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/list.h>
  28 #include <linux/mm.h>
  29 #include <linux/module.h>
  30 #include <linux/percpu.h>
  31 #include <linux/preempt.h>
  32 #include <linux/workqueue.h>
  33 #include <linux/slab.h>
  34 #include <linux/spinlock.h>
  35 #include <linux/zpool.h>
  36
  37 /*****************
  38  * Structures
  39 *****************/
  40 struct z3fold_pool;
  41 struct z3fold_ops {
  42         int (*evict)(struct z3fold_pool *pool, unsigned long handle);
  43 };
  44
  45 enum buddy {
  46         HEADLESS = 0,
  47         FIRST,
  48         MIDDLE,
  49         LAST,
  50         BUDDIES_MAX
  51 };
  52
  53 /*
  54  * struct z3fold_header - z3fold page metadata occupying first chunks of each
  55  *                      z3fold page, except for HEADLESS pages
  56  * @buddy:              links the z3fold page into the relevant list in the
  57  *                      pool
  58  * @page_lock:          per-page lock
  59  * @refcount:           reference count for the z3fold page
  60  * @work:               work_struct for page layout optimization
  61  * @pool:               pointer to the pool which this page belongs to
  62  * @cpu:                CPU which this page "belongs" to
  63  * @first_chunks:       the size of the first buddy in chunks, 0 if free
  64  * @middle_chunks:      the size of the middle buddy in chunks, 0 if free
  65  * @last_chunks:        the size of the last buddy in chunks, 0 if free
  66  * @first_num:          the starting number (for the first handle)
  67  */
  68 struct z3fold_header {
  69         struct list_head buddy;
  70         spinlock_t page_lock;
  71         struct kref refcount;
  72         struct work_struct work;
  73         struct z3fold_pool *pool;
  74         short cpu;
  75         unsigned short first_chunks;
  76         unsigned short middle_chunks;
  77         unsigned short last_chunks;
  78         unsigned short start_middle;
  79         unsigned short first_num:2;
  80 };
  81
  82 /*
  83  * NCHUNKS_ORDER determines the internal allocation granularity, effectively
  84  * adjusting internal fragmentation.  It also determines the number of
  85  * freelists maintained in each pool. NCHUNKS_ORDER of 6 means that the
  86  * allocation granularity will be in chunks of size PAGE_SIZE/64. Some chunks
  87  * in the beginning of an allocated page are occupied by z3fold header, so
  88  * NCHUNKS will be calculated to 63 (or 62 in case CONFIG_DEBUG_SPINLOCK=y),
  89  * which shows the max number of free chunks in z3fold page, also there will
  90  * be 63, or 62, respectively, freelists per pool.
  91  */
  92 #define NCHUNKS_ORDER   6
  93
  94 #define CHUNK_SHIFT     (PAGE_SHIFT - NCHUNKS_ORDER)
  95 #define CHUNK_SIZE      (1 << CHUNK_SHIFT)
  96 #define ZHDR_SIZE_ALIGNED round_up(sizeof(struct z3fold_header), CHUNK_SIZE)
  97 #define ZHDR_CHUNKS     (ZHDR_SIZE_ALIGNED >> CHUNK_SHIFT)
  98 #define TOTAL_CHUNKS    (PAGE_SIZE >> CHUNK_SHIFT)
  99 #define NCHUNKS         ((PAGE_SIZE - ZHDR_SIZE_ALIGNED) >> CHUNK_SHIFT)
 100
 101 #define BUDDY_MASK      (0x3)
 102
 103 /**
 104  * struct z3fold_pool - stores metadata for each z3fold pool
 105  * @name:       pool name
 106  * @lock:       protects pool unbuddied/lru lists
 107  * @stale_lock: protects pool stale page list
 108  * @unbuddied:  per-cpu array of lists tracking z3fold pages that contain 2-
 109  *              buddies; the list each z3fold page is added to depends on
 110  *              the size of its free region.
 111  * @lru:        list tracking the z3fold pages in LRU order by most recently
 112  *              added buddy.
 113  * @stale:      list of pages marked for freeing
 114  * @pages_nr:   number of z3fold pages in the pool.
 115  * @ops:        pointer to a structure of user defined operations specified at
 116  *              pool creation time.
 117  * @compact_wq: workqueue for page layout background optimization
 118  * @release_wq: workqueue for safe page release
 119  * @work:       work_struct for safe page release
 120  *
 121  * This structure is allocated at pool creation time and maintains metadata
 122  * pertaining to a particular z3fold pool.
 123  */
 124 struct z3fold_pool {
 125         const char *name;
 126         spinlock_t lock;
 127         spinlock_t stale_lock;
 128         struct list_head *unbuddied;
 129         struct list_head lru;
 130         struct list_head stale;
 131         atomic64_t pages_nr;
 132         const struct z3fold_ops *ops;
 133         struct zpool *zpool;
 134         const struct zpool_ops *zpool_ops;
 135         struct workqueue_struct *compact_wq;
 136         struct workqueue_struct *release_wq;
 137         struct work_struct work;
 138 };
 139
 140 /*
 141  * Internal z3fold page flags
 142  */
 143 enum z3fold_page_flags {
 144         PAGE_HEADLESS = 0,
 145         MIDDLE_CHUNK_MAPPED,
 146         NEEDS_COMPACTING,
 147         PAGE_STALE
 148 };
 149
 150 /*****************
 151  * Helpers
 152 *****************/
 153
 154 /* Converts an allocation size in bytes to size in z3fold chunks */
 155 static int size_to_chunks(size_t size)
 156 {
 157         return (size + CHUNK_SIZE - 1) >> CHUNK_SHIFT;
 158 }
 159
 160 #define for_each_unbuddied_list(_iter, _begin) \
 161         for ((_iter) = (_begin); (_iter) < NCHUNKS; (_iter)++)
 162
 163 static void compact_page_work(struct work_struct *w);
 164
 165 /* Initializes the z3fold header of a newly allocated z3fold page */
 166 static struct z3fold_header *init_z3fold_page(struct page *page,
 167                                         struct z3fold_pool *pool)
 168 {
 169         struct z3fold_header *zhdr = page_address(page);
 170
 171         INIT_LIST_HEAD(&page->lru);
 172         clear_bit(PAGE_HEADLESS, &page->private);
 173         clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
 174         clear_bit(NEEDS_COMPACTING, &page->private);
 175         clear_bit(PAGE_STALE, &page->private);
 176
 177         spin_lock_init(&zhdr->page_lock);
 178         kref_init(&zhdr->refcount);
 179         zhdr->first_chunks = 0;
 180         zhdr->middle_chunks = 0;
 181         zhdr->last_chunks = 0;
 182         zhdr->first_num = 0;
 183         zhdr->start_middle = 0;
 184         zhdr->cpu = -1;
 185         zhdr->pool = pool;
 186         INIT_LIST_HEAD(&zhdr->buddy);
 187         INIT_WORK(&zhdr->work, compact_page_work);
 188         return zhdr;
 189 }
 190
 191 /* Resets the struct page fields and frees the page */
 192 static void free_z3fold_page(struct page *page)
 193 {
 194         __free_page(page);
 195 }
 196
 197 /* Lock a z3fold page */
 198 static inline void z3fold_page_lock(struct z3fold_header *zhdr)
 199 {
 200         spin_lock(&zhdr->page_lock);
 201 }
 202
 203 /* Try to lock a z3fold page */
 204 static inline int z3fold_page_trylock(struct z3fold_header *zhdr)
 205 {
 206         return spin_trylock(&zhdr->page_lock);
 207 }
 208
 209 /* Unlock a z3fold page */
 210 static inline void z3fold_page_unlock(struct z3fold_header *zhdr)
 211 {
 212         spin_unlock(&zhdr->page_lock);
 213 }
 214
 215 /*
 216  * Encodes the handle of a particular buddy within a z3fold page
 217  * Pool lock should be held as this function accesses first_num
 218  */
 219 static unsigned long encode_handle(struct z3fold_header *zhdr, enum buddy bud)
 220 {
 221         unsigned long handle;
 222
 223         handle = (unsigned long)zhdr;
 224         if (bud != HEADLESS)
 225                 handle += (bud + zhdr->first_num) & BUDDY_MASK;
 226         return handle;
 227 }
 228
 229 /* Returns the z3fold page where a given handle is stored */
 230 static struct z3fold_header *handle_to_z3fold_header(unsigned long handle)
 231 {
 232         return (struct z3fold_header *)(handle & PAGE_MASK);
 233 }
 234
 235 /*
 236  * (handle & BUDDY_MASK) < zhdr->first_num is possible in encode_handle
 237  *  but that doesn't matter. because the masking will result in the
 238  *  correct buddy number.
 239  */
 240 static enum buddy handle_to_buddy(unsigned long handle)
 241 {
 242         struct z3fold_header *zhdr = handle_to_z3fold_header(handle);
 243         return (handle - zhdr->first_num) & BUDDY_MASK;
 244 }
 245
 246 static void __release_z3fold_page(struct z3fold_header *zhdr, bool locked)
 247 {
 248         struct page *page = virt_to_page(zhdr);
 249         struct z3fold_pool *pool = zhdr->pool;
 250
 251         WARN_ON(!list_empty(&zhdr->buddy));
 252         set_bit(PAGE_STALE, &page->private);
 253         clear_bit(NEEDS_COMPACTING, &page->private);
 254         spin_lock(&pool->lock);
 255         if (!list_empty(&page->lru))
 256                 list_del(&page->lru);
 257         spin_unlock(&pool->lock);
 258         if (locked)
 259                 z3fold_page_unlock(zhdr);
 260         spin_lock(&pool->stale_lock);
 261         list_add(&zhdr->buddy, &pool->stale);
 262         queue_work(pool->release_wq, &pool->work);
 263         spin_unlock(&pool->stale_lock);
 264 }
 265
 266 static void __attribute__((__unused__))
 267                         release_z3fold_page(struct kref *ref)
 268 {
 269         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 270                                                 refcount);
 271         __release_z3fold_page(zhdr, false);
 272 }
 273
 274 static void release_z3fold_page_locked(struct kref *ref)
 275 {
 276         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 277                                                 refcount);
 278         WARN_ON(z3fold_page_trylock(zhdr));
 279         __release_z3fold_page(zhdr, true);
 280 }
 281
 282 static void release_z3fold_page_locked_list(struct kref *ref)
 283 {
 284         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 285                                                refcount);
 286         spin_lock(&zhdr->pool->lock);
 287         list_del_init(&zhdr->buddy);
 288         spin_unlock(&zhdr->pool->lock);
 289
 290         WARN_ON(z3fold_page_trylock(zhdr));
 291         __release_z3fold_page(zhdr, true);
 292 }
 293
 294 static void free_pages_work(struct work_struct *w)
 295 {
 296         struct z3fold_pool *pool = container_of(w, struct z3fold_pool, work);
 297
 298         spin_lock(&pool->stale_lock);
 299         while (!list_empty(&pool->stale)) {
 300                 struct z3fold_header *zhdr = list_first_entry(&pool->stale,
 301                                                 struct z3fold_header, buddy);
 302                 struct page *page = virt_to_page(zhdr);
 303
 304                 list_del(&zhdr->buddy);
 305                 if (WARN_ON(!test_bit(PAGE_STALE, &page->private)))
 306                         continue;
 307                 spin_unlock(&pool->stale_lock);
 308                 cancel_work_sync(&zhdr->work);
 309                 free_z3fold_page(page);
 310                 cond_resched();
 311                 spin_lock(&pool->stale_lock);
 312         }
 313         spin_unlock(&pool->stale_lock);
 314 }
 315
 316 /*
 317  * Returns the number of free chunks in a z3fold page.
 318  * NB: can't be used with HEADLESS pages.
 319  */
 320 static int num_free_chunks(struct z3fold_header *zhdr)
 321 {
 322         int nfree;
 323         /*
 324          * If there is a middle object, pick up the bigger free space
 325          * either before or after it. Otherwise just subtract the number
 326          * of chunks occupied by the first and the last objects.
 327          */
 328         if (zhdr->middle_chunks != 0) {
 329                 int nfree_before = zhdr->first_chunks ?
 330                         0 : zhdr->start_middle - ZHDR_CHUNKS;
 331                 int nfree_after = zhdr->last_chunks ?
 332                         0 : TOTAL_CHUNKS -
 333                                 (zhdr->start_middle + zhdr->middle_chunks);
 334                 nfree = max(nfree_before, nfree_after);
 335         } else
 336                 nfree = NCHUNKS - zhdr->first_chunks - zhdr->last_chunks;
 337         return nfree;
 338 }
 339
 340 static inline void *mchunk_memmove(struct z3fold_header *zhdr,
 341                                 unsigned short dst_chunk)
 342 {
 343         void *beg = zhdr;
 344         return memmove(beg + (dst_chunk << CHUNK_SHIFT),
 345                        beg + (zhdr->start_middle << CHUNK_SHIFT),
 346                        zhdr->middle_chunks << CHUNK_SHIFT);
 347 }
 348
 349 #define BIG_CHUNK_GAP   3
 350 /* Has to be called with lock held */
 351 static int z3fold_compact_page(struct z3fold_header *zhdr)
 352 {
 353         struct page *page = virt_to_page(zhdr);
 354
 355         if (test_bit(MIDDLE_CHUNK_MAPPED, &page->private))
 356                 return 0; /* can't move middle chunk, it's used */
 357
 358         if (zhdr->middle_chunks == 0)
 359                 return 0; /* nothing to compact */
 360
 361         if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
 362                 /* move to the beginning */
 363                 mchunk_memmove(zhdr, ZHDR_CHUNKS);
 364                 zhdr->first_chunks = zhdr->middle_chunks;
 365                 zhdr->middle_chunks = 0;
 366                 zhdr->start_middle = 0;
 367                 zhdr->first_num++;
 368                 return 1;
 369         }
 370
 371         /*
 372          * moving data is expensive, so let's only do that if
 373          * there's substantial gain (at least BIG_CHUNK_GAP chunks)
 374          */
 375         if (zhdr->first_chunks != 0 && zhdr->last_chunks == 0 &&
 376             zhdr->start_middle - (zhdr->first_chunks + ZHDR_CHUNKS) >=
 377                         BIG_CHUNK_GAP) {
 378                 mchunk_memmove(zhdr, zhdr->first_chunks + ZHDR_CHUNKS);
 379                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
 380                 return 1;
 381         } else if (zhdr->last_chunks != 0 && zhdr->first_chunks == 0 &&
 382                    TOTAL_CHUNKS - (zhdr->last_chunks + zhdr->start_middle
 383                                         + zhdr->middle_chunks) >=
 384                         BIG_CHUNK_GAP) {
 385                 unsigned short new_start = TOTAL_CHUNKS - zhdr->last_chunks -
 386                         zhdr->middle_chunks;
 387                 mchunk_memmove(zhdr, new_start);
 388                 zhdr->start_middle = new_start;
 389                 return 1;
 390         }
 391
 392         return 0;
 393 }
 394
 395 static void do_compact_page(struct z3fold_header *zhdr, bool locked)
 396 {
 397         struct z3fold_pool *pool = zhdr->pool;
 398         struct page *page;
 399         struct list_head *unbuddied;
 400         int fchunks;
 401
 402         page = virt_to_page(zhdr);
 403         if (locked)
 404                 WARN_ON(z3fold_page_trylock(zhdr));
 405         else
 406                 z3fold_page_lock(zhdr);
 407         if (WARN_ON(!test_and_clear_bit(NEEDS_COMPACTING, &page->private))) {
 408                 z3fold_page_unlock(zhdr);
 409                 return;
 410         }
 411         spin_lock(&pool->lock);
 412         list_del_init(&zhdr->buddy);
 413         spin_unlock(&pool->lock);
 414
 415         if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
 416                 atomic64_dec(&pool->pages_nr);
 417                 return;
 418         }
 419
 420         z3fold_compact_page(zhdr);
 421         unbuddied = get_cpu_ptr(pool->unbuddied);
 422         fchunks = num_free_chunks(zhdr);
 423         if (fchunks < NCHUNKS &&
 424             (!zhdr->first_chunks || !zhdr->middle_chunks ||
 425                         !zhdr->last_chunks)) {
 426                 /* the page's not completely free and it's unbuddied */
 427                 spin_lock(&pool->lock);
 428                 list_add(&zhdr->buddy, &unbuddied[fchunks]);
 429                 spin_unlock(&pool->lock);
 430                 zhdr->cpu = smp_processor_id();
 431         }
 432         put_cpu_ptr(pool->unbuddied);
 433         z3fold_page_unlock(zhdr);
 434 }
 435
 436 static void compact_page_work(struct work_struct *w)
 437 {
 438         struct z3fold_header *zhdr = container_of(w, struct z3fold_header,
 439                                                 work);
 440
 441         do_compact_page(zhdr, false);
 442 }
 443
 444
 445 /*
 446  * API Functions
 447  */
 448
 449 /**
 450  * z3fold_create_pool() - create a new z3fold pool
 451  * @name:       pool name
 452  * @gfp:        gfp flags when allocating the z3fold pool structure
 453  * @ops:        user-defined operations for the z3fold pool
 454  *
 455  * Return: pointer to the new z3fold pool or NULL if the metadata allocation
 456  * failed.
 457  */
 458 static struct z3fold_pool *z3fold_create_pool(const char *name, gfp_t gfp,
 459                 const struct z3fold_ops *ops)
 460 {
 461         struct z3fold_pool *pool = NULL;
 462         int i, cpu;
 463
 464         pool = kzalloc(sizeof(struct z3fold_pool), gfp);
 465         if (!pool)
 466                 goto out;
 467         spin_lock_init(&pool->lock);
 468         spin_lock_init(&pool->stale_lock);
 469         pool->unbuddied = __alloc_percpu(sizeof(struct list_head)*NCHUNKS, 2);
 470         if (!pool->unbuddied)
 471                 goto out_pool;
 472         for_each_possible_cpu(cpu) {
 473                 struct list_head *unbuddied =
 474                                 per_cpu_ptr(pool->unbuddied, cpu);
 475                 for_each_unbuddied_list(i, 0)
 476                         INIT_LIST_HEAD(&unbuddied[i]);
 477         }
 478         INIT_LIST_HEAD(&pool->lru);
 479         INIT_LIST_HEAD(&pool->stale);
 480         atomic64_set(&pool->pages_nr, 0);
 481         pool->name = name;
 482         pool->compact_wq = create_singlethread_workqueue(pool->name);
 483         if (!pool->compact_wq)
 484                 goto out_unbuddied;
 485         pool->release_wq = create_singlethread_workqueue(pool->name);
 486         if (!pool->release_wq)
 487                 goto out_wq;
 488         INIT_WORK(&pool->work, free_pages_work);
 489         pool->ops = ops;
 490         return pool;
 491
 492 out_wq:
 493         destroy_workqueue(pool->compact_wq);
 494 out_unbuddied:
 495         free_percpu(pool->unbuddied);
 496 out_pool:
 497         kfree(pool);
 498 out:
 499         return NULL;
 500 }
 501
 502 /**
 503  * z3fold_destroy_pool() - destroys an existing z3fold pool
 504  * @pool:       the z3fold pool to be destroyed
 505  *
 506  * The pool should be emptied before this function is called.
 507  */
 508 static void z3fold_destroy_pool(struct z3fold_pool *pool)
 509 {
 510         destroy_workqueue(pool->release_wq);
 511         destroy_workqueue(pool->compact_wq);
 512         kfree(pool);
 513 }
 514
 515 /**
 516  * z3fold_alloc() - allocates a region of a given size
 517  * @pool:       z3fold pool from which to allocate
 518  * @size:       size in bytes of the desired allocation
 519  * @gfp:        gfp flags used if the pool needs to grow
 520  * @handle:     handle of the new allocation
 521  *
 522  * This function will attempt to find a free region in the pool large enough to
 523  * satisfy the allocation request.  A search of the unbuddied lists is
 524  * performed first. If no suitable free region is found, then a new page is
 525  * allocated and added to the pool to satisfy the request.
 526  *
 527  * gfp should not set __GFP_HIGHMEM as highmem pages cannot be used
 528  * as z3fold pool pages.
 529  *
 530  * Return: 0 if success and handle is set, otherwise -EINVAL if the size or
 531  * gfp arguments are invalid or -ENOMEM if the pool was unable to allocate
 532  * a new page.
 533  */
 534 static int z3fold_alloc(struct z3fold_pool *pool, size_t size, gfp_t gfp,
 535                         unsigned long *handle)
 536 {
 537         int chunks = 0, i, freechunks;
 538         struct z3fold_header *zhdr = NULL;
 539         struct page *page = NULL;
 540         enum buddy bud;
 541         bool can_sleep = gfpflags_allow_blocking(gfp);
 542
 543         if (!size || (gfp & __GFP_HIGHMEM))
 544                 return -EINVAL;
 545
 546         if (size > PAGE_SIZE)
 547                 return -ENOSPC;
 548
 549         if (size > PAGE_SIZE - ZHDR_SIZE_ALIGNED - CHUNK_SIZE)
 550                 bud = HEADLESS;
 551         else {
 552                 struct list_head *unbuddied;
 553                 chunks = size_to_chunks(size);
 554
 555 lookup:
 556                 /* First, try to find an unbuddied z3fold page. */
 557                 unbuddied = get_cpu_ptr(pool->unbuddied);
 558                 for_each_unbuddied_list(i, chunks) {
 559                         struct list_head *l = &unbuddied[i];
 560
 561                         zhdr = list_first_entry_or_null(READ_ONCE(l),
 562                                                 struct z3fold_header, buddy);
 563
 564                         if (!zhdr)
 565                                 continue;
 566
 567                         /* Re-check under lock. */
 568                         spin_lock(&pool->lock);
 569                         l = &unbuddied[i];
 570                         if (unlikely(zhdr != list_first_entry(READ_ONCE(l),
 571                                         struct z3fold_header, buddy)) ||
 572                             !z3fold_page_trylock(zhdr)) {
 573                                 spin_unlock(&pool->lock);
 574                                 put_cpu_ptr(pool->unbuddied);
 575                                 goto lookup;
 576                         }
 577                         list_del_init(&zhdr->buddy);
 578                         zhdr->cpu = -1;
 579                         spin_unlock(&pool->lock);
 580
 581                         page = virt_to_page(zhdr);
 582                         if (test_bit(NEEDS_COMPACTING, &page->private)) {
 583                                 z3fold_page_unlock(zhdr);
 584                                 zhdr = NULL;
 585                                 put_cpu_ptr(pool->unbuddied);
 586                                 if (can_sleep)
 587                                         cond_resched();
 588                                 goto lookup;
 589                         }
 590
 591                         /*
 592                          * this page could not be removed from its unbuddied
 593                          * list while pool lock was held, and then we've taken
 594                          * page lock so kref_put could not be called before
 595                          * we got here, so it's safe to just call kref_get()
 596                          */
 597                         kref_get(&zhdr->refcount);
 598                         break;
 599                 }
 600                 put_cpu_ptr(pool->unbuddied);
 601
 602                 if (zhdr) {
 603                         if (zhdr->first_chunks == 0) {
 604                                 if (zhdr->middle_chunks != 0 &&
 605                                     chunks >= zhdr->start_middle)
 606                                         bud = LAST;
 607                                 else
 608                                         bud = FIRST;
 609                         } else if (zhdr->last_chunks == 0)
 610                                 bud = LAST;
 611                         else if (zhdr->middle_chunks == 0)
 612                                 bud = MIDDLE;
 613                         else {
 614                                 if (kref_put(&zhdr->refcount,
 615                                              release_z3fold_page_locked))
 616                                         atomic64_dec(&pool->pages_nr);
 617                                 else
 618                                         z3fold_page_unlock(zhdr);
 619                                 pr_err("No free chunks in unbuddied\n");
 620                                 WARN_ON(1);
 621                                 goto lookup;
 622                         }
 623                         goto found;
 624                 }
 625                 bud = FIRST;
 626         }
 627
 628         page = NULL;
 629         if (can_sleep) {
 630                 spin_lock(&pool->stale_lock);
 631                 zhdr = list_first_entry_or_null(&pool->stale,
 632                                                 struct z3fold_header, buddy);
 633                 /*
 634                  * Before allocating a page, let's see if we can take one from
 635                  * the stale pages list. cancel_work_sync() can sleep so we
 636                  * limit this case to the contexts where we can sleep
 637                  */
 638                 if (zhdr) {
 639                         list_del(&zhdr->buddy);
 640                         spin_unlock(&pool->stale_lock);
 641                         cancel_work_sync(&zhdr->work);
 642                         page = virt_to_page(zhdr);
 643                 } else {
 644                         spin_unlock(&pool->stale_lock);
 645                 }
 646         }
 647         if (!page)
 648                 page = alloc_page(gfp);
 649
 650         if (!page)
 651                 return -ENOMEM;
 652
 653         atomic64_inc(&pool->pages_nr);
 654         zhdr = init_z3fold_page(page, pool);
 655
 656         if (bud == HEADLESS) {
 657                 set_bit(PAGE_HEADLESS, &page->private);
 658                 goto headless;
 659         }
 660         z3fold_page_lock(zhdr);
 661
 662 found:
 663         if (bud == FIRST)
 664                 zhdr->first_chunks = chunks;
 665         else if (bud == LAST)
 666                 zhdr->last_chunks = chunks;
 667         else {
 668                 zhdr->middle_chunks = chunks;
 669                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
 670         }
 671
 672         if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0 ||
 673                         zhdr->middle_chunks == 0) {
 674                 struct list_head *unbuddied = get_cpu_ptr(pool->unbuddied);
 675
 676                 /* Add to unbuddied list */
 677                 freechunks = num_free_chunks(zhdr);
 678                 spin_lock(&pool->lock);
 679                 list_add(&zhdr->buddy, &unbuddied[freechunks]);
 680                 spin_unlock(&pool->lock);
 681                 zhdr->cpu = smp_processor_id();
 682                 put_cpu_ptr(pool->unbuddied);
 683         }
 684
 685 headless:
 686         spin_lock(&pool->lock);
 687         /* Add/move z3fold page to beginning of LRU */
 688         if (!list_empty(&page->lru))
 689                 list_del(&page->lru);
 690
 691         list_add(&page->lru, &pool->lru);
 692
 693         *handle = encode_handle(zhdr, bud);
 694         spin_unlock(&pool->lock);
 695         if (bud != HEADLESS)
 696                 z3fold_page_unlock(zhdr);
 697
 698         return 0;
 699 }
 700
 701 /**
 702  * z3fold_free() - frees the allocation associated with the given handle
 703  * @pool:       pool in which the allocation resided
 704  * @handle:     handle associated with the allocation returned by z3fold_alloc()
 705  *
 706  * In the case that the z3fold page in which the allocation resides is under
 707  * reclaim, as indicated by the PG_reclaim flag being set, this function
 708  * only sets the first|last_chunks to 0.  The page is actually freed
 709  * once both buddies are evicted (see z3fold_reclaim_page() below).
 710  */
 711 static void z3fold_free(struct z3fold_pool *pool, unsigned long handle)
 712 {
 713         struct z3fold_header *zhdr;
 714         struct page *page;
 715         enum buddy bud;
 716
 717         zhdr = handle_to_z3fold_header(handle);
 718         page = virt_to_page(zhdr);
 719
 720         if (test_bit(PAGE_HEADLESS, &page->private)) {
 721                 /* HEADLESS page stored */
 722                 bud = HEADLESS;
 723         } else {
 724                 z3fold_page_lock(zhdr);
 725                 bud = handle_to_buddy(handle);
 726
 727                 switch (bud) {
 728                 case FIRST:
 729                         zhdr->first_chunks = 0;
 730                         break;
 731                 case MIDDLE:
 732                         zhdr->middle_chunks = 0;
 733                         zhdr->start_middle = 0;
 734                         break;
 735                 case LAST:
 736                         zhdr->last_chunks = 0;
 737                         break;
 738                 default:
 739                         pr_err("%s: unknown bud %d\n", __func__, bud);
 740                         WARN_ON(1);
 741                         z3fold_page_unlock(zhdr);
 742                         return;
 743                 }
 744         }
 745
 746         if (bud == HEADLESS) {
 747                 spin_lock(&pool->lock);
 748                 list_del(&page->lru);
 749                 spin_unlock(&pool->lock);
 750                 free_z3fold_page(page);
 751                 atomic64_dec(&pool->pages_nr);
 752                 return;
 753         }
 754
 755         if (kref_put(&zhdr->refcount, release_z3fold_page_locked_list)) {
 756                 atomic64_dec(&pool->pages_nr);
 757                 return;
 758         }
 759         if (test_and_set_bit(NEEDS_COMPACTING, &page->private)) {
 760                 z3fold_page_unlock(zhdr);
 761                 return;
 762         }
 763         if (zhdr->cpu < 0 || !cpu_online(zhdr->cpu)) {
 764                 spin_lock(&pool->lock);
 765                 list_del_init(&zhdr->buddy);
 766                 spin_unlock(&pool->lock);
 767                 zhdr->cpu = -1;
 768                 kref_get(&zhdr->refcount);
 769                 do_compact_page(zhdr, true);
 770                 return;
 771         }
 772         kref_get(&zhdr->refcount);
 773         queue_work_on(zhdr->cpu, pool->compact_wq, &zhdr->work);
 774         z3fold_page_unlock(zhdr);
 775 }
 776
 777 /**
 778  * z3fold_reclaim_page() - evicts allocations from a pool page and frees it
 779  * @pool:       pool from which a page will attempt to be evicted
 780  * @retries:    number of pages on the LRU list for which eviction will
 781  *              be attempted before failing
 782  *
 783  * z3fold reclaim is different from normal system reclaim in that it is done
 784  * from the bottom, up. This is because only the bottom layer, z3fold, has
 785  * information on how the allocations are organized within each z3fold page.
 786  * This has the potential to create interesting locking situations between
 787  * z3fold and the user, however.
 788  *
 789  * To avoid these, this is how z3fold_reclaim_page() should be called:
 790  *
 791  * The user detects a page should be reclaimed and calls z3fold_reclaim_page().
 792  * z3fold_reclaim_page() will remove a z3fold page from the pool LRU list and
 793  * call the user-defined eviction handler with the pool and handle as
 794  * arguments.
 795  *
 796  * If the handle can not be evicted, the eviction handler should return
 797  * non-zero. z3fold_reclaim_page() will add the z3fold page back to the
 798  * appropriate list and try the next z3fold page on the LRU up to
 799  * a user defined number of retries.
 800  *
 801  * If the handle is successfully evicted, the eviction handler should
 802  * return 0 _and_ should have called z3fold_free() on the handle. z3fold_free()
 803  * contains logic to delay freeing the page if the page is under reclaim,
 804  * as indicated by the setting of the PG_reclaim flag on the underlying page.
 805  *
 806  * If all buddies in the z3fold page are successfully evicted, then the
 807  * z3fold page can be freed.
 808  *
 809  * Returns: 0 if page is successfully freed, otherwise -EINVAL if there are
 810  * no pages to evict or an eviction handler is not registered, -EAGAIN if
 811  * the retry limit was hit.
 812  */
 813 static int z3fold_reclaim_page(struct z3fold_pool *pool, unsigned int retries)
 814 {
 815         int i, ret = 0;
 816         struct z3fold_header *zhdr = NULL;
 817         struct page *page = NULL;
 818         struct list_head *pos;
 819         unsigned long first_handle = 0, middle_handle = 0, last_handle = 0;
 820
 821         spin_lock(&pool->lock);
 822         if (!pool->ops || !pool->ops->evict || retries == 0) {
 823                 spin_unlock(&pool->lock);
 824                 return -EINVAL;
 825         }
 826         for (i = 0; i < retries; i++) {
 827                 if (list_empty(&pool->lru)) {
 828                         spin_unlock(&pool->lock);
 829                         return -EINVAL;
 830                 }
 831                 list_for_each_prev(pos, &pool->lru) {
 832                         page = list_entry(pos, struct page, lru);
 833                         if (test_bit(PAGE_HEADLESS, &page->private))
 834                                 /* candidate found */
 835                                 break;
 836
 837                         zhdr = page_address(page);
 838                         if (!z3fold_page_trylock(zhdr))
 839                                 continue; /* can't evict at this point */
 840                         kref_get(&zhdr->refcount);
 841                         list_del_init(&zhdr->buddy);
 842                         zhdr->cpu = -1;
 843                 }
 844
 845                 list_del_init(&page->lru);
 846                 spin_unlock(&pool->lock);
 847
 848                 if (!test_bit(PAGE_HEADLESS, &page->private)) {
 849                         /*
 850                          * We need encode the handles before unlocking, since
 851                          * we can race with free that will set
 852                          * (first|last)_chunks to 0
 853                          */
 854                         first_handle = 0;
 855                         last_handle = 0;
 856                         middle_handle = 0;
 857                         if (zhdr->first_chunks)
 858                                 first_handle = encode_handle(zhdr, FIRST);
 859                         if (zhdr->middle_chunks)
 860                                 middle_handle = encode_handle(zhdr, MIDDLE);
 861                         if (zhdr->last_chunks)
 862                                 last_handle = encode_handle(zhdr, LAST);
 863                         /*
 864                          * it's safe to unlock here because we hold a
 865                          * reference to this page
 866                          */
 867                         z3fold_page_unlock(zhdr);
 868                 } else {
 869                         first_handle = encode_handle(zhdr, HEADLESS);
 870                         last_handle = middle_handle = 0;
 871                 }
 872
 873                 /* Issue the eviction callback(s) */
 874                 if (middle_handle) {
 875                         ret = pool->ops->evict(pool, middle_handle);
 876                         if (ret)
 877                                 goto next;
 878                 }
 879                 if (first_handle) {
 880                         ret = pool->ops->evict(pool, first_handle);
 881                         if (ret)
 882                                 goto next;
 883                 }
 884                 if (last_handle) {
 885                         ret = pool->ops->evict(pool, last_handle);
 886                         if (ret)
 887                                 goto next;
 888                 }
 889 next:
 890                 spin_lock(&pool->lock);
 891                 if (test_bit(PAGE_HEADLESS, &page->private)) {
 892                         if (ret == 0) {
 893                                 spin_unlock(&pool->lock);
 894                                 free_z3fold_page(page);
 895                                 return 0;
 896                         }
 897                 } else if (kref_put(&zhdr->refcount, release_z3fold_page)) {
 898                         atomic64_dec(&pool->pages_nr);
 899                         spin_unlock(&pool->lock);
 900                         return 0;
 901                 }
 902
 903                 /*
 904                  * Add to the beginning of LRU.
 905                  * Pool lock has to be kept here to ensure the page has
 906                  * not already been released
 907                  */
 908                 list_add(&page->lru, &pool->lru);
 909         }
 910         spin_unlock(&pool->lock);
 911         return -EAGAIN;
 912 }
 913
 914 /**
 915  * z3fold_map() - maps the allocation associated with the given handle
 916  * @pool:       pool in which the allocation resides
 917  * @handle:     handle associated with the allocation to be mapped
 918  *
 919  * Extracts the buddy number from handle and constructs the pointer to the
 920  * correct starting chunk within the page.
 921  *
 922  * Returns: a pointer to the mapped allocation
 923  */
 924 static void *z3fold_map(struct z3fold_pool *pool, unsigned long handle)
 925 {
 926         struct z3fold_header *zhdr;
 927         struct page *page;
 928         void *addr;
 929         enum buddy buddy;
 930
 931         zhdr = handle_to_z3fold_header(handle);
 932         addr = zhdr;
 933         page = virt_to_page(zhdr);
 934
 935         if (test_bit(PAGE_HEADLESS, &page->private))
 936                 goto out;
 937
 938         z3fold_page_lock(zhdr);
 939         buddy = handle_to_buddy(handle);
 940         switch (buddy) {
 941         case FIRST:
 942                 addr += ZHDR_SIZE_ALIGNED;
 943                 break;
 944         case MIDDLE:
 945                 addr += zhdr->start_middle << CHUNK_SHIFT;
 946                 set_bit(MIDDLE_CHUNK_MAPPED, &page->private);
 947                 break;
 948         case LAST:
 949                 addr += PAGE_SIZE - (zhdr->last_chunks << CHUNK_SHIFT);
 950                 break;
 951         default:
 952                 pr_err("unknown buddy id %d\n", buddy);
 953                 WARN_ON(1);
 954                 addr = NULL;
 955                 break;
 956         }
 957
 958         z3fold_page_unlock(zhdr);
 959 out:
 960         return addr;
 961 }
 962
 963 /**
 964  * z3fold_unmap() - unmaps the allocation associated with the given handle
 965  * @pool:       pool in which the allocation resides
 966  * @handle:     handle associated with the allocation to be unmapped
 967  */
 968 static void z3fold_unmap(struct z3fold_pool *pool, unsigned long handle)
 969 {
 970         struct z3fold_header *zhdr;
 971         struct page *page;
 972         enum buddy buddy;
 973
 974         zhdr = handle_to_z3fold_header(handle);
 975         page = virt_to_page(zhdr);
 976
 977         if (test_bit(PAGE_HEADLESS, &page->private))
 978                 return;
 979
 980         z3fold_page_lock(zhdr);
 981         buddy = handle_to_buddy(handle);
 982         if (buddy == MIDDLE)
 983                 clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
 984         z3fold_page_unlock(zhdr);
 985 }
 986
 987 /**
 988  * z3fold_get_pool_size() - gets the z3fold pool size in pages
 989  * @pool:       pool whose size is being queried
 990  *
 991  * Returns: size in pages of the given pool.
 992  */
 993 static u64 z3fold_get_pool_size(struct z3fold_pool *pool)
 994 {
 995         return atomic64_read(&pool->pages_nr);
 996 }
 997
 998 /*****************
 999  * zpool
1000  ****************/
1001
1002 static int z3fold_zpool_evict(struct z3fold_pool *pool, unsigned long handle)
1003 {
1004         if (pool->zpool && pool->zpool_ops && pool->zpool_ops->evict)
1005                 return pool->zpool_ops->evict(pool->zpool, handle);
1006         else
1007                 return -ENOENT;
1008 }
1009
1010 static const struct z3fold_ops z3fold_zpool_ops = {
1011         .evict =        z3fold_zpool_evict
1012 };
1013
1014 static void *z3fold_zpool_create(const char *name, gfp_t gfp,
1015                                const struct zpool_ops *zpool_ops,
1016                                struct zpool *zpool)
1017 {
1018         struct z3fold_pool *pool;
1019
1020         pool = z3fold_create_pool(name, gfp,
1021                                 zpool_ops ? &z3fold_zpool_ops : NULL);
1022         if (pool) {
1023                 pool->zpool = zpool;
1024                 pool->zpool_ops = zpool_ops;
1025         }
1026         return pool;
1027 }
1028
1029 static void z3fold_zpool_destroy(void *pool)
1030 {
1031         z3fold_destroy_pool(pool);
1032 }
1033
1034 static int z3fold_zpool_malloc(void *pool, size_t size, gfp_t gfp,
1035                         unsigned long *handle)
1036 {
1037         return z3fold_alloc(pool, size, gfp, handle);
1038 }
1039 static void z3fold_zpool_free(void *pool, unsigned long handle)
1040 {
1041         z3fold_free(pool, handle);
1042 }
1043
1044 static int z3fold_zpool_shrink(void *pool, unsigned int pages,
1045                         unsigned int *reclaimed)
1046 {
1047         unsigned int total = 0;
1048         int ret = -EINVAL;
1049
1050         while (total < pages) {
1051                 ret = z3fold_reclaim_page(pool, 8);
1052                 if (ret < 0)
1053                         break;
1054                 total++;
1055         }
1056
1057         if (reclaimed)
1058                 *reclaimed = total;
1059
1060         return ret;
1061 }
1062
1063 static void *z3fold_zpool_map(void *pool, unsigned long handle,
1064                         enum zpool_mapmode mm)
1065 {
1066         return z3fold_map(pool, handle);
1067 }
1068 static void z3fold_zpool_unmap(void *pool, unsigned long handle)
1069 {
1070         z3fold_unmap(pool, handle);
1071 }
1072
1073 static u64 z3fold_zpool_total_size(void *pool)
1074 {
1075         return z3fold_get_pool_size(pool) * PAGE_SIZE;
1076 }
1077
1078 static struct zpool_driver z3fold_zpool_driver = {
1079         .type =         "z3fold",
1080         .owner =        THIS_MODULE,
1081         .create =       z3fold_zpool_create,
1082         .destroy =      z3fold_zpool_destroy,
1083         .malloc =       z3fold_zpool_malloc,
1084         .free =         z3fold_zpool_free,
1085         .shrink =       z3fold_zpool_shrink,
1086         .map =          z3fold_zpool_map,
1087         .unmap =        z3fold_zpool_unmap,
1088         .total_size =   z3fold_zpool_total_size,
1089 };
1090
1091 MODULE_ALIAS("zpool-z3fold");
1092
1093 static int __init init_z3fold(void)
1094 {
1095         /* Make sure the z3fold header is not larger than the page size */
1096         BUILD_BUG_ON(ZHDR_SIZE_ALIGNED > PAGE_SIZE);
1097         zpool_register_driver(&z3fold_zpool_driver);
1098
1099         return 0;
1100 }
1101
1102 static void __exit exit_z3fold(void)
1103 {
1104         zpool_unregister_driver(&z3fold_zpool_driver);
1105 }
1106
1107 module_init(init_z3fold);
1108 module_exit(exit_z3fold);
1109
1110 MODULE_LICENSE("GPL");
1111 MODULE_AUTHOR("Vitaly Wool <vitalywool@gmail.com>");
1112 MODULE_DESCRIPTION("3-Fold Allocator for Compressed Pages");