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 #define BUDDY_SHIFT     2
 103
 104 /**
 105  * struct z3fold_pool - stores metadata for each z3fold pool
 106  * @name:       pool name
 107  * @lock:       protects pool unbuddied/lru lists
 108  * @stale_lock: protects pool stale page list
 109  * @unbuddied:  per-cpu array of lists tracking z3fold pages that contain 2-
 110  *              buddies; the list each z3fold page is added to depends on
 111  *              the size of its free region.
 112  * @lru:        list tracking the z3fold pages in LRU order by most recently
 113  *              added buddy.
 114  * @stale:      list of pages marked for freeing
 115  * @pages_nr:   number of z3fold pages in the pool.
 116  * @ops:        pointer to a structure of user defined operations specified at
 117  *              pool creation time.
 118  * @compact_wq: workqueue for page layout background optimization
 119  * @release_wq: workqueue for safe page release
 120  * @work:       work_struct for safe page release
 121  *
 122  * This structure is allocated at pool creation time and maintains metadata
 123  * pertaining to a particular z3fold pool.
 124  */
 125 struct z3fold_pool {
 126         const char *name;
 127         spinlock_t lock;
 128         spinlock_t stale_lock;
 129         struct list_head *unbuddied;
 130         struct list_head lru;
 131         struct list_head stale;
 132         atomic64_t pages_nr;
 133         const struct z3fold_ops *ops;
 134         struct zpool *zpool;
 135         const struct zpool_ops *zpool_ops;
 136         struct workqueue_struct *compact_wq;
 137         struct workqueue_struct *release_wq;
 138         struct work_struct work;
 139 };
 140
 141 /*
 142  * Internal z3fold page flags
 143  */
 144 enum z3fold_page_flags {
 145         PAGE_HEADLESS = 0,
 146         MIDDLE_CHUNK_MAPPED,
 147         NEEDS_COMPACTING,
 148         PAGE_STALE,
 149         PAGE_CLAIMED, /* by either reclaim or free */
 150 };
 151
 152 /*****************
 153  * Helpers
 154 *****************/
 155
 156 /* Converts an allocation size in bytes to size in z3fold chunks */
 157 static int size_to_chunks(size_t size)
 158 {
 159         return (size + CHUNK_SIZE - 1) >> CHUNK_SHIFT;
 160 }
 161
 162 #define for_each_unbuddied_list(_iter, _begin) \
 163         for ((_iter) = (_begin); (_iter) < NCHUNKS; (_iter)++)
 164
 165 static void compact_page_work(struct work_struct *w);
 166
 167 /* Initializes the z3fold header of a newly allocated z3fold page */
 168 static struct z3fold_header *init_z3fold_page(struct page *page,
 169                                         struct z3fold_pool *pool)
 170 {
 171         struct z3fold_header *zhdr = page_address(page);
 172
 173         INIT_LIST_HEAD(&page->lru);
 174         clear_bit(PAGE_HEADLESS, &page->private);
 175         clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
 176         clear_bit(NEEDS_COMPACTING, &page->private);
 177         clear_bit(PAGE_STALE, &page->private);
 178         clear_bit(PAGE_CLAIMED, &page->private);
 179
 180         spin_lock_init(&zhdr->page_lock);
 181         kref_init(&zhdr->refcount);
 182         zhdr->first_chunks = 0;
 183         zhdr->middle_chunks = 0;
 184         zhdr->last_chunks = 0;
 185         zhdr->first_num = 0;
 186         zhdr->start_middle = 0;
 187         zhdr->cpu = -1;
 188         zhdr->pool = pool;
 189         INIT_LIST_HEAD(&zhdr->buddy);
 190         INIT_WORK(&zhdr->work, compact_page_work);
 191         return zhdr;
 192 }
 193
 194 /* Resets the struct page fields and frees the page */
 195 static void free_z3fold_page(struct page *page)
 196 {
 197         __free_page(page);
 198 }
 199
 200 /* Lock a z3fold page */
 201 static inline void z3fold_page_lock(struct z3fold_header *zhdr)
 202 {
 203         spin_lock(&zhdr->page_lock);
 204 }
 205
 206 /* Try to lock a z3fold page */
 207 static inline int z3fold_page_trylock(struct z3fold_header *zhdr)
 208 {
 209         return spin_trylock(&zhdr->page_lock);
 210 }
 211
 212 /* Unlock a z3fold page */
 213 static inline void z3fold_page_unlock(struct z3fold_header *zhdr)
 214 {
 215         spin_unlock(&zhdr->page_lock);
 216 }
 217
 218 /*
 219  * Encodes the handle of a particular buddy within a z3fold page
 220  * Pool lock should be held as this function accesses first_num
 221  */
 222 static unsigned long encode_handle(struct z3fold_header *zhdr, enum buddy bud)
 223 {
 224         unsigned long handle;
 225
 226         handle = (unsigned long)zhdr;
 227         if (bud != HEADLESS) {
 228                 handle |= (bud + zhdr->first_num) & BUDDY_MASK;
 229                 if (bud == LAST)
 230                         handle |= (zhdr->last_chunks << BUDDY_SHIFT);
 231         }
 232         return handle;
 233 }
 234
 235 /* Returns the z3fold page where a given handle is stored */
 236 static struct z3fold_header *handle_to_z3fold_header(unsigned long handle)
 237 {
 238         return (struct z3fold_header *)(handle & PAGE_MASK);
 239 }
 240
 241 /* only for LAST bud, returns zero otherwise */
 242 static unsigned short handle_to_chunks(unsigned long handle)
 243 {
 244         return (handle & ~PAGE_MASK) >> BUDDY_SHIFT;
 245 }
 246
 247 /*
 248  * (handle & BUDDY_MASK) < zhdr->first_num is possible in encode_handle
 249  *  but that doesn't matter. because the masking will result in the
 250  *  correct buddy number.
 251  */
 252 static enum buddy handle_to_buddy(unsigned long handle)
 253 {
 254         struct z3fold_header *zhdr = handle_to_z3fold_header(handle);
 255         return (handle - zhdr->first_num) & BUDDY_MASK;
 256 }
 257
 258 static void __release_z3fold_page(struct z3fold_header *zhdr, bool locked)
 259 {
 260         struct page *page = virt_to_page(zhdr);
 261         struct z3fold_pool *pool = zhdr->pool;
 262
 263         WARN_ON(!list_empty(&zhdr->buddy));
 264         set_bit(PAGE_STALE, &page->private);
 265         clear_bit(NEEDS_COMPACTING, &page->private);
 266         spin_lock(&pool->lock);
 267         if (!list_empty(&page->lru))
 268                 list_del(&page->lru);
 269         spin_unlock(&pool->lock);
 270         if (locked)
 271                 z3fold_page_unlock(zhdr);
 272         spin_lock(&pool->stale_lock);
 273         list_add(&zhdr->buddy, &pool->stale);
 274         queue_work(pool->release_wq, &pool->work);
 275         spin_unlock(&pool->stale_lock);
 276 }
 277
 278 static void __attribute__((__unused__))
 279                         release_z3fold_page(struct kref *ref)
 280 {
 281         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 282                                                 refcount);
 283         __release_z3fold_page(zhdr, false);
 284 }
 285
 286 static void release_z3fold_page_locked(struct kref *ref)
 287 {
 288         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 289                                                 refcount);
 290         WARN_ON(z3fold_page_trylock(zhdr));
 291         __release_z3fold_page(zhdr, true);
 292 }
 293
 294 static void release_z3fold_page_locked_list(struct kref *ref)
 295 {
 296         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 297                                                refcount);
 298         spin_lock(&zhdr->pool->lock);
 299         list_del_init(&zhdr->buddy);
 300         spin_unlock(&zhdr->pool->lock);
 301
 302         WARN_ON(z3fold_page_trylock(zhdr));
 303         __release_z3fold_page(zhdr, true);
 304 }
 305
 306 static void free_pages_work(struct work_struct *w)
 307 {
 308         struct z3fold_pool *pool = container_of(w, struct z3fold_pool, work);
 309
 310         spin_lock(&pool->stale_lock);
 311         while (!list_empty(&pool->stale)) {
 312                 struct z3fold_header *zhdr = list_first_entry(&pool->stale,
 313                                                 struct z3fold_header, buddy);
 314                 struct page *page = virt_to_page(zhdr);
 315
 316                 list_del(&zhdr->buddy);
 317                 if (WARN_ON(!test_bit(PAGE_STALE, &page->private)))
 318                         continue;
 319                 spin_unlock(&pool->stale_lock);
 320                 cancel_work_sync(&zhdr->work);
 321                 free_z3fold_page(page);
 322                 cond_resched();
 323                 spin_lock(&pool->stale_lock);
 324         }
 325         spin_unlock(&pool->stale_lock);
 326 }
 327
 328 /*
 329  * Returns the number of free chunks in a z3fold page.
 330  * NB: can't be used with HEADLESS pages.
 331  */
 332 static int num_free_chunks(struct z3fold_header *zhdr)
 333 {
 334         int nfree;
 335         /*
 336          * If there is a middle object, pick up the bigger free space
 337          * either before or after it. Otherwise just subtract the number
 338          * of chunks occupied by the first and the last objects.
 339          */
 340         if (zhdr->middle_chunks != 0) {
 341                 int nfree_before = zhdr->first_chunks ?
 342                         0 : zhdr->start_middle - ZHDR_CHUNKS;
 343                 int nfree_after = zhdr->last_chunks ?
 344                         0 : TOTAL_CHUNKS -
 345                                 (zhdr->start_middle + zhdr->middle_chunks);
 346                 nfree = max(nfree_before, nfree_after);
 347         } else
 348                 nfree = NCHUNKS - zhdr->first_chunks - zhdr->last_chunks;
 349         return nfree;
 350 }
 351
 352 static inline void *mchunk_memmove(struct z3fold_header *zhdr,
 353                                 unsigned short dst_chunk)
 354 {
 355         void *beg = zhdr;
 356         return memmove(beg + (dst_chunk << CHUNK_SHIFT),
 357                        beg + (zhdr->start_middle << CHUNK_SHIFT),
 358                        zhdr->middle_chunks << CHUNK_SHIFT);
 359 }
 360
 361 #define BIG_CHUNK_GAP   3
 362 /* Has to be called with lock held */
 363 static int z3fold_compact_page(struct z3fold_header *zhdr)
 364 {
 365         struct page *page = virt_to_page(zhdr);
 366
 367         if (test_bit(MIDDLE_CHUNK_MAPPED, &page->private))
 368                 return 0; /* can't move middle chunk, it's used */
 369
 370         if (zhdr->middle_chunks == 0)
 371                 return 0; /* nothing to compact */
 372
 373         if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
 374                 /* move to the beginning */
 375                 mchunk_memmove(zhdr, ZHDR_CHUNKS);
 376                 zhdr->first_chunks = zhdr->middle_chunks;
 377                 zhdr->middle_chunks = 0;
 378                 zhdr->start_middle = 0;
 379                 zhdr->first_num++;
 380                 return 1;
 381         }
 382
 383         /*
 384          * moving data is expensive, so let's only do that if
 385          * there's substantial gain (at least BIG_CHUNK_GAP chunks)
 386          */
 387         if (zhdr->first_chunks != 0 && zhdr->last_chunks == 0 &&
 388             zhdr->start_middle - (zhdr->first_chunks + ZHDR_CHUNKS) >=
 389                         BIG_CHUNK_GAP) {
 390                 mchunk_memmove(zhdr, zhdr->first_chunks + ZHDR_CHUNKS);
 391                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
 392                 return 1;
 393         } else if (zhdr->last_chunks != 0 && zhdr->first_chunks == 0 &&
 394                    TOTAL_CHUNKS - (zhdr->last_chunks + zhdr->start_middle
 395                                         + zhdr->middle_chunks) >=
 396                         BIG_CHUNK_GAP) {
 397                 unsigned short new_start = TOTAL_CHUNKS - zhdr->last_chunks -
 398                         zhdr->middle_chunks;
 399                 mchunk_memmove(zhdr, new_start);
 400                 zhdr->start_middle = new_start;
 401                 return 1;
 402         }
 403
 404         return 0;
 405 }
 406
 407 static void do_compact_page(struct z3fold_header *zhdr, bool locked)
 408 {
 409         struct z3fold_pool *pool = zhdr->pool;
 410         struct page *page;
 411         struct list_head *unbuddied;
 412         int fchunks;
 413
 414         page = virt_to_page(zhdr);
 415         if (locked)
 416                 WARN_ON(z3fold_page_trylock(zhdr));
 417         else
 418                 z3fold_page_lock(zhdr);
 419         if (WARN_ON(!test_and_clear_bit(NEEDS_COMPACTING, &page->private))) {
 420                 z3fold_page_unlock(zhdr);
 421                 return;
 422         }
 423         spin_lock(&pool->lock);
 424         list_del_init(&zhdr->buddy);
 425         spin_unlock(&pool->lock);
 426
 427         if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
 428                 atomic64_dec(&pool->pages_nr);
 429                 return;
 430         }
 431
 432         z3fold_compact_page(zhdr);
 433         unbuddied = get_cpu_ptr(pool->unbuddied);
 434         fchunks = num_free_chunks(zhdr);
 435         if (fchunks < NCHUNKS &&
 436             (!zhdr->first_chunks || !zhdr->middle_chunks ||
 437                         !zhdr->last_chunks)) {
 438                 /* the page's not completely free and it's unbuddied */
 439                 spin_lock(&pool->lock);
 440                 list_add(&zhdr->buddy, &unbuddied[fchunks]);
 441                 spin_unlock(&pool->lock);
 442                 zhdr->cpu = smp_processor_id();
 443         }
 444         put_cpu_ptr(pool->unbuddied);
 445         z3fold_page_unlock(zhdr);
 446 }
 447
 448 static void compact_page_work(struct work_struct *w)
 449 {
 450         struct z3fold_header *zhdr = container_of(w, struct z3fold_header,
 451                                                 work);
 452
 453         do_compact_page(zhdr, false);
 454 }
 455
 456
 457 /*
 458  * API Functions
 459  */
 460
 461 /**
 462  * z3fold_create_pool() - create a new z3fold pool
 463  * @name:       pool name
 464  * @gfp:        gfp flags when allocating the z3fold pool structure
 465  * @ops:        user-defined operations for the z3fold pool
 466  *
 467  * Return: pointer to the new z3fold pool or NULL if the metadata allocation
 468  * failed.
 469  */
 470 static struct z3fold_pool *z3fold_create_pool(const char *name, gfp_t gfp,
 471                 const struct z3fold_ops *ops)
 472 {
 473         struct z3fold_pool *pool = NULL;
 474         int i, cpu;
 475
 476         pool = kzalloc(sizeof(struct z3fold_pool), gfp);
 477         if (!pool)
 478                 goto out;
 479         spin_lock_init(&pool->lock);
 480         spin_lock_init(&pool->stale_lock);
 481         pool->unbuddied = __alloc_percpu(sizeof(struct list_head)*NCHUNKS, 2);
 482         if (!pool->unbuddied)
 483                 goto out_pool;
 484         for_each_possible_cpu(cpu) {
 485                 struct list_head *unbuddied =
 486                                 per_cpu_ptr(pool->unbuddied, cpu);
 487                 for_each_unbuddied_list(i, 0)
 488                         INIT_LIST_HEAD(&unbuddied[i]);
 489         }
 490         INIT_LIST_HEAD(&pool->lru);
 491         INIT_LIST_HEAD(&pool->stale);
 492         atomic64_set(&pool->pages_nr, 0);
 493         pool->name = name;
 494         pool->compact_wq = create_singlethread_workqueue(pool->name);
 495         if (!pool->compact_wq)
 496                 goto out_unbuddied;
 497         pool->release_wq = create_singlethread_workqueue(pool->name);
 498         if (!pool->release_wq)
 499                 goto out_wq;
 500         INIT_WORK(&pool->work, free_pages_work);
 501         pool->ops = ops;
 502         return pool;
 503
 504 out_wq:
 505         destroy_workqueue(pool->compact_wq);
 506 out_unbuddied:
 507         free_percpu(pool->unbuddied);
 508 out_pool:
 509         kfree(pool);
 510 out:
 511         return NULL;
 512 }
 513
 514 /**
 515  * z3fold_destroy_pool() - destroys an existing z3fold pool
 516  * @pool:       the z3fold pool to be destroyed
 517  *
 518  * The pool should be emptied before this function is called.
 519  */
 520 static void z3fold_destroy_pool(struct z3fold_pool *pool)
 521 {
 522         destroy_workqueue(pool->release_wq);
 523         destroy_workqueue(pool->compact_wq);
 524         kfree(pool);
 525 }
 526
 527 /**
 528  * z3fold_alloc() - allocates a region of a given size
 529  * @pool:       z3fold pool from which to allocate
 530  * @size:       size in bytes of the desired allocation
 531  * @gfp:        gfp flags used if the pool needs to grow
 532  * @handle:     handle of the new allocation
 533  *
 534  * This function will attempt to find a free region in the pool large enough to
 535  * satisfy the allocation request.  A search of the unbuddied lists is
 536  * performed first. If no suitable free region is found, then a new page is
 537  * allocated and added to the pool to satisfy the request.
 538  *
 539  * gfp should not set __GFP_HIGHMEM as highmem pages cannot be used
 540  * as z3fold pool pages.
 541  *
 542  * Return: 0 if success and handle is set, otherwise -EINVAL if the size or
 543  * gfp arguments are invalid or -ENOMEM if the pool was unable to allocate
 544  * a new page.
 545  */
 546 static int z3fold_alloc(struct z3fold_pool *pool, size_t size, gfp_t gfp,
 547                         unsigned long *handle)
 548 {
 549         int chunks = 0, i, freechunks;
 550         struct z3fold_header *zhdr = NULL;
 551         struct page *page = NULL;
 552         enum buddy bud;
 553         bool can_sleep = gfpflags_allow_blocking(gfp);
 554
 555         if (!size || (gfp & __GFP_HIGHMEM))
 556                 return -EINVAL;
 557
 558         if (size > PAGE_SIZE)
 559                 return -ENOSPC;
 560
 561         if (size > PAGE_SIZE - ZHDR_SIZE_ALIGNED - CHUNK_SIZE)
 562                 bud = HEADLESS;
 563         else {
 564                 struct list_head *unbuddied;
 565                 chunks = size_to_chunks(size);
 566
 567 lookup:
 568                 /* First, try to find an unbuddied z3fold page. */
 569                 unbuddied = get_cpu_ptr(pool->unbuddied);
 570                 for_each_unbuddied_list(i, chunks) {
 571                         struct list_head *l = &unbuddied[i];
 572
 573                         zhdr = list_first_entry_or_null(READ_ONCE(l),
 574                                                 struct z3fold_header, buddy);
 575
 576                         if (!zhdr)
 577                                 continue;
 578
 579                         /* Re-check under lock. */
 580                         spin_lock(&pool->lock);
 581                         l = &unbuddied[i];
 582                         if (unlikely(zhdr != list_first_entry(READ_ONCE(l),
 583                                         struct z3fold_header, buddy)) ||
 584                             !z3fold_page_trylock(zhdr)) {
 585                                 spin_unlock(&pool->lock);
 586                                 put_cpu_ptr(pool->unbuddied);
 587                                 goto lookup;
 588                         }
 589                         list_del_init(&zhdr->buddy);
 590                         zhdr->cpu = -1;
 591                         spin_unlock(&pool->lock);
 592
 593                         page = virt_to_page(zhdr);
 594                         if (test_bit(NEEDS_COMPACTING, &page->private)) {
 595                                 z3fold_page_unlock(zhdr);
 596                                 zhdr = NULL;
 597                                 put_cpu_ptr(pool->unbuddied);
 598                                 if (can_sleep)
 599                                         cond_resched();
 600                                 goto lookup;
 601                         }
 602
 603                         /*
 604                          * this page could not be removed from its unbuddied
 605                          * list while pool lock was held, and then we've taken
 606                          * page lock so kref_put could not be called before
 607                          * we got here, so it's safe to just call kref_get()
 608                          */
 609                         kref_get(&zhdr->refcount);
 610                         break;
 611                 }
 612                 put_cpu_ptr(pool->unbuddied);
 613
 614                 if (zhdr) {
 615                         if (zhdr->first_chunks == 0) {
 616                                 if (zhdr->middle_chunks != 0 &&
 617                                     chunks >= zhdr->start_middle)
 618                                         bud = LAST;
 619                                 else
 620                                         bud = FIRST;
 621                         } else if (zhdr->last_chunks == 0)
 622                                 bud = LAST;
 623                         else if (zhdr->middle_chunks == 0)
 624                                 bud = MIDDLE;
 625                         else {
 626                                 if (kref_put(&zhdr->refcount,
 627                                              release_z3fold_page_locked))
 628                                         atomic64_dec(&pool->pages_nr);
 629                                 else
 630                                         z3fold_page_unlock(zhdr);
 631                                 pr_err("No free chunks in unbuddied\n");
 632                                 WARN_ON(1);
 633                                 goto lookup;
 634                         }
 635                         goto found;
 636                 }
 637                 bud = FIRST;
 638         }
 639
 640         page = NULL;
 641         if (can_sleep) {
 642                 spin_lock(&pool->stale_lock);
 643                 zhdr = list_first_entry_or_null(&pool->stale,
 644                                                 struct z3fold_header, buddy);
 645                 /*
 646                  * Before allocating a page, let's see if we can take one from
 647                  * the stale pages list. cancel_work_sync() can sleep so we
 648                  * limit this case to the contexts where we can sleep
 649                  */
 650                 if (zhdr) {
 651                         list_del(&zhdr->buddy);
 652                         spin_unlock(&pool->stale_lock);
 653                         cancel_work_sync(&zhdr->work);
 654                         page = virt_to_page(zhdr);
 655                 } else {
 656                         spin_unlock(&pool->stale_lock);
 657                 }
 658         }
 659         if (!page)
 660                 page = alloc_page(gfp);
 661
 662         if (!page)
 663                 return -ENOMEM;
 664
 665         atomic64_inc(&pool->pages_nr);
 666         zhdr = init_z3fold_page(page, pool);
 667
 668         if (bud == HEADLESS) {
 669                 set_bit(PAGE_HEADLESS, &page->private);
 670                 goto headless;
 671         }
 672         z3fold_page_lock(zhdr);
 673
 674 found:
 675         if (bud == FIRST)
 676                 zhdr->first_chunks = chunks;
 677         else if (bud == LAST)
 678                 zhdr->last_chunks = chunks;
 679         else {
 680                 zhdr->middle_chunks = chunks;
 681                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
 682         }
 683
 684         if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0 ||
 685                         zhdr->middle_chunks == 0) {
 686                 struct list_head *unbuddied = get_cpu_ptr(pool->unbuddied);
 687
 688                 /* Add to unbuddied list */
 689                 freechunks = num_free_chunks(zhdr);
 690                 spin_lock(&pool->lock);
 691                 list_add(&zhdr->buddy, &unbuddied[freechunks]);
 692                 spin_unlock(&pool->lock);
 693                 zhdr->cpu = smp_processor_id();
 694                 put_cpu_ptr(pool->unbuddied);
 695         }
 696
 697 headless:
 698         spin_lock(&pool->lock);
 699         /* Add/move z3fold page to beginning of LRU */
 700         if (!list_empty(&page->lru))
 701                 list_del(&page->lru);
 702
 703         list_add(&page->lru, &pool->lru);
 704
 705         *handle = encode_handle(zhdr, bud);
 706         spin_unlock(&pool->lock);
 707         if (bud != HEADLESS)
 708                 z3fold_page_unlock(zhdr);
 709
 710         return 0;
 711 }
 712
 713 /**
 714  * z3fold_free() - frees the allocation associated with the given handle
 715  * @pool:       pool in which the allocation resided
 716  * @handle:     handle associated with the allocation returned by z3fold_alloc()
 717  *
 718  * In the case that the z3fold page in which the allocation resides is under
 719  * reclaim, as indicated by the PG_reclaim flag being set, this function
 720  * only sets the first|last_chunks to 0.  The page is actually freed
 721  * once both buddies are evicted (see z3fold_reclaim_page() below).
 722  */
 723 static void z3fold_free(struct z3fold_pool *pool, unsigned long handle)
 724 {
 725         struct z3fold_header *zhdr;
 726         struct page *page;
 727         enum buddy bud;
 728
 729         zhdr = handle_to_z3fold_header(handle);
 730         page = virt_to_page(zhdr);
 731
 732         if (test_bit(PAGE_HEADLESS, &page->private)) {
 733                 /* if a headless page is under reclaim, just leave.
 734                  * NB: we use test_and_set_bit for a reason: if the bit
 735                  * has not been set before, we release this page
 736                  * immediately so we don't care about its value any more.
 737                  */
 738                 if (!test_and_set_bit(PAGE_CLAIMED, &page->private)) {
 739                         spin_lock(&pool->lock);
 740                         list_del(&page->lru);
 741                         spin_unlock(&pool->lock);
 742                         free_z3fold_page(page);
 743                         atomic64_dec(&pool->pages_nr);
 744                 }
 745                 return;
 746         }
 747
 748         /* Non-headless case */
 749         z3fold_page_lock(zhdr);
 750         bud = handle_to_buddy(handle);
 751
 752         switch (bud) {
 753         case FIRST:
 754                 zhdr->first_chunks = 0;
 755                 break;
 756         case MIDDLE:
 757                 zhdr->middle_chunks = 0;
 758                 break;
 759         case LAST:
 760                 zhdr->last_chunks = 0;
 761                 break;
 762         default:
 763                 pr_err("%s: unknown bud %d\n", __func__, bud);
 764                 WARN_ON(1);
 765                 z3fold_page_unlock(zhdr);
 766                 return;
 767         }
 768
 769         if (kref_put(&zhdr->refcount, release_z3fold_page_locked_list)) {
 770                 atomic64_dec(&pool->pages_nr);
 771                 return;
 772         }
 773         if (test_bit(PAGE_CLAIMED, &page->private)) {
 774                 z3fold_page_unlock(zhdr);
 775                 return;
 776         }
 777         if (test_and_set_bit(NEEDS_COMPACTING, &page->private)) {
 778                 z3fold_page_unlock(zhdr);
 779                 return;
 780         }
 781         if (zhdr->cpu < 0 || !cpu_online(zhdr->cpu)) {
 782                 spin_lock(&pool->lock);
 783                 list_del_init(&zhdr->buddy);
 784                 spin_unlock(&pool->lock);
 785                 zhdr->cpu = -1;
 786                 kref_get(&zhdr->refcount);
 787                 do_compact_page(zhdr, true);
 788                 return;
 789         }
 790         kref_get(&zhdr->refcount);
 791         queue_work_on(zhdr->cpu, pool->compact_wq, &zhdr->work);
 792         z3fold_page_unlock(zhdr);
 793 }
 794
 795 /**
 796  * z3fold_reclaim_page() - evicts allocations from a pool page and frees it
 797  * @pool:       pool from which a page will attempt to be evicted
 798  * @retries:    number of pages on the LRU list for which eviction will
 799  *              be attempted before failing
 800  *
 801  * z3fold reclaim is different from normal system reclaim in that it is done
 802  * from the bottom, up. This is because only the bottom layer, z3fold, has
 803  * information on how the allocations are organized within each z3fold page.
 804  * This has the potential to create interesting locking situations between
 805  * z3fold and the user, however.
 806  *
 807  * To avoid these, this is how z3fold_reclaim_page() should be called:
 808  *
 809  * The user detects a page should be reclaimed and calls z3fold_reclaim_page().
 810  * z3fold_reclaim_page() will remove a z3fold page from the pool LRU list and
 811  * call the user-defined eviction handler with the pool and handle as
 812  * arguments.
 813  *
 814  * If the handle can not be evicted, the eviction handler should return
 815  * non-zero. z3fold_reclaim_page() will add the z3fold page back to the
 816  * appropriate list and try the next z3fold page on the LRU up to
 817  * a user defined number of retries.
 818  *
 819  * If the handle is successfully evicted, the eviction handler should
 820  * return 0 _and_ should have called z3fold_free() on the handle. z3fold_free()
 821  * contains logic to delay freeing the page if the page is under reclaim,
 822  * as indicated by the setting of the PG_reclaim flag on the underlying page.
 823  *
 824  * If all buddies in the z3fold page are successfully evicted, then the
 825  * z3fold page can be freed.
 826  *
 827  * Returns: 0 if page is successfully freed, otherwise -EINVAL if there are
 828  * no pages to evict or an eviction handler is not registered, -EAGAIN if
 829  * the retry limit was hit.
 830  */
 831 static int z3fold_reclaim_page(struct z3fold_pool *pool, unsigned int retries)
 832 {
 833         int i, ret = 0;
 834         struct z3fold_header *zhdr = NULL;
 835         struct page *page = NULL;
 836         struct list_head *pos;
 837         unsigned long first_handle = 0, middle_handle = 0, last_handle = 0;
 838
 839         spin_lock(&pool->lock);
 840         if (!pool->ops || !pool->ops->evict || retries == 0) {
 841                 spin_unlock(&pool->lock);
 842                 return -EINVAL;
 843         }
 844         for (i = 0; i < retries; i++) {
 845                 if (list_empty(&pool->lru)) {
 846                         spin_unlock(&pool->lock);
 847                         return -EINVAL;
 848                 }
 849                 list_for_each_prev(pos, &pool->lru) {
 850                         page = list_entry(pos, struct page, lru);
 851
 852                         /* this bit could have been set by free, in which case
 853                          * we pass over to the next page in the pool.
 854                          */
 855                         if (test_and_set_bit(PAGE_CLAIMED, &page->private))
 856                                 continue;
 857
 858                         zhdr = page_address(page);
 859                         if (test_bit(PAGE_HEADLESS, &page->private))
 860                                 break;
 861
 862                         if (!z3fold_page_trylock(zhdr)) {
 863                                 zhdr = NULL;
 864                                 continue; /* can't evict at this point */
 865                         }
 866                         kref_get(&zhdr->refcount);
 867                         list_del_init(&zhdr->buddy);
 868                         zhdr->cpu = -1;
 869                         break;
 870                 }
 871
 872                 if (!zhdr)
 873                         break;
 874
 875                 list_del_init(&page->lru);
 876                 spin_unlock(&pool->lock);
 877
 878                 if (!test_bit(PAGE_HEADLESS, &page->private)) {
 879                         /*
 880                          * We need encode the handles before unlocking, since
 881                          * we can race with free that will set
 882                          * (first|last)_chunks to 0
 883                          */
 884                         first_handle = 0;
 885                         last_handle = 0;
 886                         middle_handle = 0;
 887                         if (zhdr->first_chunks)
 888                                 first_handle = encode_handle(zhdr, FIRST);
 889                         if (zhdr->middle_chunks)
 890                                 middle_handle = encode_handle(zhdr, MIDDLE);
 891                         if (zhdr->last_chunks)
 892                                 last_handle = encode_handle(zhdr, LAST);
 893                         /*
 894                          * it's safe to unlock here because we hold a
 895                          * reference to this page
 896                          */
 897                         z3fold_page_unlock(zhdr);
 898                 } else {
 899                         first_handle = encode_handle(zhdr, HEADLESS);
 900                         last_handle = middle_handle = 0;
 901                 }
 902
 903                 /* Issue the eviction callback(s) */
 904                 if (middle_handle) {
 905                         ret = pool->ops->evict(pool, middle_handle);
 906                         if (ret)
 907                                 goto next;
 908                 }
 909                 if (first_handle) {
 910                         ret = pool->ops->evict(pool, first_handle);
 911                         if (ret)
 912                                 goto next;
 913                 }
 914                 if (last_handle) {
 915                         ret = pool->ops->evict(pool, last_handle);
 916                         if (ret)
 917                                 goto next;
 918                 }
 919 next:
 920                 if (test_bit(PAGE_HEADLESS, &page->private)) {
 921                         if (ret == 0) {
 922                                 free_z3fold_page(page);
 923                                 atomic64_dec(&pool->pages_nr);
 924                                 return 0;
 925                         }
 926                         spin_lock(&pool->lock);
 927                         list_add(&page->lru, &pool->lru);
 928                         spin_unlock(&pool->lock);
 929                 } else {
 930                         z3fold_page_lock(zhdr);
 931                         clear_bit(PAGE_CLAIMED, &page->private);
 932                         if (kref_put(&zhdr->refcount,
 933                                         release_z3fold_page_locked)) {
 934                                 atomic64_dec(&pool->pages_nr);
 935                                 return 0;
 936                         }
 937                         /*
 938                          * if we are here, the page is still not completely
 939                          * free. Take the global pool lock then to be able
 940                          * to add it back to the lru list
 941                          */
 942                         spin_lock(&pool->lock);
 943                         list_add(&page->lru, &pool->lru);
 944                         spin_unlock(&pool->lock);
 945                         z3fold_page_unlock(zhdr);
 946                 }
 947
 948                 /* We started off locked to we need to lock the pool back */
 949                 spin_lock(&pool->lock);
 950         }
 951         spin_unlock(&pool->lock);
 952         return -EAGAIN;
 953 }
 954
 955 /**
 956  * z3fold_map() - maps the allocation associated with the given handle
 957  * @pool:       pool in which the allocation resides
 958  * @handle:     handle associated with the allocation to be mapped
 959  *
 960  * Extracts the buddy number from handle and constructs the pointer to the
 961  * correct starting chunk within the page.
 962  *
 963  * Returns: a pointer to the mapped allocation
 964  */
 965 static void *z3fold_map(struct z3fold_pool *pool, unsigned long handle)
 966 {
 967         struct z3fold_header *zhdr;
 968         struct page *page;
 969         void *addr;
 970         enum buddy buddy;
 971
 972         zhdr = handle_to_z3fold_header(handle);
 973         addr = zhdr;
 974         page = virt_to_page(zhdr);
 975
 976         if (test_bit(PAGE_HEADLESS, &page->private))
 977                 goto out;
 978
 979         z3fold_page_lock(zhdr);
 980         buddy = handle_to_buddy(handle);
 981         switch (buddy) {
 982         case FIRST:
 983                 addr += ZHDR_SIZE_ALIGNED;
 984                 break;
 985         case MIDDLE:
 986                 addr += zhdr->start_middle << CHUNK_SHIFT;
 987                 set_bit(MIDDLE_CHUNK_MAPPED, &page->private);
 988                 break;
 989         case LAST:
 990                 addr += PAGE_SIZE - (handle_to_chunks(handle) << CHUNK_SHIFT);
 991                 break;
 992         default:
 993                 pr_err("unknown buddy id %d\n", buddy);
 994                 WARN_ON(1);
 995                 addr = NULL;
 996                 break;
 997         }
 998
 999         z3fold_page_unlock(zhdr);
1000 out:
1001         return addr;
1002 }
1003
1004 /**
1005  * z3fold_unmap() - unmaps the allocation associated with the given handle
1006  * @pool:       pool in which the allocation resides
1007  * @handle:     handle associated with the allocation to be unmapped
1008  */
1009 static void z3fold_unmap(struct z3fold_pool *pool, unsigned long handle)
1010 {
1011         struct z3fold_header *zhdr;
1012         struct page *page;
1013         enum buddy buddy;
1014
1015         zhdr = handle_to_z3fold_header(handle);
1016         page = virt_to_page(zhdr);
1017
1018         if (test_bit(PAGE_HEADLESS, &page->private))
1019                 return;
1020
1021         z3fold_page_lock(zhdr);
1022         buddy = handle_to_buddy(handle);
1023         if (buddy == MIDDLE)
1024                 clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1025         z3fold_page_unlock(zhdr);
1026 }
1027
1028 /**
1029  * z3fold_get_pool_size() - gets the z3fold pool size in pages
1030  * @pool:       pool whose size is being queried
1031  *
1032  * Returns: size in pages of the given pool.
1033  */
1034 static u64 z3fold_get_pool_size(struct z3fold_pool *pool)
1035 {
1036         return atomic64_read(&pool->pages_nr);
1037 }
1038
1039 /*****************
1040  * zpool
1041  ****************/
1042
1043 static int z3fold_zpool_evict(struct z3fold_pool *pool, unsigned long handle)
1044 {
1045         if (pool->zpool && pool->zpool_ops && pool->zpool_ops->evict)
1046                 return pool->zpool_ops->evict(pool->zpool, handle);
1047         else
1048                 return -ENOENT;
1049 }
1050
1051 static const struct z3fold_ops z3fold_zpool_ops = {
1052         .evict =        z3fold_zpool_evict
1053 };
1054
1055 static void *z3fold_zpool_create(const char *name, gfp_t gfp,
1056                                const struct zpool_ops *zpool_ops,
1057                                struct zpool *zpool)
1058 {
1059         struct z3fold_pool *pool;
1060
1061         pool = z3fold_create_pool(name, gfp,
1062                                 zpool_ops ? &z3fold_zpool_ops : NULL);
1063         if (pool) {
1064                 pool->zpool = zpool;
1065                 pool->zpool_ops = zpool_ops;
1066         }
1067         return pool;
1068 }
1069
1070 static void z3fold_zpool_destroy(void *pool)
1071 {
1072         z3fold_destroy_pool(pool);
1073 }
1074
1075 static int z3fold_zpool_malloc(void *pool, size_t size, gfp_t gfp,
1076                         unsigned long *handle)
1077 {
1078         return z3fold_alloc(pool, size, gfp, handle);
1079 }
1080 static void z3fold_zpool_free(void *pool, unsigned long handle)
1081 {
1082         z3fold_free(pool, handle);
1083 }
1084
1085 static int z3fold_zpool_shrink(void *pool, unsigned int pages,
1086                         unsigned int *reclaimed)
1087 {
1088         unsigned int total = 0;
1089         int ret = -EINVAL;
1090
1091         while (total < pages) {
1092                 ret = z3fold_reclaim_page(pool, 8);
1093                 if (ret < 0)
1094                         break;
1095                 total++;
1096         }
1097
1098         if (reclaimed)
1099                 *reclaimed = total;
1100
1101         return ret;
1102 }
1103
1104 static void *z3fold_zpool_map(void *pool, unsigned long handle,
1105                         enum zpool_mapmode mm)
1106 {
1107         return z3fold_map(pool, handle);
1108 }
1109 static void z3fold_zpool_unmap(void *pool, unsigned long handle)
1110 {
1111         z3fold_unmap(pool, handle);
1112 }
1113
1114 static u64 z3fold_zpool_total_size(void *pool)
1115 {
1116         return z3fold_get_pool_size(pool) * PAGE_SIZE;
1117 }
1118
1119 static struct zpool_driver z3fold_zpool_driver = {
1120         .type =         "z3fold",
1121         .owner =        THIS_MODULE,
1122         .create =       z3fold_zpool_create,
1123         .destroy =      z3fold_zpool_destroy,
1124         .malloc =       z3fold_zpool_malloc,
1125         .free =         z3fold_zpool_free,
1126         .shrink =       z3fold_zpool_shrink,
1127         .map =          z3fold_zpool_map,
1128         .unmap =        z3fold_zpool_unmap,
1129         .total_size =   z3fold_zpool_total_size,
1130 };
1131
1132 MODULE_ALIAS("zpool-z3fold");
1133
1134 static int __init init_z3fold(void)
1135 {
1136         /* Make sure the z3fold header is not larger than the page size */
1137         BUILD_BUG_ON(ZHDR_SIZE_ALIGNED > PAGE_SIZE);
1138         zpool_register_driver(&z3fold_zpool_driver);
1139
1140         return 0;
1141 }
1142
1143 static void __exit exit_z3fold(void)
1144 {
1145         zpool_unregister_driver(&z3fold_zpool_driver);
1146 }
1147
1148 module_init(init_z3fold);
1149 module_exit(exit_z3fold);
1150
1151 MODULE_LICENSE("GPL");
1152 MODULE_AUTHOR("Vitaly Wool <vitalywool@gmail.com>");
1153 MODULE_DESCRIPTION("3-Fold Allocator for Compressed Pages");