at compile time and is determined by NCHUNKS_ORDER below.
  19  *
  20  * z3fold doesn't export any API and is meant to be used via zpool API.
  21  */
  22
  23 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  24
  25 #include <linux/atomic.h>
  26 #include <linux/sched.h>
  27 #include <linux/cpumask.h>
  28 #include <linux/dcache.h>
  29 #include <linux/list.h>
  30 #include <linux/mm.h>
  31 #include <linux/module.h>
  32 #include <linux/page-flags.h>
  33 #include <linux/migrate.h>
  34 #include <linux/node.h>
  35 #include <linux/compaction.h>
  36 #include <linux/percpu.h>
  37 #include <linux/mount.h>
  38 #include <linux/fs.h>
  39 #include <linux/preempt.h>
  40 #include <linux/workqueue.h>
  41 #include <linux/slab.h>
  42 #include <linux/spinlock.h>
  43 #include <linux/zpool.h>
  44
  45 /*
  46  * NCHUNKS_ORDER determines the internal allocation granularity, effectively
  47  * adjusting internal fragmentation.  It also determines the number of
  48  * freelists maintained in each pool. NCHUNKS_ORDER of 6 means that the
  49  * allocation granularity will be in chunks of size PAGE_SIZE/64. Some chunks
  50  * in the beginning of an allocated page are occupied by z3fold header, so
  51  * NCHUNKS will be calculated to 63 (or 62 in case CONFIG_DEBUG_SPINLOCK=y),
  52  * which shows the max number of free chunks in z3fold page, also there will
  53  * be 63, or 62, respectively, freelists per pool.
  54  */
  55 #define NCHUNKS_ORDER   6
  56
  57 #define CHUNK_SHIFT     (PAGE_SHIFT - NCHUNKS_ORDER)
  58 #define CHUNK_SIZE      (1 << CHUNK_SHIFT)
  59 #define ZHDR_SIZE_ALIGNED round_up(sizeof(struct z3fold_header), CHUNK_SIZE)
  60 #define ZHDR_CHUNKS     (ZHDR_SIZE_ALIGNED >> CHUNK_SHIFT)
  61 #define TOTAL_CHUNKS    (PAGE_SIZE >> CHUNK_SHIFT)
  62 #define NCHUNKS         ((PAGE_SIZE - ZHDR_SIZE_ALIGNED) >> CHUNK_SHIFT)
  63
  64 #define BUDDY_MASK      (0x3)
  65 #define BUDDY_SHIFT     2
  66 #define SLOTS_ALIGN     (0x40)
  67
  68 /*****************
  69  * Structures
  70 *****************/
  71 struct z3fold_pool;
  72 struct z3fold_ops {
  73         int (*evict)(struct z3fold_pool *pool, unsigned long handle);
  74 };
  75
  76 enum buddy {
  77         HEADLESS = 0,
  78         FIRST,
  79         MIDDLE,
  80         LAST,
  81         BUDDIES_MAX = LAST
  82 };
  83
  84 struct z3fold_buddy_slots {
  85         /*
  86          * we are using BUDDY_MASK in handle_to_buddy etc. so there should
  87          * be enough slots to hold all possible variants
  88          */
  89         unsigned long slot[BUDDY_MASK + 1];
  90         unsigned long pool; /* back link + flags */
  91 };
  92 #define HANDLE_FLAG_MASK        (0x03)
  93
  94 /*
  95  * struct z3fold_header - z3fold page metadata occupying first chunks of each
  96  *                      z3fold page, except for HEADLESS pages
  97  * @buddy:              links the z3fold page into the relevant list in the
  98  *                      pool
  99  * @page_lock:          per-page lock
 100  * @refcount:           reference count for the z3fold page
 101  * @work:               work_struct for page layout optimization
 102  * @slots:              pointer to the structure holding buddy slots
 103  * @cpu:                CPU which this page "belongs" to
 104  * @first_chunks:       the size of the first buddy in chunks, 0 if free
 105  * @middle_chunks:      the size of the middle buddy in chunks, 0 if free
 106  * @last_chunks:        the size of the last buddy in chunks, 0 if free
 107  * @first_num:          the starting number (for the first handle)
 108  * @mapped_count:       the number of objects currently mapped
 109  */
 110 struct z3fold_header {
 111         struct list_head buddy;
 112         spinlock_t page_lock;
 113         struct kref refcount;
 114         struct work_struct work;
 115         struct z3fold_buddy_slots *slots;
 116         short cpu;
 117         unsigned short first_chunks;
 118         unsigned short middle_chunks;
 119         unsigned short last_chunks;
 120         unsigned short start_middle;
 121         unsigned short first_num:2;
 122         unsigned short mapped_count:2;
 123 };
 124
 125 /**
 126  * struct z3fold_pool - stores metadata for each z3fold pool
 127  * @name:       pool name
 128  * @lock:       protects pool unbuddied/lru lists
 129  * @stale_lock: protects pool stale page list
 130  * @unbuddied:  per-cpu array of lists tracking z3fold pages that contain 2-
 131  *              buddies; the list each z3fold page is added to depends on
 132  *              the size of its free region.
 133  * @lru:        list tracking the z3fold pages in LRU order by most recently
 134  *              added buddy.
 135  * @stale:      list of pages marked for freeing
 136  * @pages_nr:   number of z3fold pages in the pool.
 137  * @c_handle:   cache for z3fold_buddy_slots allocation
 138  * @ops:        pointer to a structure of user defined operations specified at
 139  *              pool creation time.
 140  * @compact_wq: workqueue for page layout background optimization
 141  * @release_wq: workqueue for safe page release
 142  * @work:       work_struct for safe page release
 143  * @inode:      inode for z3fold pseudo filesystem
 144  *
 145  * This structure is allocated at pool creation time and maintains metadata
 146  * pertaining to a particular z3fold pool.
 147  */
 148 struct z3fold_pool {
 149         const char *name;
 150         spinlock_t lock;
 151         spinlock_t stale_lock;
 152         struct list_head *unbuddied;
 153         struct list_head lru;
 154         struct list_head stale;
 155         atomic64_t pages_nr;
 156         struct kmem_cache *c_handle;
 157         const struct z3fold_ops *ops;
 158         struct zpool *zpool;
 159         const struct zpool_ops *zpool_ops;
 160         struct workqueue_struct *compact_wq;
 161         struct workqueue_struct *release_wq;
 162         struct work_struct work;
 163         struct inode *inode;
 164 };
 165
 166 /*
 167  * Internal z3fold page flags
 168  */
 169 enum z3fold_page_flags {
 170         PAGE_HEADLESS = 0,
 171         MIDDLE_CHUNK_MAPPED,
 172         NEEDS_COMPACTING,
 173         PAGE_STALE,
 174         PAGE_CLAIMED, /* by either reclaim or free */
 175 };
 176
 177 /*****************
 178  * Helpers
 179 *****************/
 180
 181 /* Converts an allocation size in bytes to size in z3fold chunks */
 182 static int size_to_chunks(size_t size)
 183 {
 184         return (size + CHUNK_SIZE - 1) >> CHUNK_SHIFT;
 185 }
 186
 187 #define for_each_unbuddied_list(_iter, _begin) \
 188         for ((_iter) = (_begin); (_iter) < NCHUNKS; (_iter)++)
 189
 190 static void compact_page_work(struct work_struct *w);
 191
 192 static inline struct z3fold_buddy_slots *alloc_slots(struct z3fold_pool *pool)
 193 {
 194         struct z3fold_buddy_slots *slots = kmem_cache_alloc(pool->c_handle,
 195                                                         GFP_KERNEL);
 196
 197         if (slots) {
 198                 memset(slots->slot, 0, sizeof(slots->slot));
 199                 slots->pool = (unsigned long)pool;
 200         }
 201
 202         return slots;
 203 }
 204
 205 static inline struct z3fold_pool *slots_to_pool(struct z3fold_buddy_slots *s)
 206 {
 207         return (struct z3fold_pool *)(s->pool & ~HANDLE_FLAG_MASK);
 208 }
 209
 210 static inline struct z3fold_buddy_slots *handle_to_slots(unsigned long handle)
 211 {
 212         return (struct z3fold_buddy_slots *)(handle & ~(SLOTS_ALIGN - 1));
 213 }
 214
 215 static inline void free_handle(unsigned long handle)
 216 {
 217         struct z3fold_buddy_slots *slots;
 218         int i;
 219         bool is_free;
 220
 221         if (handle & (1 << PAGE_HEADLESS))
 222                 return;
 223
 224         WARN_ON(*(unsigned long *)handle == 0);
 225         *(unsigned long *)handle = 0;
 226         slots = handle_to_slots(handle);
 227         is_free = true;
 228         for (i = 0; i <= BUDDY_MASK; i++) {
 229                 if (slots->slot[i]) {
 230                         is_free = false;
 231                         break;
 232                 }
 233         }
 234
 235         if (is_free) {
 236                 struct z3fold_pool *pool = slots_to_pool(slots);
 237
 238                 kmem_cache_free(pool->c_handle, slots);
 239         }
 240 }
 241
 242 static struct dentry *z3fold_do_mount(struct file_system_type *fs_type,
 243                                 int flags, const char *dev_name, void *data)
 244 {
 245         static const struct dentry_operations ops = {
 246                 .d_dname = simple_dname,
 247         };
 248
 249         return mount_pseudo(fs_type, "z3fold:", NULL, &ops, 0x33);
 250 }
 251
 252 static struct file_system_type z3fold_fs = {
 253         .name           = "z3fold",
 254         .mount          = z3fold_do_mount,
 255         .kill_sb        = kill_anon_super,
 256 };
 257
 258 static struct vfsmount *z3fold_mnt;
 259 static int z3fold_mount(void)
 260 {
 261         int ret = 0;
 262
 263         z3fold_mnt = kern_mount(&z3fold_fs);
 264         if (IS_ERR(z3fold_mnt))
 265                 ret = PTR_ERR(z3fold_mnt);
 266
 267         return ret;
 268 }
 269
 270 static void z3fold_unmount(void)
 271 {
 272         kern_unmount(z3fold_mnt);
 273 }
 274
 275 static const struct address_space_operations z3fold_aops;
 276 static int z3fold_register_migration(struct z3fold_pool *pool)
 277 {
 278         pool->inode = alloc_anon_inode(z3fold_mnt->mnt_sb);
 279         if (IS_ERR(pool->inode)) {
 280                 pool->inode = NULL;
 281                 return 1;
 282         }
 283
 284         pool->inode->i_mapping->private_data = pool;
 285         pool->inode->i_mapping->a_ops = &z3fold_aops;
 286         return 0;
 287 }
 288
 289 static void z3fold_unregister_migration(struct z3fold_pool *pool)
 290 {
 291         if (pool->inode)
 292                 iput(pool->inode);
 293  }
 294
 295 /* Initializes the z3fold header of a newly allocated z3fold page */
 296 static struct z3fold_header *init_z3fold_page(struct page *page,
 297                                         struct z3fold_pool *pool)
 298 {
 299         struct z3fold_header *zhdr = page_address(page);
 300         struct z3fold_buddy_slots *slots = alloc_slots(pool);
 301
 302         if (!slots)
 303                 return NULL;
 304
 305         INIT_LIST_HEAD(&page->lru);
 306         clear_bit(PAGE_HEADLESS, &page->private);
 307         clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
 308         clear_bit(NEEDS_COMPACTING, &page->private);
 309         clear_bit(PAGE_STALE, &page->private);
 310         clear_bit(PAGE_CLAIMED, &page->private);
 311
 312         spin_lock_init(&zhdr->page_lock);
 313         kref_init(&zhdr->refcount);
 314         zhdr->first_chunks = 0;
 315         zhdr->middle_chunks = 0;
 316         zhdr->last_chunks = 0;
 317         zhdr->first_num = 0;
 318         zhdr->start_middle = 0;
 319         zhdr->cpu = -1;
 320         zhdr->slots = slots;
 321         INIT_LIST_HEAD(&zhdr->buddy);
 322         INIT_WORK(&zhdr->work, compact_page_work);
 323         return zhdr;
 324 }
 325
 326 /* Resets the struct page fields and frees the page */
 327 static void free_z3fold_page(struct page *page, bool headless)
 328 {
 329         if (!headless) {
 330                 lock_page(page);
 331                 __ClearPageMovable(page);
 332                 unlock_page(page);
 333         }
 334         ClearPagePrivate(page);
 335         __free_page(page);
 336 }
 337
 338 /* Lock a z3fold page */
 339 static inline void z3fold_page_lock(struct z3fold_header *zhdr)
 340 {
 341         spin_lock(&zhdr->page_lock);
 342 }
 343
 344 /* Try to lock a z3fold page */
 345 static inline int z3fold_page_trylock(struct z3fold_header *zhdr)
 346 {
 347         return spin_trylock(&zhdr->page_lock);
 348 }
 349
 350 /* Unlock a z3fold page */
 351 static inline void z3fold_page_unlock(struct z3fold_header *zhdr)
 352 {
 353         spin_unlock(&zhdr->page_lock);
 354 }
 355
 356 /* Helper function to build the index */
 357 static inline int __idx(struct z3fold_header *zhdr, enum buddy bud)
 358 {
 359         return (bud + zhdr->first_num) & BUDDY_MASK;
 360 }
 361
 362 /*
 363  * Encodes the handle of a particular buddy within a z3fold page
 364  * Pool lock should be held as this function accesses first_num
 365  */
 366 static unsigned long encode_handle(struct z3fold_header *zhdr, enum buddy bud)
 367 {
 368         struct z3fold_buddy_slots *slots;
 369         unsigned long h = (unsigned long)zhdr;
 370         int idx = 0;
 371
 372         /*
 373          * For a headless page, its handle is its pointer with the extra
 374          * PAGE_HEADLESS bit set
 375          */
 376         if (bud == HEADLESS)
 377                 return h | (1 << PAGE_HEADLESS);
 378
 379         /* otherwise, return pointer to encoded handle */
 380         idx = __idx(zhdr, bud);
 381         h += idx;
 382         if (bud == LAST)
 383                 h |= (zhdr->last_chunks << BUDDY_SHIFT);
 384
 385         slots = zhdr->slots;
 386         slots->slot[idx] = h;
 387         return (unsigned long)&slots->slot[idx];
 388 }
 389
 390 /* Returns the z3fold page where a given handle is stored */
 391 static inline struct z3fold_header *handle_to_z3fold_header(unsigned long h)
 392 {
 393         unsigned long addr = h;
 394
 395         if (!(addr & (1 << PAGE_HEADLESS)))
 396                 addr = *(unsigned long *)h;
 397
 398         return (struct z3fold_header *)(addr & PAGE_MASK);
 399 }
 400
 401 /* only for LAST bud, returns zero otherwise */
 402 static unsigned short handle_to_chunks(unsigned long handle)
 403 {
 404         unsigned long addr = *(unsigned long *)handle;
 405
 406         return (addr & ~PAGE_MASK) >> BUDDY_SHIFT;
 407 }
 408
 409 /*
 410  * (handle & BUDDY_MASK) < zhdr->first_num is possible in encode_handle
 411  *  but that doesn't matter. because the masking will result in the
 412  *  correct buddy number.
 413  */
 414 static enum buddy handle_to_buddy(unsigned long handle)
 415 {
 416         struct z3fold_header *zhdr;
 417         unsigned long addr;
 418
 419         WARN_ON(handle & (1 << PAGE_HEADLESS));
 420         addr = *(unsigned long *)handle;
 421         zhdr = (struct z3fold_header *)(addr & PAGE_MASK);
 422         return (addr - zhdr->first_num) & BUDDY_MASK;
 423 }
 424
 425 static inline struct z3fold_pool *zhdr_to_pool(struct z3fold_header *zhdr)
 426 {
 427         return slots_to_pool(zhdr->slots);
 428 }
 429
 430 static void __release_z3fold_page(struct z3fold_header *zhdr, bool locked)
 431 {
 432         struct page *page = virt_to_page(zhdr);
 433         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 434
 435         WARN_ON(!list_empty(&zhdr->buddy));
 436         set_bit(PAGE_STALE, &page->private);
 437         clear_bit(NEEDS_COMPACTING, &page->private);
 438         spin_lock(&pool->lock);
 439         if (!list_empty(&page->lru))
 440                 list_del_init(&page->lru);
 441         spin_unlock(&pool->lock);
 442         if (locked)
 443                 z3fold_page_unlock(zhdr);
 444         spin_lock(&pool->stale_lock);
 445         list_add(&zhdr->buddy, &pool->stale);
 446         queue_work(pool->release_wq, &pool->work);
 447         spin_unlock(&pool->stale_lock);
 448 }
 449
 450 static void __attribute__((__unused__))
 451                         release_z3fold_page(struct kref *ref)
 452 {
 453         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 454                                                 refcount);
 455         __release_z3fold_page(zhdr, false);
 456 }
 457
 458 static void release_z3fold_page_locked(struct kref *ref)
 459 {
 460         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 461                                                 refcount);
 462         WARN_ON(z3fold_page_trylock(zhdr));
 463         __release_z3fold_page(zhdr, true);
 464 }
 465
 466 static void release_z3fold_page_locked_list(struct kref *ref)
 467 {
 468         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 469                                                refcount);
 470         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 471         spin_lock(&pool->lock);
 472         list_del_init(&zhdr->buddy);
 473         spin_unlock(&pool->lock);
 474
 475         WARN_ON(z3fold_page_trylock(zhdr));
 476         __release_z3fold_page(zhdr, true);
 477 }
 478
 479 static void free_pages_work(struct work_struct *w)
 480 {
 481         struct z3fold_pool *pool = container_of(w, struct z3fold_pool, work);
 482
 483         spin_lock(&pool->stale_lock);
 484         while (!list_empty(&pool->stale)) {
 485                 struct z3fold_header *zhdr = list_first_entry(&pool->stale,
 486                                                 struct z3fold_header, buddy);
 487                 struct page *page = virt_to_page(zhdr);
 488
 489                 list_del(&zhdr->buddy);
 490                 if (WARN_ON(!test_bit(PAGE_STALE, &page->private)))
 491                         continue;
 492                 spin_unlock(&pool->stale_lock);
 493                 cancel_work_sync(&zhdr->work);
 494                 free_z3fold_page(page, false);
 495                 cond_resched();
 496                 spin_lock(&pool->stale_lock);
 497         }
 498         spin_unlock(&pool->stale_lock);
 499 }
 500
 501 /*
 502  * Returns the number of free chunks in a z3fold page.
 503  * NB: can't be used with HEADLESS pages.
 504  */
 505 static int num_free_chunks(struct z3fold_header *zhdr)
 506 {
 507         int nfree;
 508         /*
 509          * If there is a middle object, pick up the bigger free space
 510          * either before or after it. Otherwise just subtract the number
 511          * of chunks occupied by the first and the last objects.
 512          */
 513         if (zhdr->middle_chunks != 0) {
 514                 int nfree_before = zhdr->first_chunks ?
 515                         0 : zhdr->start_middle - ZHDR_CHUNKS;
 516                 int nfree_after = zhdr->last_chunks ?
 517                         0 : TOTAL_CHUNKS -
 518                                 (zhdr->start_middle + zhdr->middle_chunks);
 519                 nfree = max(nfree_before, nfree_after);
 520         } else
 521                 nfree = NCHUNKS - zhdr->first_chunks - zhdr->last_chunks;
 522         return nfree;
 523 }
 524
 525 /* Add to the appropriate unbuddied list */
 526 static inline void add_to_unbuddied(struct z3fold_pool *pool,
 527                                 struct z3fold_header *zhdr)
 528 {
 529         if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0 ||
 530                         zhdr->middle_chunks == 0) {
 531                 struct list_head *unbuddied = get_cpu_ptr(pool->unbuddied);
 532
 533                 int freechunks = num_free_chunks(zhdr);
 534                 spin_lock(&pool->lock);
 535                 list_add(&zhdr->buddy, &unbuddied[freechunks]);
 536                 spin_unlock(&pool->lock);
 537                 zhdr->cpu = smp_processor_id();
 538                 put_cpu_ptr(pool->unbuddied);
 539         }
 540 }
 541
 542 static inline void *mchunk_memmove(struct z3fold_header *zhdr,
 543                                 unsigned short dst_chunk)
 544 {
 545         void *beg = zhdr;
 546         return memmove(beg + (dst_chunk << CHUNK_SHIFT),
 547                        beg + (zhdr->start_middle << CHUNK_SHIFT),
 548                        zhdr->middle_chunks << CHUNK_SHIFT);
 549 }
 550
 551 #define BIG_CHUNK_GAP   3
 552 /* Has to be called with lock held */
 553 static int z3fold_compact_page(struct z3fold_header *zhdr)
 554 {
 555         struct page *page = virt_to_page(zhdr);
 556
 557         if (test_bit(MIDDLE_CHUNK_MAPPED, &page->private))
 558                 return 0; /* can't move middle chunk, it's used */
 559
 560         if (unlikely(PageIsolated(page)))
 561                 return 0;
 562
 563         if (zhdr->middle_chunks == 0)
 564                 return 0; /* nothing to compact */
 565
 566         if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
 567                 /* move to the beginning */
 568                 mchunk_memmove(zhdr, ZHDR_CHUNKS);
 569                 zhdr->first_chunks = zhdr->middle_chunks;
 570                 zhdr->middle_chunks = 0;
 571                 zhdr->start_middle = 0;
 572                 zhdr->first_num++;
 573                 return 1;
 574         }
 575
 576         /*
 577          * moving data is expensive, so let's only do that if
 578          * there's substantial gain (at least BIG_CHUNK_GAP chunks)
 579          */
 580         if (zhdr->first_chunks != 0 && zhdr->last_chunks == 0 &&
 581             zhdr->start_middle - (zhdr->first_chunks + ZHDR_CHUNKS) >=
 582                         BIG_CHUNK_GAP) {
 583                 mchunk_memmove(zhdr, zhdr->first_chunks + ZHDR_CHUNKS);
 584                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
 585                 return 1;
 586         } else if (zhdr->last_chunks != 0 && zhdr->first_chunks == 0 &&
 587                    TOTAL_CHUNKS - (zhdr->last_chunks + zhdr->start_middle
 588                                         + zhdr->middle_chunks) >=
 589                         BIG_CHUNK_GAP) {
 590                 unsigned short new_start = TOTAL_CHUNKS - zhdr->last_chunks -
 591                         zhdr->middle_chunks;
 592                 mchunk_memmove(zhdr, new_start);
 593                 zhdr->start_middle = new_start;
 594                 return 1;
 595         }
 596
 597         return 0;
 598 }
 599
 600 static void do_compact_page(struct z3fold_header *zhdr, bool locked)
 601 {
 602         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 603         struct page *page;
 604
 605         page = virt_to_page(zhdr);
 606         if (locked)
 607                 WARN_ON(z3fold_page_trylock(zhdr));
 608         else
 609                 z3fold_page_lock(zhdr);
 610         if (WARN_ON(!test_and_clear_bit(NEEDS_COMPACTING, &page->private))) {
 611                 z3fold_page_unlock(zhdr);
 612                 return;
 613         }
 614         spin_lock(&pool->lock);
 615         list_del_init(&zhdr->buddy);
 616         spin_unlock(&pool->lock);
 617
 618         if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
 619                 atomic64_dec(&pool->pages_nr);
 620                 return;
 621         }
 622
 623         if (unlikely(PageIsolated(page) ||
 624                      test_bit(PAGE_STALE, &page->private))) {
 625                 z3fold_page_unlock(zhdr);
 626                 return;
 627         }
 628
 629         z3fold_compact_page(zhdr);
 630         add_to_unbuddied(pool, zhdr);
 631         z3fold_page_unlock(zhdr);
 632 }
 633
 634 static void compact_page_work(struct work_struct *w)
 635 {
 636         struct z3fold_header *zhdr = container_of(w, struct z3fold_header,
 637                                                 work);
 638
 639         do_compact_page(zhdr, false);
 640 }
 641
 642 /* returns _locked_ z3fold page header or NULL */
 643 static inline struct z3fold_header *__z3fold_alloc(struct z3fold_pool *pool,
 644                                                 size_t size, bool can_sleep)
 645 {
 646         struct z3fold_header *zhdr = NULL;
 647         struct page *page;
 648         struct list_head *unbuddied;
 649         int chunks = size_to_chunks(size), i;
 650
 651 lookup:
 652         /* First, try to find an unbuddied z3fold page. */
 653         unbuddied = get_cpu_ptr(pool->unbuddied);
 654         for_each_unbuddied_list(i, chunks) {
 655                 struct list_head *l = &unbuddied[i];
 656
 657                 zhdr = list_first_entry_or_null(READ_ONCE(l),
 658                                         struct z3fold_header, buddy);
 659
 660                 if (!zhdr)
 661                         continue;
 662
 663                 /* Re-check under lock. */
 664                 spin_lock(&pool->lock);
 665                 l = &unbuddied[i];
 666                 if (unlikely(zhdr != list_first_entry(READ_ONCE(l),
 667                                                 struct z3fold_header, buddy)) ||
 668                     !z3fold_page_trylock(zhdr)) {
 669                         spin_unlock(&pool->lock);
 670                         zhdr = NULL;
 671                         put_cpu_ptr(pool->unbuddied);
 672                         if (can_sleep)
 673                                 cond_resched();
 674                         goto lookup;
 675                 }
 676                 list_del_init(&zhdr->buddy);
 677                 zhdr->cpu = -1;
 678                 spin_unlock(&pool->lock);
 679
 680                 page = virt_to_page(zhdr);
 681                 if (test_bit(NEEDS_COMPACTING, &page->private)) {
 682                         z3fold_page_unlock(zhdr);
 683                         zhdr = NULL;
 684                         put_cpu_ptr(pool->unbuddied);
 685                         if (can_sleep)
 686                                 cond_resched();
 687                         goto lookup;
 688                 }
 689
 690                 /*
 691                  * this page could not be removed from its unbuddied
 692                  * list while pool lock was held, and then we've taken
 693                  * page lock so kref_put could not be called before
 694                  * we got here, so it's safe to just call kref_get()
 695                  */
 696                 kref_get(&zhdr->refcount);
 697                 break;
 698         }
 699         put_cpu_ptr(pool->unbuddied);
 700
 701         if (!zhdr) {
 702                 int cpu;
 703
 704                 /* look for _exact_ match on other cpus' lists */
 705                 for_each_online_cpu(cpu) {
 706                         struct list_head *l;
 707
 708                         unbuddied = per_cpu_ptr(pool->unbuddied, cpu);
 709                         spin_lock(&pool->lock);
 710                         l = &unbuddied[chunks];
 711
 712                         zhdr = list_first_entry_or_null(READ_ONCE(l),
 713                                                 struct z3fold_header, buddy);
 714
 715                         if (!zhdr || !z3fold_page_trylock(zhdr)) {
 716                                 spin_unlock(&pool->lock);
 717                                 zhdr = NULL;
 718                                 continue;
 719                         }
 720                         list_del_init(&zhdr->buddy);
 721                         zhdr->cpu = -1;
 722                         spin_unlock(&pool->lock);
 723
 724                         page = virt_to_page(zhdr);
 725                         if (test_bit(NEEDS_COMPACTING, &page->private)) {
 726                                 z3fold_page_unlock(zhdr);
 727                                 zhdr = NULL;
 728                                 if (can_sleep)
 729                                         cond_resched();
 730                                 continue;
 731                         }
 732                         kref_get(&zhdr->refcount);
 733                         break;
 734                 }
 735         }
 736
 737         return zhdr;
 738 }
 739
 740 /*
 741  * API Functions
 742  */
 743
 744 /**
 745  * z3fold_create_pool() - create a new z3fold pool
 746  * @name:       pool name
 747  * @gfp:        gfp flags when allocating the z3fold pool structure
 748  * @ops:        user-defined operations for the z3fold pool
 749  *
 750  * Return: pointer to the new z3fold pool or NULL if the metadata allocation
 751  * failed.
 752  */
 753 static struct z3fold_pool *z3fold_create_pool(const char *name, gfp_t gfp,
 754                 const struct z3fold_ops *ops)
 755 {
 756         struct z3fold_pool *pool = NULL;
 757         int i, cpu;
 758
 759         pool = kzalloc(sizeof(struct z3fold_pool), gfp);
 760         if (!pool)
 761                 goto out;
 762         pool->c_handle = kmem_cache_create("z3fold_handle",
 763                                 sizeof(struct z3fold_buddy_slots),
 764                                 SLOTS_ALIGN, 0, NULL);
 765         if (!pool->c_handle)
 766                 goto out_c;
 767         spin_lock_init(&pool->lock);
 768         spin_lock_init(&pool->stale_lock);
 769         pool->unbuddied = __alloc_percpu(sizeof(struct list_head)*NCHUNKS, 2);
 770         if (!pool->unbuddied)
 771                 goto out_pool;
 772         for_each_possible_cpu(cpu) {
 773                 struct list_head *unbuddied =
 774                                 per_cpu_ptr(pool->unbuddied, cpu);
 775                 for_each_unbuddied_list(i, 0)
 776                         INIT_LIST_HEAD(&unbuddied[i]);
 777         }
 778         INIT_LIST_HEAD(&pool->lru);
 779         INIT_LIST_HEAD(&pool->stale);
 780         atomic64_set(&pool->pages_nr, 0);
 781         pool->name = name;
 782         pool->compact_wq = create_singlethread_workqueue(pool->name);
 783         if (!pool->compact_wq)
 784                 goto out_unbuddied;
 785         pool->release_wq = create_singlethread_workqueue(pool->name);
 786         if (!pool->release_wq)
 787                 goto out_wq;
 788         if (z3fold_register_migration(pool))
 789                 goto out_rwq;
 790         INIT_WORK(&pool->work, free_pages_work);
 791         pool->ops = ops;
 792         return pool;
 793
 794 out_rwq:
 795         destroy_workqueue(pool->release_wq);
 796 out_wq:
 797         destroy_workqueue(pool->compact_wq);
 798 out_unbuddied:
 799         free_percpu(pool->unbuddied);
 800 out_pool:
 801         kmem_cache_destroy(pool->c_handle);
 802 out_c:
 803         kfree(pool);
 804 out:
 805         return NULL;
 806 }
 807
 808 /**
 809  * z3fold_destroy_pool() - destroys an existing z3fold pool
 810  * @pool:       the z3fold pool to be destroyed
 811  *
 812  * The pool should be emptied before this function is called.
 813  */
 814 static void z3fold_destroy_pool(struct z3fold_pool *pool)
 815 {
 816         kmem_cache_destroy(pool->c_handle);
 817         z3fold_unregister_migration(pool);
 818         destroy_workqueue(pool->release_wq);
 819         destroy_workqueue(pool->compact_wq);
 820         kfree(pool);
 821 }
 822
 823 /**
 824  * z3fold_alloc() - allocates a region of a given size
 825  * @pool:       z3fold pool from which to allocate
 826  * @size:       size in bytes of the desired allocation
 827  * @gfp:        gfp flags used if the pool needs to grow
 828  * @handle:     handle of the new allocation
 829  *
 830  * This function will attempt to find a free region in the pool large enough to
 831  * satisfy the allocation request.  A search of the unbuddied lists is
 832  * performed first. If no suitable free region is found, then a new page is
 833  * allocated and added to the pool to satisfy the request.
 834  *
 835  * gfp should not set __GFP_HIGHMEM as highmem pages cannot be used
 836  * as z3fold pool pages.
 837  *
 838  * Return: 0 if success and handle is set, otherwise -EINVAL if the size or
 839  * gfp arguments are invalid or -ENOMEM if the pool was unable to allocate
 840  * a new page.
 841  */
 842 static int z3fold_alloc(struct z3fold_pool *pool, size_t size, gfp_t gfp,
 843                         unsigned long *handle)
 844 {
 845         int chunks = size_to_chunks(size);
 846         struct z3fold_header *zhdr = NULL;
 847         struct page *page = NULL;
 848         enum buddy bud;
 849         bool can_sleep = gfpflags_allow_blocking(gfp);
 850
 851         if (!size || (gfp & __GFP_HIGHMEM))
 852                 return -EINVAL;
 853
 854         if (size > PAGE_SIZE)
 855                 return -ENOSPC;
 856
 857         if (size > PAGE_SIZE - ZHDR_SIZE_ALIGNED - CHUNK_SIZE)
 858                 bud = HEADLESS;
 859         else {
 860 retry:
 861                 zhdr = __z3fold_alloc(pool, size, can_sleep);
 862                 if (zhdr) {
 863                         if (zhdr->first_chunks == 0) {
 864                                 if (zhdr->middle_chunks != 0 &&
 865                                     chunks >= zhdr->start_middle)
 866                                         bud = LAST;
 867                                 else
 868                                         bud = FIRST;
 869                         } else if (zhdr->last_chunks == 0)
 870                                 bud = LAST;
 871                         else if (zhdr->middle_chunks == 0)
 872                                 bud = MIDDLE;
 873                         else {
 874                                 if (kref_put(&zhdr->refcount,
 875                                              release_z3fold_page_locked))
 876                                         atomic64_dec(&pool->pages_nr);
 877                                 else
 878                                         z3fold_page_unlock(zhdr);
 879                                 pr_err("No free chunks in unbuddied\n");
 880                                 WARN_ON(1);
 881                                 goto retry;
 882                         }
 883                         page = virt_to_page(zhdr);
 884                         goto found;
 885                 }
 886                 bud = FIRST;
 887         }
 888
 889         page = NULL;
 890         if (can_sleep) {
 891                 spin_lock(&pool->stale_lock);
 892                 zhdr = list_first_entry_or_null(&pool->stale,
 893                                                 struct z3fold_header, buddy);
 894                 /*
 895                  * Before allocating a page, let's see if we can take one from
 896                  * the stale pages list. cancel_work_sync() can sleep so we
 897                  * limit this case to the contexts where we can sleep
 898                  */
 899                 if (zhdr) {
 900                         list_del(&zhdr->buddy);
 901                         spin_unlock(&pool->stale_lock);
 902                         cancel_work_sync(&zhdr->work);
 903                         page = virt_to_page(zhdr);
 904                 } else {
 905                         spin_unlock(&pool->stale_lock);
 906                 }
 907         }
 908         if (!page)
 909                 page = alloc_page(gfp);
 910
 911         if (!page)
 912                 return -ENOMEM;
 913
 914         zhdr = init_z3fold_page(page, pool);
 915         if (!zhdr) {
 916                 __free_page(page);
 917                 return -ENOMEM;
 918         }
 919         atomic64_inc(&pool->pages_nr);
 920
 921         if (bud == HEADLESS) {
 922                 set_bit(PAGE_HEADLESS, &page->private);
 923                 goto headless;
 924         }
 925         __SetPageMovable(page, pool->inode->i_mapping);
 926         z3fold_page_lock(zhdr);
 927
 928 found:
 929         if (bud == FIRST)
 930                 zhdr->first_chunks = chunks;
 931         else if (bud == LAST)
 932                 zhdr->last_chunks = chunks;
 933         else {
 934                 zhdr->middle_chunks = chunks;
 935                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
 936         }
 937         add_to_unbuddied(pool, zhdr);
 938
 939 headless:
 940         spin_lock(&pool->lock);
 941         /* Add/move z3fold page to beginning of LRU */
 942         if (!list_empty(&page->lru))
 943                 list_del(&page->lru);
 944
 945         list_add(&page->lru, &pool->lru);
 946
 947         *handle = encode_handle(zhdr, bud);
 948         spin_unlock(&pool->lock);
 949         if (bud != HEADLESS)
 950                 z3fold_page_unlock(zhdr);
 951
 952         return 0;
 953 }
 954
 955 /**
 956  * z3fold_free() - frees the allocation associated with the given handle
 957  * @pool:       pool in which the allocation resided
 958  * @handle:     handle associated with the allocation returned by z3fold_alloc()
 959  *
 960  * In the case that the z3fold page in which the allocation resides is under
 961  * reclaim, as indicated by the PG_reclaim flag being set, this function
 962  * only sets the first|last_chunks to 0.  The page is actually freed
 963  * once both buddies are evicted (see z3fold_reclaim_page() below).
 964  */
 965 static void z3fold_free(struct z3fold_pool *pool, unsigned long handle)
 966 {
 967         struct z3fold_header *zhdr;
 968         struct page *page;
 969         enum buddy bud;
 970
 971         zhdr = handle_to_z3fold_header(handle);
 972         page = virt_to_page(zhdr);
 973
 974         if (test_bit(PAGE_HEADLESS, &page->private)) {
 975                 /* if a headless page is under reclaim, just leave.
 976                  * NB: we use test_and_set_bit for a reason: if the bit
 977                  * has not been set before, we release this page
 978                  * immediately so we don't care about its value any more.
 979                  */
 980                 if (!test_and_set_bit(PAGE_CLAIMED, &page->private)) {
 981                         spin_lock(&pool->lock);
 982                         list_del(&page->lru);
 983                         spin_unlock(&pool->lock);
 984                         free_z3fold_page(page, true);
 985                         atomic64_dec(&pool->pages_nr);
 986                 }
 987                 return;
 988         }
 989
 990         /* Non-headless case */
 991         z3fold_page_lock(zhdr);
 992         bud = handle_to_buddy(handle);
 993
 994         switch (bud) {
 995         case FIRST:
 996                 zhdr->first_chunks = 0;
 997                 break;
 998         case MIDDLE:
 999                 zhdr->middle_chunks = 0;
1000                 break;
1001         case LAST:
1002                 zhdr->last_chunks = 0;
1003                 break;
1004         default:
1005                 pr_err("%s: unknown bud %d\n", __func__, bud);
1006                 WARN_ON(1);
1007                 z3fold_page_unlock(zhdr);
1008                 return;
1009         }
1010
1011         free_handle(handle);
1012         if (kref_put(&zhdr->refcount, release_z3fold_page_locked_list)) {
1013                 atomic64_dec(&pool->pages_nr);
1014                 return;
1015         }
1016         if (test_bit(PAGE_CLAIMED, &page->private)) {
1017                 z3fold_page_unlock(zhdr);
1018                 return;
1019         }
1020         if (unlikely(PageIsolated(page)) ||
1021             test_and_set_bit(NEEDS_COMPACTING, &page->private)) {
1022                 z3fold_page_unlock(zhdr);
1023                 return;
1024         }
1025         if (zhdr->cpu < 0 || !cpu_online(zhdr->cpu)) {
1026                 spin_lock(&pool->lock);
1027                 list_del_init(&zhdr->buddy);
1028                 spin_unlock(&pool->lock);
1029                 zhdr->cpu = -1;
1030                 kref_get(&zhdr->refcount);
1031                 do_compact_page(zhdr, true);
1032                 return;
1033         }
1034         kref_get(&zhdr->refcount);
1035         queue_work_on(zhdr->cpu, pool->compact_wq, &zhdr->work);
1036         z3fold_page_unlock(zhdr);
1037 }
1038
1039 /**
1040  * z3fold_reclaim_page() - evicts allocations from a pool page and frees it
1041  * @pool:       pool from which a page will attempt to be evicted
1042  * @retries:    number of pages on the LRU list for which eviction will
1043  *              be attempted before failing
1044  *
1045  * z3fold reclaim is different from normal system reclaim in that it is done
1046  * from the bottom, up. This is because only the bottom layer, z3fold, has
1047  * information on how the allocations are organized within each z3fold page.
1048  * This has the potential to create interesting locking situations between
1049  * z3fold and the user, however.
1050  *
1051  * To avoid these, this is how z3fold_reclaim_page() should be called:
1052  *
1053  * The user detects a page should be reclaimed and calls z3fold_reclaim_page().
1054  * z3fold_reclaim_page() will remove a z3fold page from the pool LRU list and
1055  * call the user-defined eviction handler with the pool and handle as
1056  * arguments.
1057  *
1058  * If the handle can not be evicted, the eviction handler should return
1059  * non-zero. z3fold_reclaim_page() will add the z3fold page back to the
1060  * appropriate list and try the next z3fold page on the LRU up to
1061  * a user defined number of retries.
1062  *
1063  * If the handle is successfully evicted, the eviction handler should
1064  * return 0 _and_ should have called z3fold_free() on the handle. z3fold_free()
1065  * contains logic to delay freeing the page if the page is under reclaim,
1066  * as indicated by the setting of the PG_reclaim flag on the underlying page.
1067  *
1068  * If all buddies in the z3fold page are successfully evicted, then the
1069  * z3fold page can be freed.
1070  *
1071  * Returns: 0 if page is successfully freed, otherwise -EINVAL if there are
1072  * no pages to evict or an eviction handler is not registered, -EAGAIN if
1073  * the retry limit was hit.
1074  */
1075 static int z3fold_reclaim_page(struct z3fold_pool *pool, unsigned int retries)
1076 {
1077         int i, ret = 0;
1078         struct z3fold_header *zhdr = NULL;
1079         struct page *page = NULL;
1080         struct list_head *pos;
1081         unsigned long first_handle = 0, middle_handle = 0, last_handle = 0;
1082
1083         spin_lock(&pool->lock);
1084         if (!pool->ops || !pool->ops->evict || retries == 0) {
1085                 spin_unlock(&pool->lock);
1086                 return -EINVAL;
1087         }
1088         for (i = 0; i < retries; i++) {
1089                 if (list_empty(&pool->lru)) {
1090                         spin_unlock(&pool->lock);
1091                         return -EINVAL;
1092                 }
1093                 list_for_each_prev(pos, &pool->lru) {
1094                         page = list_entry(pos, struct page, lru);
1095
1096                         /* this bit could have been set by free, in which case
1097                          * we pass over to the next page in the pool.
1098                          */
1099                         if (test_and_set_bit(PAGE_CLAIMED, &page->private))
1100                                 continue;
1101
1102                         if (unlikely(PageIsolated(page)))
1103                                 continue;
1104                         if (test_bit(PAGE_HEADLESS, &page->private))
1105                                 break;
1106
1107                         zhdr = page_address(page);
1108                         if (!z3fold_page_trylock(zhdr)) {
1109                                 zhdr = NULL;
1110                                 continue; /* can't evict at this point */
1111                         }
1112                         kref_get(&zhdr->refcount);
1113                         list_del_init(&zhdr->buddy);
1114                         zhdr->cpu = -1;
1115                         break;
1116                 }
1117
1118                 if (!zhdr)
1119                         break;
1120
1121                 list_del_init(&page->lru);
1122                 spin_unlock(&pool->lock);
1123
1124                 if (!test_bit(PAGE_HEADLESS, &page->private)) {
1125                         /*
1126                          * We need encode the handles before unlocking, since
1127                          * we can race with free that will set
1128                          * (first|last)_chunks to 0
1129                          */
1130                         first_handle = 0;
1131                         last_handle = 0;
1132                         middle_handle = 0;
1133                         if (zhdr->first_chunks)
1134                                 first_handle = encode_handle(zhdr, FIRST);
1135                         if (zhdr->middle_chunks)
1136                                 middle_handle = encode_handle(zhdr, MIDDLE);
1137                         if (zhdr->last_chunks)
1138                                 last_handle = encode_handle(zhdr, LAST);
1139                         /*
1140                          * it's safe to unlock here because we hold a
1141                          * reference to this page
1142                          */
1143                         z3fold_page_unlock(zhdr);
1144                 } else {
1145                         first_handle = encode_handle(zhdr, HEADLESS);
1146                         last_handle = middle_handle = 0;
1147                 }
1148
1149                 /* Issue the eviction callback(s) */
1150                 if (middle_handle) {
1151                         ret = pool->ops->evict(pool, middle_handle);
1152                         if (ret)
1153                                 goto next;
1154                 }
1155                 if (first_handle) {
1156                         ret = pool->ops->evict(pool, first_handle);
1157                         if (ret)
1158                                 goto next;
1159                 }
1160                 if (last_handle) {
1161                         ret = pool->ops->evict(pool, last_handle);
1162                         if (ret)
1163                                 goto next;
1164                 }
1165 next:
1166                 if (test_bit(PAGE_HEADLESS, &page->private)) {
1167                         if (ret == 0) {
1168                                 free_z3fold_page(page, true);
1169                                 atomic64_dec(&pool->pages_nr);
1170                                 return 0;
1171                         }
1172                         spin_lock(&pool->lock);
1173                         list_add(&page->lru, &pool->lru);
1174                         spin_unlock(&pool->lock);
1175                 } else {
1176                         z3fold_page_lock(zhdr);
1177                         clear_bit(PAGE_CLAIMED, &page->private);
1178                         if (kref_put(&zhdr->refcount,
1179                                         release_z3fold_page_locked)) {
1180                                 atomic64_dec(&pool->pages_nr);
1181                                 return 0;
1182                         }
1183                         /*
1184                          * if we are here, the page is still not completely
1185                          * free. Take the global pool lock then to be able
1186                          * to add it back to the lru list
1187                          */
1188                         spin_lock(&pool->lock);
1189                         list_add(&page->lru, &pool->lru);
1190                         spin_unlock(&pool->lock);
1191                         z3fold_page_unlock(zhdr);
1192                 }
1193
1194                 /* We started off locked to we need to lock the pool back */
1195                 spin_lock(&pool->lock);
1196         }
1197         spin_unlock(&pool->lock);
1198         return -EAGAIN;
1199 }
1200
1201 /**
1202  * z3fold_map() - maps the allocation associated with the given handle
1203  * @pool:       pool in which the allocation resides
1204  * @handle:     handle associated with the allocation to be mapped
1205  *
1206  * Extracts the buddy number from handle and constructs the pointer to the
1207  * correct starting chunk within the page.
1208  *
1209  * Returns: a pointer to the mapped allocation
1210  */
1211 static void *z3fold_map(struct z3fold_pool *pool, unsigned long handle)
1212 {
1213         struct z3fold_header *zhdr;
1214         struct page *page;
1215         void *addr;
1216         enum buddy buddy;
1217
1218         zhdr = handle_to_z3fold_header(handle);
1219         addr = zhdr;
1220         page = virt_to_page(zhdr);
1221
1222         if (test_bit(PAGE_HEADLESS, &page->private))
1223                 goto out;
1224
1225         z3fold_page_lock(zhdr);
1226         buddy = handle_to_buddy(handle);
1227         switch (buddy) {
1228         case FIRST:
1229                 addr += ZHDR_SIZE_ALIGNED;
1230                 break;
1231         case MIDDLE:
1232                 addr += zhdr->start_middle << CHUNK_SHIFT;
1233                 set_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1234                 break;
1235         case LAST:
1236                 addr += PAGE_SIZE - (handle_to_chunks(handle) << CHUNK_SHIFT);
1237                 break;
1238         default:
1239                 pr_err("unknown buddy id %d\n", buddy);
1240                 WARN_ON(1);
1241                 addr = NULL;
1242                 break;
1243         }
1244
1245         if (addr)
1246                 zhdr->mapped_count++;
1247         z3fold_page_unlock(zhdr);
1248 out:
1249         return addr;
1250 }
1251
1252 /**
1253  * z3fold_unmap() - unmaps the allocation associated with the given handle
1254  * @pool:       pool in which the allocation resides
1255  * @handle:     handle associated with the allocation to be unmapped
1256  */
1257 static void z3fold_unmap(struct z3fold_pool *pool, unsigned long handle)
1258 {
1259         struct z3fold_header *zhdr;
1260         struct page *page;
1261         enum buddy buddy;
1262
1263         zhdr = handle_to_z3fold_header(handle);
1264         page = virt_to_page(zhdr);
1265
1266         if (test_bit(PAGE_HEADLESS, &page->private))
1267                 return;
1268
1269         z3fold_page_lock(zhdr);
1270         buddy = handle_to_buddy(handle);
1271         if (buddy == MIDDLE)
1272                 clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1273         zhdr->mapped_count--;
1274         z3fold_page_unlock(zhdr);
1275 }
1276
1277 /**
1278  * z3fold_get_pool_size() - gets the z3fold pool size in pages
1279  * @pool:       pool whose size is being queried
1280  *
1281  * Returns: size in pages of the given pool.
1282  */
1283 static u64 z3fold_get_pool_size(struct z3fold_pool *pool)
1284 {
1285         return atomic64_read(&pool->pages_nr);
1286 }
1287
1288 static bool z3fold_page_isolate(struct page *page, isolate_mode_t mode)
1289 {
1290         struct z3fold_header *zhdr;
1291         struct z3fold_pool *pool;
1292
1293         VM_BUG_ON_PAGE(!PageMovable(page), page);
1294         VM_BUG_ON_PAGE(PageIsolated(page), page);
1295
1296         if (test_bit(PAGE_HEADLESS, &page->private))
1297                 return false;
1298
1299         zhdr = page_address(page);
1300         z3fold_page_lock(zhdr);
1301         if (test_bit(NEEDS_COMPACTING, &page->private) ||
1302             test_bit(PAGE_STALE, &page->private))
1303                 goto out;
1304
1305         pool = zhdr_to_pool(zhdr);
1306
1307         if (zhdr->mapped_count == 0) {
1308                 kref_get(&zhdr->refcount);
1309                 if (!list_empty(&zhdr->buddy))
1310                         list_del_init(&zhdr->buddy);
1311                 spin_lock(&pool->lock);
1312                 if (!list_empty(&page->lru))
1313                         list_del(&page->lru);
1314                 spin_unlock(&pool->lock);
1315                 z3fold_page_unlock(zhdr);
1316                 return true;
1317         }
1318 out:
1319         z3fold_page_unlock(zhdr);
1320         return false;
1321 }
1322
1323 static int z3fold_page_migrate(struct address_space *mapping, struct page *newpage,
1324                                struct page *page, enum migrate_mode mode)
1325 {
1326         struct z3fold_header *zhdr, *new_zhdr;
1327         struct z3fold_pool *pool;
1328         struct address_space *new_mapping;
1329
1330         VM_BUG_ON_PAGE(!PageMovable(page), page);
1331         VM_BUG_ON_PAGE(!PageIsolated(page), page);
1332
1333         zhdr = page_address(page);
1334         pool = zhdr_to_pool(zhdr);
1335
1336         if (!trylock_page(page))
1337                 return -EAGAIN;
1338
1339         if (!z3fold_page_trylock(zhdr)) {
1340                 unlock_page(page);
1341                 return -EAGAIN;
1342         }
1343         if (zhdr->mapped_count != 0) {
1344                 z3fold_page_unlock(zhdr);
1345                 unlock_page(page);
1346                 return -EBUSY;
1347         }
1348         new_zhdr = page_address(newpage);
1349         memcpy(new_zhdr, zhdr, PAGE_SIZE);
1350         newpage->private = page->private;
1351         page->private = 0;
1352         z3fold_page_unlock(zhdr);
1353         spin_lock_init(&new_zhdr->page_lock);
1354         new_mapping = page_mapping(page);
1355         __ClearPageMovable(page);
1356         ClearPagePrivate(page);
1357
1358         get_page(newpage);
1359         z3fold_page_lock(new_zhdr);
1360         if (new_zhdr->first_chunks)
1361                 encode_handle(new_zhdr, FIRST);
1362         if (new_zhdr->last_chunks)
1363                 encode_handle(new_zhdr, LAST);
1364         if (new_zhdr->middle_chunks)
1365                 encode_handle(new_zhdr, MIDDLE);
1366         set_bit(NEEDS_COMPACTING, &newpage->private);
1367         new_zhdr->cpu = smp_processor_id();
1368         spin_lock(&pool->lock);
1369         list_add(&newpage->lru, &pool->lru);
1370         spin_unlock(&pool->lock);
1371         __SetPageMovable(newpage, new_mapping);
1372         z3fold_page_unlock(new_zhdr);
1373
1374         queue_work_on(new_zhdr->cpu, pool->compact_wq, &new_zhdr->work);
1375
1376         page_mapcount_reset(page);
1377         unlock_page(page);
1378         put_page(page);
1379         return 0;
1380 }
1381
1382 static void z3fold_page_putback(struct page *page)
1383 {
1384         struct z3fold_header *zhdr;
1385         struct z3fold_pool *pool;
1386
1387         zhdr = page_address(page);
1388         pool = zhdr_to_pool(zhdr);
1389
1390         z3fold_page_lock(zhdr);
1391         if (!list_empty(&zhdr->buddy))
1392                 list_del_init(&zhdr->buddy);
1393         INIT_LIST_HEAD(&page->lru);
1394         if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
1395                 atomic64_dec(&pool->pages_nr);
1396                 return;
1397         }
1398         spin_lock(&pool->lock);
1399         list_add(&page->lru, &pool->lru);
1400         spin_unlock(&pool->lock);
1401         z3fold_page_unlock(zhdr);
1402 }
1403
1404 static const struct address_space_operations z3fold_aops = {
1405         .isolate_page = z3fold_page_isolate,
1406         .migratepage = z3fold_page_migrate,
1407         .putback_page = z3fold_page_putback,
1408 };
1409
1410 /*****************
1411  * zpool
1412  ****************/
1413
1414 static int z3fold_zpool_evict(struct z3fold_pool *pool, unsigned long handle)
1415 {
1416         if (pool->zpool && pool->zpool_ops && pool->zpool_ops->evict)
1417                 return pool->zpool_ops->evict(pool->zpool, handle);
1418         else
1419                 return -ENOENT;
1420 }
1421
1422 static const struct z3fold_ops z3fold_zpool_ops = {
1423         .evict =        z3fold_zpool_evict
1424 };
1425
1426 static void *z3fold_zpool_create(const char *name, gfp_t gfp,
1427                                const struct zpool_ops *zpool_ops,
1428                                struct zpool *zpool)
1429 {
1430         struct z3fold_pool *pool;
1431
1432         pool = z3fold_create_pool(name, gfp,
1433                                 zpool_ops ? &z3fold_zpool_ops : NULL);
1434         if (pool) {
1435                 pool->zpool = zpool;
1436                 pool->zpool_ops = zpool_ops;
1437         }
1438         return pool;
1439 }
1440
1441 static void z3fold_zpool_destroy(void *pool)
1442 {
1443         z3fold_destroy_pool(pool);
1444 }
1445
1446 static int z3fold_zpool_malloc(void *pool, size_t size, gfp_t gfp,
1447                         unsigned long *handle)
1448 {
1449         return z3fold_alloc(pool, size, gfp, handle);
1450 }
1451 static void z3fold_zpool_free(void *pool, unsigned long handle)
1452 {
1453         z3fold_free(pool, handle);
1454 }
1455
1456 static int z3fold_zpool_shrink(void *pool, unsigned int pages,
1457                         unsigned int *reclaimed)
1458 {
1459         unsigned int total = 0;
1460         int ret = -EINVAL;
1461
1462         while (total < pages) {
1463                 ret = z3fold_reclaim_page(pool, 8);
1464                 if (ret < 0)
1465                         break;
1466                 total++;
1467         }
1468
1469         if (reclaimed)
1470                 *reclaimed = total;
1471
1472         return ret;
1473 }
1474
1475 static void *z3fold_zpool_map(void *pool, unsigned long handle,
1476                         enum zpool_mapmode mm)
1477 {
1478         return z3fold_map(pool, handle);
1479 }
1480 static void z3fold_zpool_unmap(void *pool, unsigned long handle)
1481 {
1482         z3fold_unmap(pool, handle);
1483 }
1484
1485 static u64 z3fold_zpool_total_size(void *pool)
1486 {
1487         return z3fold_get_pool_size(pool) * PAGE_SIZE;
1488 }
1489
1490 static struct zpool_driver z3fold_zpool_driver = {
1491         .type =         "z3fold",
1492         .owner =        THIS_MODULE,
1493         .create =       z3fold_zpool_create,
1494         .destroy =      z3fold_zpool_destroy,
1495         .malloc =       z3fold_zpool_malloc,
1496         .free =         z3fold_zpool_free,
1497         .shrink =       z3fold_zpool_shrink,
1498         .map =          z3fold_zpool_map,
1499         .unmap =        z3fold_zpool_unmap,
1500         .total_size =   z3fold_zpool_total_size,
1501 };
1502
1503 MODULE_ALIAS("zpool-z3fold");
1504
1505 static int __init init_z3fold(void)
1506 {
1507         int ret;
1508
1509         /* Make sure the z3fold header is not larger than the page size */
1510         BUILD_BUG_ON(ZHDR_SIZE_ALIGNED > PAGE_SIZE);
1511         ret = z3fold_mount();
1512         if (ret)
1513                 return ret;
1514
1515         zpool_register_driver(&z3fold_zpool_driver);
1516
1517         return 0;
1518 }
1519
1520 static void __exit exit_z3fold(void)
1521 {
1522         z3fold_unmount();
1523         zpool_unregister_driver(&z3fold_zpool_driver);
1524 }
1525
1526 module_init(init_z3fold);
1527 module_exit(exit_z3fold);
1528
1529 MODULE_LICENSE("GPL");
1530 MODULE_AUTHOR("Vitaly Wool <vitalywool@gmail.com>");
1531 MODULE_DESCRIPTION("3-Fold Allocator for Compressed Pages");