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