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/list.h>
  27 #include <linux/mm.h>
  28 #include <linux/module.h>
  29 #include <linux/preempt.h>
  30 #include <linux/slab.h>
  31 #include <linux/spinlock.h>
  32 #include <linux/zpool.h>
  33
  34 /*****************
  35  * Structures
  36 *****************/
  37 struct z3fold_pool;
  38 struct z3fold_ops {
  39         int (*evict)(struct z3fold_pool *pool, unsigned long handle);
  40 };
  41
  42 enum buddy {
  43         HEADLESS = 0,
  44         FIRST,
  45         MIDDLE,
  46         LAST,
  47         BUDDIES_MAX
  48 };
  49
  50 /*
  51  * struct z3fold_header - z3fold page metadata occupying the first chunk of each
  52  *                      z3fold page, except for HEADLESS pages
  53  * @buddy:      links the z3fold page into the relevant list in the pool
  54  * @page_lock:          per-page lock
  55  * @refcount:           reference cound for the z3fold page
  56  * @first_chunks:       the size of the first buddy in chunks, 0 if free
  57  * @middle_chunks:      the size of the middle buddy in chunks, 0 if free
  58  * @last_chunks:        the size of the last buddy in chunks, 0 if free
  59  * @first_num:          the starting number (for the first handle)
  60  */
  61 struct z3fold_header {
  62         struct list_head buddy;
  63         spinlock_t page_lock;
  64         struct kref refcount;
  65         unsigned short first_chunks;
  66         unsigned short middle_chunks;
  67         unsigned short last_chunks;
  68         unsigned short start_middle;
  69         unsigned short first_num:2;
  70 };
  71
  72 /*
  73  * NCHUNKS_ORDER determines the internal allocation granularity, effectively
  74  * adjusting internal fragmentation.  It also determines the number of
  75  * freelists maintained in each pool. NCHUNKS_ORDER of 6 means that the
  76  * allocation granularity will be in chunks of size PAGE_SIZE/64. Some chunks
  77  * in the beginning of an allocated page are occupied by z3fold header, so
  78  * NCHUNKS will be calculated to 63 (or 62 in case CONFIG_DEBUG_SPINLOCK=y),
  79  * which shows the max number of free chunks in z3fold page, also there will
  80  * be 63, or 62, respectively, freelists per pool.
  81  */
  82 #define NCHUNKS_ORDER   6
  83
  84 #define CHUNK_SHIFT     (PAGE_SHIFT - NCHUNKS_ORDER)
  85 #define CHUNK_SIZE      (1 << CHUNK_SHIFT)
  86 #define ZHDR_SIZE_ALIGNED round_up(sizeof(struct z3fold_header), CHUNK_SIZE)
  87 #define ZHDR_CHUNKS     (ZHDR_SIZE_ALIGNED >> CHUNK_SHIFT)
  88 #define TOTAL_CHUNKS    (PAGE_SIZE >> CHUNK_SHIFT)
  89 #define NCHUNKS         ((PAGE_SIZE - ZHDR_SIZE_ALIGNED) >> CHUNK_SHIFT)
  90
  91 #define BUDDY_MASK      (0x3)
  92
  93 /**
  94  * struct z3fold_pool - stores metadata for each z3fold pool
  95  * @lock:       protects all pool fields and first|last_chunk fields of any
  96  *              z3fold page in the pool
  97  * @unbuddied:  array of lists tracking z3fold pages that contain 2- buddies;
  98  *              the lists each z3fold page is added to depends on the size of
  99  *              its free region.
 100  * @lru:        list tracking the z3fold pages in LRU order by most recently
 101  *              added buddy.
 102  * @pages_nr:   number of z3fold pages in the pool.
 103  * @ops:        pointer to a structure of user defined operations specified at
 104  *              pool creation time.
 105  *
 106  * This structure is allocated at pool creation time and maintains metadata
 107  * pertaining to a particular z3fold pool.
 108  */
 109 struct z3fold_pool {
 110         spinlock_t lock;
 111         struct list_head unbuddied[NCHUNKS];
 112         struct list_head lru;
 113         atomic64_t pages_nr;
 114         const struct z3fold_ops *ops;
 115         struct zpool *zpool;
 116         const struct zpool_ops *zpool_ops;
 117 };
 118
 119 /*
 120  * Internal z3fold page flags
 121  */
 122 enum z3fold_page_flags {
 123         PAGE_HEADLESS = 0,
 124         MIDDLE_CHUNK_MAPPED,
 125 };
 126
 127
 128 /*****************
 129  * Helpers
 130 *****************/
 131
 132 /* Converts an allocation size in bytes to size in z3fold chunks */
 133 static int size_to_chunks(size_t size)
 134 {
 135         return (size + CHUNK_SIZE - 1) >> CHUNK_SHIFT;
 136 }
 137
 138 #define for_each_unbuddied_list(_iter, _begin) \
 139         for ((_iter) = (_begin); (_iter) < NCHUNKS; (_iter)++)
 140
 141 /* Initializes the z3fold header of a newly allocated z3fold page */
 142 static struct z3fold_header *init_z3fold_page(struct page *page)
 143 {
 144         struct z3fold_header *zhdr = page_address(page);
 145
 146         INIT_LIST_HEAD(&page->lru);
 147         clear_bit(PAGE_HEADLESS, &page->private);
 148         clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
 149
 150         spin_lock_init(&zhdr->page_lock);
 151         kref_init(&zhdr->refcount);
 152         zhdr->first_chunks = 0;
 153         zhdr->middle_chunks = 0;
 154         zhdr->last_chunks = 0;
 155         zhdr->first_num = 0;
 156         zhdr->start_middle = 0;
 157         INIT_LIST_HEAD(&zhdr->buddy);
 158         return zhdr;
 159 }
 160
 161 /* Resets the struct page fields and frees the page */
 162 static void free_z3fold_page(struct page *page)
 163 {
 164         __free_page(page);
 165 }
 166
 167 static void release_z3fold_page(struct kref *ref)
 168 {
 169         struct z3fold_header *zhdr;
 170         struct page *page;
 171
 172         zhdr = container_of(ref, struct z3fold_header, refcount);
 173         page = virt_to_page(zhdr);
 174
 175         if (!list_empty(&zhdr->buddy))
 176                 list_del(&zhdr->buddy);
 177         if (!list_empty(&page->lru))
 178                 list_del(&page->lru);
 179         free_z3fold_page(page);
 180 }
 181
 182 /* Lock a z3fold page */
 183 static inline void z3fold_page_lock(struct z3fold_header *zhdr)
 184 {
 185         spin_lock(&zhdr->page_lock);
 186 }
 187
 188 /* Try to lock a z3fold page */
 189 static inline int z3fold_page_trylock(struct z3fold_header *zhdr)
 190 {
 191         return spin_trylock(&zhdr->page_lock);
 192 }
 193
 194 /* Unlock a z3fold page */
 195 static inline void z3fold_page_unlock(struct z3fold_header *zhdr)
 196 {
 197         spin_unlock(&zhdr->page_lock);
 198 }
 199
 200 /*
 201  * Encodes the handle of a particular buddy within a z3fold page
 202  * Pool lock should be held as this function accesses first_num
 203  */
 204 static unsigned long encode_handle(struct z3fold_header *zhdr, enum buddy bud)
 205 {
 206         unsigned long handle;
 207
 208         handle = (unsigned long)zhdr;
 209         if (bud != HEADLESS)
 210                 handle += (bud + zhdr->first_num) & BUDDY_MASK;
 211         return handle;
 212 }
 213
 214 /* Returns the z3fold page where a given handle is stored */
 215 static struct z3fold_header *handle_to_z3fold_header(unsigned long handle)
 216 {
 217         return (struct z3fold_header *)(handle & PAGE_MASK);
 218 }
 219
 220 /*
 221  * (handle & BUDDY_MASK) < zhdr->first_num is possible in encode_handle
 222  *  but that doesn't matter. because the masking will result in the
 223  *  correct buddy number.
 224  */
 225 static enum buddy handle_to_buddy(unsigned long handle)
 226 {
 227         struct z3fold_header *zhdr = handle_to_z3fold_header(handle);
 228         return (handle - zhdr->first_num) & BUDDY_MASK;
 229 }
 230
 231 /*
 232  * Returns the number of free chunks in a z3fold page.
 233  * NB: can't be used with HEADLESS pages.
 234  */
 235 static int num_free_chunks(struct z3fold_header *zhdr)
 236 {
 237         int nfree;
 238         /*
 239          * If there is a middle object, pick up the bigger free space
 240          * either before or after it. Otherwise just subtract the number
 241          * of chunks occupied by the first and the last objects.
 242          */
 243         if (zhdr->middle_chunks != 0) {
 244                 int nfree_before = zhdr->first_chunks ?
 245                         0 : zhdr->start_middle - ZHDR_CHUNKS;
 246                 int nfree_after = zhdr->last_chunks ?
 247                         0 : TOTAL_CHUNKS -
 248                                 (zhdr->start_middle + zhdr->middle_chunks);
 249                 nfree = max(nfree_before, nfree_after);
 250         } else
 251                 nfree = NCHUNKS - zhdr->first_chunks - zhdr->last_chunks;
 252         return nfree;
 253 }
 254
 255 /*****************
 256  * API Functions
 257 *****************/
 258 /**
 259  * z3fold_create_pool() - create a new z3fold pool
 260  * @gfp:        gfp flags when allocating the z3fold pool structure
 261  * @ops:        user-defined operations for the z3fold pool
 262  *
 263  * Return: pointer to the new z3fold pool or NULL if the metadata allocation
 264  * failed.
 265  */
 266 static struct z3fold_pool *z3fold_create_pool(gfp_t gfp,
 267                 const struct z3fold_ops *ops)
 268 {
 269         struct z3fold_pool *pool;
 270         int i;
 271
 272         pool = kzalloc(sizeof(struct z3fold_pool), gfp);
 273         if (!pool)
 274                 return NULL;
 275         spin_lock_init(&pool->lock);
 276         for_each_unbuddied_list(i, 0)
 277                 INIT_LIST_HEAD(&pool->unbuddied[i]);
 278         INIT_LIST_HEAD(&pool->lru);
 279         atomic64_set(&pool->pages_nr, 0);
 280         pool->ops = ops;
 281         return pool;
 282 }
 283
 284 /**
 285  * z3fold_destroy_pool() - destroys an existing z3fold pool
 286  * @pool:       the z3fold pool to be destroyed
 287  *
 288  * The pool should be emptied before this function is called.
 289  */
 290 static void z3fold_destroy_pool(struct z3fold_pool *pool)
 291 {
 292         kfree(pool);
 293 }
 294
 295 static inline void *mchunk_memmove(struct z3fold_header *zhdr,
 296                                 unsigned short dst_chunk)
 297 {
 298         void *beg = zhdr;
 299         return memmove(beg + (dst_chunk << CHUNK_SHIFT),
 300                        beg + (zhdr->start_middle << CHUNK_SHIFT),
 301                        zhdr->middle_chunks << CHUNK_SHIFT);
 302 }
 303
 304 #define BIG_CHUNK_GAP   3
 305 /* Has to be called with lock held */
 306 static int z3fold_compact_page(struct z3fold_header *zhdr)
 307 {
 308         struct page *page = virt_to_page(zhdr);
 309
 310         if (test_bit(MIDDLE_CHUNK_MAPPED, &page->private))
 311                 return 0; /* can't move middle chunk, it's used */
 312
 313         if (zhdr->middle_chunks == 0)
 314                 return 0; /* nothing to compact */
 315
 316         if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
 317                 /* move to the beginning */
 318                 mchunk_memmove(zhdr, ZHDR_CHUNKS);
 319                 zhdr->first_chunks = zhdr->middle_chunks;
 320                 zhdr->middle_chunks = 0;
 321                 zhdr->start_middle = 0;
 322                 zhdr->first_num++;
 323                 return 1;
 324         }
 325
 326         /*
 327          * moving data is expensive, so let's only do that if
 328          * there's substantial gain (at least BIG_CHUNK_GAP chunks)
 329          */
 330         if (zhdr->first_chunks != 0 && zhdr->last_chunks == 0 &&
 331             zhdr->start_middle - (zhdr->first_chunks + ZHDR_CHUNKS) >=
 332                         BIG_CHUNK_GAP) {
 333                 mchunk_memmove(zhdr, zhdr->first_chunks + ZHDR_CHUNKS);
 334                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
 335                 return 1;
 336         } else if (zhdr->last_chunks != 0 && zhdr->first_chunks == 0 &&
 337                    TOTAL_CHUNKS - (zhdr->last_chunks + zhdr->start_middle
 338                                         + zhdr->middle_chunks) >=
 339                         BIG_CHUNK_GAP) {
 340                 unsigned short new_start = TOTAL_CHUNKS - zhdr->last_chunks -
 341                         zhdr->middle_chunks;
 342                 mchunk_memmove(zhdr, new_start);
 343                 zhdr->start_middle = new_start;
 344                 return 1;
 345         }
 346
 347         return 0;
 348 }
 349
 350 /**
 351  * z3fold_alloc() - allocates a region of a given size
 352  * @pool:       z3fold pool from which to allocate
 353  * @size:       size in bytes of the desired allocation
 354  * @gfp:        gfp flags used if the pool needs to grow
 355  * @handle:     handle of the new allocation
 356  *
 357  * This function will attempt to find a free region in the pool large enough to
 358  * satisfy the allocation request.  A search of the unbuddied lists is
 359  * performed first. If no suitable free region is found, then a new page is
 360  * allocated and added to the pool to satisfy the request.
 361  *
 362  * gfp should not set __GFP_HIGHMEM as highmem pages cannot be used
 363  * as z3fold pool pages.
 364  *
 365  * Return: 0 if success and handle is set, otherwise -EINVAL if the size or
 366  * gfp arguments are invalid or -ENOMEM if the pool was unable to allocate
 367  * a new page.
 368  */
 369 static int z3fold_alloc(struct z3fold_pool *pool, size_t size, gfp_t gfp,
 370                         unsigned long *handle)
 371 {
 372         int chunks = 0, i, freechunks;
 373         struct z3fold_header *zhdr = NULL;
 374         enum buddy bud;
 375         struct page *page;
 376
 377         if (!size || (gfp & __GFP_HIGHMEM))
 378                 return -EINVAL;
 379
 380         if (size > PAGE_SIZE)
 381                 return -ENOSPC;
 382
 383         if (size > PAGE_SIZE - ZHDR_SIZE_ALIGNED - CHUNK_SIZE)
 384                 bud = HEADLESS;
 385         else {
 386                 chunks = size_to_chunks(size);
 387
 388                 /* First, try to find an unbuddied z3fold page. */
 389                 zhdr = NULL;
 390                 for_each_unbuddied_list(i, chunks) {
 391                         spin_lock(&pool->lock);
 392                         zhdr = list_first_entry_or_null(&pool->unbuddied[i],
 393                                                 struct z3fold_header, buddy);
 394                         if (!zhdr || !z3fold_page_trylock(zhdr)) {
 395                                 spin_unlock(&pool->lock);
 396                                 continue;
 397                         }
 398                         kref_get(&zhdr->refcount);
 399                         list_del_init(&zhdr->buddy);
 400                         spin_unlock(&pool->lock);
 401
 402                         page = virt_to_page(zhdr);
 403                         if (zhdr->first_chunks == 0) {
 404                                 if (zhdr->middle_chunks != 0 &&
 405                                     chunks >= zhdr->start_middle)
 406                                         bud = LAST;
 407                                 else
 408                                         bud = FIRST;
 409                         } else if (zhdr->last_chunks == 0)
 410                                 bud = LAST;
 411                         else if (zhdr->middle_chunks == 0)
 412                                 bud = MIDDLE;
 413                         else {
 414                                 z3fold_page_unlock(zhdr);
 415                                 spin_lock(&pool->lock);
 416                                 if (kref_put(&zhdr->refcount,
 417                                              release_z3fold_page))
 418                                         atomic64_dec(&pool->pages_nr);
 419                                 spin_unlock(&pool->lock);
 420                                 pr_err("No free chunks in unbuddied\n");
 421                                 WARN_ON(1);
 422                                 continue;
 423                         }
 424                         goto found;
 425                 }
 426                 bud = FIRST;
 427         }
 428
 429         /* Couldn't find unbuddied z3fold page, create new one */
 430         page = alloc_page(gfp);
 431         if (!page)
 432                 return -ENOMEM;
 433
 434         atomic64_inc(&pool->pages_nr);
 435         zhdr = init_z3fold_page(page);
 436
 437         if (bud == HEADLESS) {
 438                 set_bit(PAGE_HEADLESS, &page->private);
 439                 spin_lock(&pool->lock);
 440                 goto headless;
 441         }
 442         z3fold_page_lock(zhdr);
 443
 444 found:
 445         if (bud == FIRST)
 446                 zhdr->first_chunks = chunks;
 447         else if (bud == LAST)
 448                 zhdr->last_chunks = chunks;
 449         else {
 450                 zhdr->middle_chunks = chunks;
 451                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
 452         }
 453
 454         spin_lock(&pool->lock);
 455         if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0 ||
 456                         zhdr->middle_chunks == 0) {
 457                 /* Add to unbuddied list */
 458                 freechunks = num_free_chunks(zhdr);
 459                 list_add(&zhdr->buddy, &pool->unbuddied[freechunks]);
 460         }
 461
 462 headless:
 463         /* Add/move z3fold page to beginning of LRU */
 464         if (!list_empty(&page->lru))
 465                 list_del(&page->lru);
 466
 467         list_add(&page->lru, &pool->lru);
 468
 469         *handle = encode_handle(zhdr, bud);
 470         spin_unlock(&pool->lock);
 471         if (bud != HEADLESS)
 472                 z3fold_page_unlock(zhdr);
 473
 474         return 0;
 475 }
 476
 477 /**
 478  * z3fold_free() - frees the allocation associated with the given handle
 479  * @pool:       pool in which the allocation resided
 480  * @handle:     handle associated with the allocation returned by z3fold_alloc()
 481  *
 482  * In the case that the z3fold page in which the allocation resides is under
 483  * reclaim, as indicated by the PG_reclaim flag being set, this function
 484  * only sets the first|last_chunks to 0.  The page is actually freed
 485  * once both buddies are evicted (see z3fold_reclaim_page() below).
 486  */
 487 static void z3fold_free(struct z3fold_pool *pool, unsigned long handle)
 488 {
 489         struct z3fold_header *zhdr;
 490         int freechunks;
 491         struct page *page;
 492         enum buddy bud;
 493
 494         zhdr = handle_to_z3fold_header(handle);
 495         page = virt_to_page(zhdr);
 496
 497         if (test_bit(PAGE_HEADLESS, &page->private)) {
 498                 /* HEADLESS page stored */
 499                 bud = HEADLESS;
 500         } else {
 501                 z3fold_page_lock(zhdr);
 502                 bud = handle_to_buddy(handle);
 503
 504                 switch (bud) {
 505                 case FIRST:
 506                         zhdr->first_chunks = 0;
 507                         break;
 508                 case MIDDLE:
 509                         zhdr->middle_chunks = 0;
 510                         zhdr->start_middle = 0;
 511                         break;
 512                 case LAST:
 513                         zhdr->last_chunks = 0;
 514                         break;
 515                 default:
 516                         pr_err("%s: unknown bud %d\n", __func__, bud);
 517                         WARN_ON(1);
 518                         z3fold_page_unlock(zhdr);
 519                         return;
 520                 }
 521         }
 522
 523         if (bud == HEADLESS) {
 524                 spin_lock(&pool->lock);
 525                 list_del(&page->lru);
 526                 spin_unlock(&pool->lock);
 527                 free_z3fold_page(page);
 528                 atomic64_dec(&pool->pages_nr);
 529         } else {
 530                 if (zhdr->first_chunks != 0 || zhdr->middle_chunks != 0 ||
 531                     zhdr->last_chunks != 0) {
 532                         z3fold_compact_page(zhdr);
 533                         /* Add to the unbuddied list */
 534                         spin_lock(&pool->lock);
 535                         if (!list_empty(&zhdr->buddy))
 536                                 list_del(&zhdr->buddy);
 537                         freechunks = num_free_chunks(zhdr);
 538                         list_add(&zhdr->buddy, &pool->unbuddied[freechunks]);
 539                         spin_unlock(&pool->lock);
 540                 }
 541                 z3fold_page_unlock(zhdr);
 542                 spin_lock(&pool->lock);
 543                 if (kref_put(&zhdr->refcount, release_z3fold_page))
 544                         atomic64_dec(&pool->pages_nr);
 545                 spin_unlock(&pool->lock);
 546         }
 547
 548 }
 549
 550 /**
 551  * z3fold_reclaim_page() - evicts allocations from a pool page and frees it
 552  * @pool:       pool from which a page will attempt to be evicted
 553  * @retires:    number of pages on the LRU list for which eviction will
 554  *              be attempted before failing
 555  *
 556  * z3fold reclaim is different from normal system reclaim in that it is done
 557  * from the bottom, up. This is because only the bottom layer, z3fold, has
 558  * information on how the allocations are organized within each z3fold page.
 559  * This has the potential to create interesting locking situations between
 560  * z3fold and the user, however.
 561  *
 562  * To avoid these, this is how z3fold_reclaim_page() should be called:
 563
 564  * The user detects a page should be reclaimed and calls z3fold_reclaim_page().
 565  * z3fold_reclaim_page() will remove a z3fold page from the pool LRU list and
 566  * call the user-defined eviction handler with the pool and handle as
 567  * arguments.
 568  *
 569  * If the handle can not be evicted, the eviction handler should return
 570  * non-zero. z3fold_reclaim_page() will add the z3fold page back to the
 571  * appropriate list and try the next z3fold page on the LRU up to
 572  * a user defined number of retries.
 573  *
 574  * If the handle is successfully evicted, the eviction handler should
 575  * return 0 _and_ should have called z3fold_free() on the handle. z3fold_free()
 576  * contains logic to delay freeing the page if the page is under reclaim,
 577  * as indicated by the setting of the PG_reclaim flag on the underlying page.
 578  *
 579  * If all buddies in the z3fold page are successfully evicted, then the
 580  * z3fold page can be freed.
 581  *
 582  * Returns: 0 if page is successfully freed, otherwise -EINVAL if there are
 583  * no pages to evict or an eviction handler is not registered, -EAGAIN if
 584  * the retry limit was hit.
 585  */
 586 static int z3fold_reclaim_page(struct z3fold_pool *pool, unsigned int retries)
 587 {
 588         int i, ret = 0, freechunks;
 589         struct z3fold_header *zhdr;
 590         struct page *page;
 591         unsigned long first_handle = 0, middle_handle = 0, last_handle = 0;
 592
 593         spin_lock(&pool->lock);
 594         if (!pool->ops || !pool->ops->evict || retries == 0) {
 595                 spin_unlock(&pool->lock);
 596                 return -EINVAL;
 597         }
 598         for (i = 0; i < retries; i++) {
 599                 if (list_empty(&pool->lru)) {
 600                         spin_unlock(&pool->lock);
 601                         return -EINVAL;
 602                 }
 603                 page = list_last_entry(&pool->lru, struct page, lru);
 604                 list_del_init(&page->lru);
 605
 606                 zhdr = page_address(page);
 607                 if (!test_bit(PAGE_HEADLESS, &page->private)) {
 608                         if (!list_empty(&zhdr->buddy))
 609                                 list_del_init(&zhdr->buddy);
 610                         kref_get(&zhdr->refcount);
 611                         spin_unlock(&pool->lock);
 612                         z3fold_page_lock(zhdr);
 613                         /*
 614                          * We need encode the handles before unlocking, since
 615                          * we can race with free that will set
 616                          * (first|last)_chunks to 0
 617                          */
 618                         first_handle = 0;
 619                         last_handle = 0;
 620                         middle_handle = 0;
 621                         if (zhdr->first_chunks)
 622                                 first_handle = encode_handle(zhdr, FIRST);
 623                         if (zhdr->middle_chunks)
 624                                 middle_handle = encode_handle(zhdr, MIDDLE);
 625                         if (zhdr->last_chunks)
 626                                 last_handle = encode_handle(zhdr, LAST);
 627                         z3fold_page_unlock(zhdr);
 628                 } else {
 629                         first_handle = encode_handle(zhdr, HEADLESS);
 630                         last_handle = middle_handle = 0;
 631                         spin_unlock(&pool->lock);
 632                 }
 633
 634                 /* Issue the eviction callback(s) */
 635                 if (middle_handle) {
 636                         ret = pool->ops->evict(pool, middle_handle);
 637                         if (ret)
 638                                 goto next;
 639                 }
 640                 if (first_handle) {
 641                         ret = pool->ops->evict(pool, first_handle);
 642                         if (ret)
 643                                 goto next;
 644                 }
 645                 if (last_handle) {
 646                         ret = pool->ops->evict(pool, last_handle);
 647                         if (ret)
 648                                 goto next;
 649                 }
 650 next:
 651                 if (test_bit(PAGE_HEADLESS, &page->private)) {
 652                         if (ret == 0) {
 653                                 free_z3fold_page(page);
 654                                 return 0;
 655                         } else {
 656                                 spin_lock(&pool->lock);
 657                         }
 658                 } else {
 659                         z3fold_page_lock(zhdr);
 660                         if ((zhdr->first_chunks || zhdr->last_chunks ||
 661                              zhdr->middle_chunks) &&
 662                             !(zhdr->first_chunks && zhdr->last_chunks &&
 663                               zhdr->middle_chunks)) {
 664                                 z3fold_compact_page(zhdr);
 665                                 /* add to unbuddied list */
 666                                 spin_lock(&pool->lock);
 667                                 freechunks = num_free_chunks(zhdr);
 668                                 list_add(&zhdr->buddy,
 669                                          &pool->unbuddied[freechunks]);
 670                                 spin_unlock(&pool->lock);
 671                         }
 672                         z3fold_page_unlock(zhdr);
 673                         spin_lock(&pool->lock);
 674                         if (kref_put(&zhdr->refcount, release_z3fold_page)) {
 675                                 spin_unlock(&pool->lock);
 676                                 atomic64_dec(&pool->pages_nr);
 677                                 return 0;
 678                         }
 679                 }
 680
 681                 /*
 682                  * Add to the beginning of LRU.
 683                  * Pool lock has to be kept here to ensure the page has
 684                  * not already been released
 685                  */
 686                 list_add(&page->lru, &pool->lru);
 687         }
 688         spin_unlock(&pool->lock);
 689         return -EAGAIN;
 690 }
 691
 692 /**
 693  * z3fold_map() - maps the allocation associated with the given handle
 694  * @pool:       pool in which the allocation resides
 695  * @handle:     handle associated with the allocation to be mapped
 696  *
 697  * Extracts the buddy number from handle and constructs the pointer to the
 698  * correct starting chunk within the page.
 699  *
 700  * Returns: a pointer to the mapped allocation
 701  */
 702 static void *z3fold_map(struct z3fold_pool *pool, unsigned long handle)
 703 {
 704         struct z3fold_header *zhdr;
 705         struct page *page;
 706         void *addr;
 707         enum buddy buddy;
 708
 709         zhdr = handle_to_z3fold_header(handle);
 710         addr = zhdr;
 711         page = virt_to_page(zhdr);
 712
 713         if (test_bit(PAGE_HEADLESS, &page->private))
 714                 goto out;
 715
 716         z3fold_page_lock(zhdr);
 717         buddy = handle_to_buddy(handle);
 718         switch (buddy) {
 719         case FIRST:
 720                 addr += ZHDR_SIZE_ALIGNED;
 721                 break;
 722         case MIDDLE:
 723                 addr += zhdr->start_middle << CHUNK_SHIFT;
 724                 set_bit(MIDDLE_CHUNK_MAPPED, &page->private);
 725                 break;
 726         case LAST:
 727                 addr += PAGE_SIZE - (zhdr->last_chunks << CHUNK_SHIFT);
 728                 break;
 729         default:
 730                 pr_err("unknown buddy id %d\n", buddy);
 731                 WARN_ON(1);
 732                 addr = NULL;
 733                 break;
 734         }
 735
 736         z3fold_page_unlock(zhdr);
 737 out:
 738         return addr;
 739 }
 740
 741 /**
 742  * z3fold_unmap() - unmaps the allocation associated with the given handle
 743  * @pool:       pool in which the allocation resides
 744  * @handle:     handle associated with the allocation to be unmapped
 745  */
 746 static void z3fold_unmap(struct z3fold_pool *pool, unsigned long handle)
 747 {
 748         struct z3fold_header *zhdr;
 749         struct page *page;
 750         enum buddy buddy;
 751
 752         zhdr = handle_to_z3fold_header(handle);
 753         page = virt_to_page(zhdr);
 754
 755         if (test_bit(PAGE_HEADLESS, &page->private))
 756                 return;
 757
 758         z3fold_page_lock(zhdr);
 759         buddy = handle_to_buddy(handle);
 760         if (buddy == MIDDLE)
 761                 clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
 762         z3fold_page_unlock(zhdr);
 763 }
 764
 765 /**
 766  * z3fold_get_pool_size() - gets the z3fold pool size in pages
 767  * @pool:       pool whose size is being queried
 768  *
 769  * Returns: size in pages of the given pool.
 770  */
 771 static u64 z3fold_get_pool_size(struct z3fold_pool *pool)
 772 {
 773         return atomic64_read(&pool->pages_nr);
 774 }
 775
 776 /*****************
 777  * zpool
 778  ****************/
 779
 780 static int z3fold_zpool_evict(struct z3fold_pool *pool, unsigned long handle)
 781 {
 782         if (pool->zpool && pool->zpool_ops && pool->zpool_ops->evict)
 783                 return pool->zpool_ops->evict(pool->zpool, handle);
 784         else
 785                 return -ENOENT;
 786 }
 787
 788 static const struct z3fold_ops z3fold_zpool_ops = {
 789         .evict =        z3fold_zpool_evict
 790 };
 791
 792 static void *z3fold_zpool_create(const char *name, gfp_t gfp,
 793                                const struct zpool_ops *zpool_ops,
 794                                struct zpool *zpool)
 795 {
 796         struct z3fold_pool *pool;
 797
 798         pool = z3fold_create_pool(gfp, zpool_ops ? &z3fold_zpool_ops : NULL);
 799         if (pool) {
 800                 pool->zpool = zpool;
 801                 pool->zpool_ops = zpool_ops;
 802         }
 803         return pool;
 804 }
 805
 806 static void z3fold_zpool_destroy(void *pool)
 807 {
 808         z3fold_destroy_pool(pool);
 809 }
 810
 811 static int z3fold_zpool_malloc(void *pool, size_t size, gfp_t gfp,
 812                         unsigned long *handle)
 813 {
 814         return z3fold_alloc(pool, size, gfp, handle);
 815 }
 816 static void z3fold_zpool_free(void *pool, unsigned long handle)
 817 {
 818         z3fold_free(pool, handle);
 819 }
 820
 821 static int z3fold_zpool_shrink(void *pool, unsigned int pages,
 822                         unsigned int *reclaimed)
 823 {
 824         unsigned int total = 0;
 825         int ret = -EINVAL;
 826
 827         while (total < pages) {
 828                 ret = z3fold_reclaim_page(pool, 8);
 829                 if (ret < 0)
 830                         break;
 831                 total++;
 832         }
 833
 834         if (reclaimed)
 835                 *reclaimed = total;
 836
 837         return ret;
 838 }
 839
 840 static void *z3fold_zpool_map(void *pool, unsigned long handle,
 841                         enum zpool_mapmode mm)
 842 {
 843         return z3fold_map(pool, handle);
 844 }
 845 static void z3fold_zpool_unmap(void *pool, unsigned long handle)
 846 {
 847         z3fold_unmap(pool, handle);
 848 }
 849
 850 static u64 z3fold_zpool_total_size(void *pool)
 851 {
 852         return z3fold_get_pool_size(pool) * PAGE_SIZE;
 853 }
 854
 855 static struct zpool_driver z3fold_zpool_driver = {
 856         .type =         "z3fold",
 857         .owner =        THIS_MODULE,
 858         .create =       z3fold_zpool_create,
 859         .destroy =      z3fold_zpool_destroy,
 860         .malloc =       z3fold_zpool_malloc,
 861         .free =         z3fold_zpool_free,
 862         .shrink =       z3fold_zpool_shrink,
 863         .map =          z3fold_zpool_map,
 864         .unmap =        z3fold_zpool_unmap,
 865         .total_size =   z3fold_zpool_total_size,
 866 };
 867
 868 MODULE_ALIAS("zpool-z3fold");
 869
 870 static int __init init_z3fold(void)
 871 {
 872         /* Make sure the z3fold header is not larger than the page size */
 873         BUILD_BUG_ON(ZHDR_SIZE_ALIGNED > PAGE_SIZE);
 874         zpool_register_driver(&z3fold_zpool_driver);
 875
 876         return 0;
 877 }
 878
 879 static void __exit exit_z3fold(void)
 880 {
 881         zpool_unregister_driver(&z3fold_zpool_driver);
 882 }
 883
 884 module_init(init_z3fold);
 885 module_exit(exit_z3fold);
 886
 887 MODULE_LICENSE("GPL");
 888 MODULE_AUTHOR("Vitaly Wool <vitalywool@gmail.com>");
 889 MODULE_DESCRIPTION("3-Fold Allocator for Compressed Pages");