mm/z3fold.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * z3fold.c
   4  *
   5  * Author: Vitaly Wool <vitaly.wool@konsulko.com>
   6  * Copyright (C) 2016, Sony Mobile Communications Inc.
   7  *
   8  * This implementation is based on zbud written by Seth Jennings.
   9  *
  10  * z3fold is an special purpose allocator for storing compressed pages. It
  11  * can store up to three compressed pages per page which improves the
  12  * compression ratio of zbud while retaining its main concepts (e. g. always
  13  * storing an integral number of objects per page) and simplicity.
  14  * It still has simple and deterministic reclaim properties that make it
  15  * preferable to a higher density approach (with no requirement on integral
  16  * number of object per page) when reclaim is used.
  17  *
  18  * As in zbud, pages are divided into "chunks".  The size of the chunks is
  19  * fixed at compile time and is determined by NCHUNKS_ORDER below.
  20  *
  21  * z3fold doesn't export any API and is meant to be used via zpool API.
  22  */
  23
  24 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  25
  26 #include <linux/atomic.h>
  27 #include <linux/sched.h>
  28 #include <linux/cpumask.h>
  29 #include <linux/dcache.h>
  30 #include <linux/list.h>
  31 #include <linux/mm.h>
  32 #include <linux/module.h>
  33 #include <linux/page-flags.h>
  34 #include <linux/migrate.h>
  35 #include <linux/node.h>
  36 #include <linux/compaction.h>
  37 #include <linux/percpu.h>
  38 #include <linux/mount.h>
  39 #include <linux/fs.h>
  40 #include <linux/preempt.h>
  41 #include <linux/workqueue.h>
  42 #include <linux/slab.h>
  43 #include <linux/spinlock.h>
  44 #include <linux/zpool.h>
  45
  46 /*
  47  * NCHUNKS_ORDER determines the internal allocation granularity, effectively
  48  * adjusting internal fragmentation.  It also determines the number of
  49  * freelists maintained in each pool. NCHUNKS_ORDER of 6 means that the
  50  * allocation granularity will be in chunks of size PAGE_SIZE/64. Some chunks
  51  * in the beginning of an allocated page are occupied by z3fold header, so
  52  * NCHUNKS will be calculated to 63 (or 62 in case CONFIG_DEBUG_SPINLOCK=y),
  53  * which shows the max number of free chunks in z3fold page, also there will
  54  * be 63, or 62, respectively, freelists per pool.
  55  */
  56 #define NCHUNKS_ORDER   6
  57
  58 #define CHUNK_SHIFT     (PAGE_SHIFT - NCHUNKS_ORDER)
  59 #define CHUNK_SIZE      (1 << CHUNK_SHIFT)
  60 #define ZHDR_SIZE_ALIGNED round_up(sizeof(struct z3fold_header), CHUNK_SIZE)
  61 #define ZHDR_CHUNKS     (ZHDR_SIZE_ALIGNED >> CHUNK_SHIFT)
  62 #define TOTAL_CHUNKS    (PAGE_SIZE >> CHUNK_SHIFT)
  63 #define NCHUNKS         ((PAGE_SIZE - ZHDR_SIZE_ALIGNED) >> CHUNK_SHIFT)
  64
  65 #define BUDDY_MASK      (0x3)
  66 #define BUDDY_SHIFT     2
  67 #define SLOTS_ALIGN     (0x40)
  68
  69 /*****************
  70  * Structures
  71 *****************/
  72 struct z3fold_pool;
  73 struct z3fold_ops {
  74         int (*evict)(struct z3fold_pool *pool, unsigned long handle);
  75 };
  76
  77 enum buddy {
  78         HEADLESS = 0,
  79         FIRST,
  80         MIDDLE,
  81         LAST,
  82         BUDDIES_MAX = LAST
  83 };
  84
  85 struct z3fold_buddy_slots {
  86         /*
  87          * we are using BUDDY_MASK in handle_to_buddy etc. so there should
  88          * be enough slots to hold all possible variants
  89          */
  90         unsigned long slot[BUDDY_MASK + 1];
  91         unsigned long pool; /* back link + flags */
  92 };
  93 #define HANDLE_FLAG_MASK        (0x03)
  94
  95 /*
  96  * struct z3fold_header - z3fold page metadata occupying first chunks of each
  97  *                      z3fold page, except for HEADLESS pages
  98  * @buddy:              links the z3fold page into the relevant list in the
  99  *                      pool
 100  * @page_lock:          per-page lock
 101  * @refcount:           reference count for the z3fold page
 102  * @work:               work_struct for page layout optimization
 103  * @slots:              pointer to the structure holding buddy slots
 104  * @cpu:                CPU which this page "belongs" to
 105  * @first_chunks:       the size of the first buddy in chunks, 0 if free
 106  * @middle_chunks:      the size of the middle buddy in chunks, 0 if free
 107  * @last_chunks:        the size of the last buddy in chunks, 0 if free
 108  * @first_num:          the starting number (for the first handle)
 109  * @mapped_count:       the number of objects currently mapped
 110  */
 111 struct z3fold_header {
 112         struct list_head buddy;
 113         spinlock_t page_lock;
 114         struct kref refcount;
 115         struct work_struct work;
 116         struct z3fold_buddy_slots *slots;
 117         short cpu;
 118         unsigned short first_chunks;
 119         unsigned short middle_chunks;
 120         unsigned short last_chunks;
 121         unsigned short start_middle;
 122         unsigned short first_num:2;
 123         unsigned short mapped_count:2;
 124 };
 125
 126 /**
 127  * struct z3fold_pool - stores metadata for each z3fold pool
 128  * @name:       pool name
 129  * @lock:       protects pool unbuddied/lru lists
 130  * @stale_lock: protects pool stale page list
 131  * @unbuddied:  per-cpu array of lists tracking z3fold pages that contain 2-
 132  *              buddies; the list each z3fold page is added to depends on
 133  *              the size of its free region.
 134  * @lru:        list tracking the z3fold pages in LRU order by most recently
 135  *              added buddy.
 136  * @stale:      list of pages marked for freeing
 137  * @pages_nr:   number of z3fold pages in the pool.
 138  * @c_handle:   cache for z3fold_buddy_slots allocation
 139  * @ops:        pointer to a structure of user defined operations specified at
 140  *              pool creation time.
 141  * @compact_wq: workqueue for page layout background optimization
 142  * @release_wq: workqueue for safe page release
 143  * @work:       work_struct for safe page release
 144  * @inode:      inode for z3fold pseudo filesystem
 145  *
 146  * This structure is allocated at pool creation time and maintains metadata
 147  * pertaining to a particular z3fold pool.
 148  */
 149 struct z3fold_pool {
 150         const char *name;
 151         spinlock_t lock;
 152         spinlock_t stale_lock;
 153         struct list_head *unbuddied;
 154         struct list_head lru;
 155         struct list_head stale;
 156         atomic64_t pages_nr;
 157         struct kmem_cache *c_handle;
 158         const struct z3fold_ops *ops;
 159         struct zpool *zpool;
 160         const struct zpool_ops *zpool_ops;
 161         struct workqueue_struct *compact_wq;
 162         struct workqueue_struct *release_wq;
 163         struct work_struct work;
 164         struct inode *inode;
 165 };
 166
 167 /*
 168  * Internal z3fold page flags
 169  */
 170 enum z3fold_page_flags {
 171         PAGE_HEADLESS = 0,
 172         MIDDLE_CHUNK_MAPPED,
 173         NEEDS_COMPACTING,
 174         PAGE_STALE,
 175         PAGE_CLAIMED, /* by either reclaim or free */
 176 };
 177
 178 /*****************
 179  * Helpers
 180 *****************/
 181
 182 /* Converts an allocation size in bytes to size in z3fold chunks */
 183 static int size_to_chunks(size_t size)
 184 {
 185         return (size + CHUNK_SIZE - 1) >> CHUNK_SHIFT;
 186 }
 187
 188 #define for_each_unbuddied_list(_iter, _begin) \
 189         for ((_iter) = (_begin); (_iter) < NCHUNKS; (_iter)++)
 190
 191 static void compact_page_work(struct work_struct *w);
 192
 193 static inline struct z3fold_buddy_slots *alloc_slots(struct z3fold_pool *pool,
 194                                                         gfp_t gfp)
 195 {
 196         struct z3fold_buddy_slots *slots = kmem_cache_alloc(pool->c_handle,
 197                                                             gfp);
 198
 199         if (slots) {
 200                 memset(slots->slot, 0, sizeof(slots->slot));
 201                 slots->pool = (unsigned long)pool;
 202         }
 203
 204         return slots;
 205 }
 206
 207 static inline struct z3fold_pool *slots_to_pool(struct z3fold_buddy_slots *s)
 208 {
 209         return (struct z3fold_pool *)(s->pool & ~HANDLE_FLAG_MASK);
 210 }
 211
 212 static inline struct z3fold_buddy_slots *handle_to_slots(unsigned long handle)
 213 {
 214         return (struct z3fold_buddy_slots *)(handle & ~(SLOTS_ALIGN - 1));
 215 }
 216
 217 static inline void free_handle(unsigned long handle)
 218 {
 219         struct z3fold_buddy_slots *slots;
 220         int i;
 221         bool is_free;
 222
 223         if (handle & (1 << PAGE_HEADLESS))
 224                 return;
 225
 226         WARN_ON(*(unsigned long *)handle == 0);
 227         *(unsigned long *)handle = 0;
 228         slots = handle_to_slots(handle);
 229         is_free = true;
 230         for (i = 0; i <= BUDDY_MASK; i++) {
 231                 if (slots->slot[i]) {
 232                         is_free = false;
 233                         break;
 234                 }
 235         }
 236
 237         if (is_free) {
 238                 struct z3fold_pool *pool = slots_to_pool(slots);
 239
 240                 kmem_cache_free(pool->c_handle, slots);
 241         }
 242 }
 243
 244 static struct dentry *z3fold_do_mount(struct file_system_type *fs_type,
 245                                 int flags, const char *dev_name, void *data)
 246 {
 247         static const struct dentry_operations ops = {
 248                 .d_dname = simple_dname,
 249         };
 250
 251         return mount_pseudo(fs_type, "z3fold:", NULL, &ops, 0x33);
 252 }
 253
 254 static struct file_system_type z3fold_fs = {
 255         .name           = "z3fold",
 256         .mount          = z3fold_do_mount,
 257         .kill_sb        = kill_anon_super,
 258 };
 259
 260 static struct vfsmount *z3fold_mnt;
 261 static int z3fold_mount(void)
 262 {
 263         int ret = 0;
 264
 265         z3fold_mnt = kern_mount(&z3fold_fs);
 266         if (IS_ERR(z3fold_mnt))
 267                 ret = PTR_ERR(z3fold_mnt);
 268
 269         return ret;
 270 }
 271
 272 static void z3fold_unmount(void)
 273 {
 274         kern_unmount(z3fold_mnt);
 275 }
 276
 277 static const struct address_space_operations z3fold_aops;
 278 static int z3fold_register_migration(struct z3fold_pool *pool)
 279 {
 280         pool->inode = alloc_anon_inode(z3fold_mnt->mnt_sb);
 281         if (IS_ERR(pool->inode)) {
 282                 pool->inode = NULL;
 283                 return 1;
 284         }
 285
 286         pool->inode->i_mapping->private_data = pool;
 287         pool->inode->i_mapping->a_ops = &z3fold_aops;
 288         return 0;
 289 }
 290
 291 static void z3fold_unregister_migration(struct z3fold_pool *pool)
 292 {
 293         if (pool->inode)
 294                 iput(pool->inode);
 295  }
 296
 297 /* Initializes the z3fold header of a newly allocated z3fold page */
 298 static struct z3fold_header *init_z3fold_page(struct page *page,
 299                                         struct z3fold_pool *pool, gfp_t gfp)
 300 {
 301         struct z3fold_header *zhdr = page_address(page);
 302         struct z3fold_buddy_slots *slots = alloc_slots(pool, gfp);
 303
 304         if (!slots)
 305                 return NULL;
 306
 307         INIT_LIST_HEAD(&page->lru);
 308         clear_bit(PAGE_HEADLESS, &page->private);
 309         clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
 310         clear_bit(NEEDS_COMPACTING, &page->private);
 311         clear_bit(PAGE_STALE, &page->private);
 312         clear_bit(PAGE_CLAIMED, &page->private);
 313
 314         spin_lock_init(&zhdr->page_lock);
 315         kref_init(&zhdr->refcount);
 316         zhdr->first_chunks = 0;
 317         zhdr->middle_chunks = 0;
 318         zhdr->last_chunks = 0;
 319         zhdr->first_num = 0;
 320         zhdr->start_middle = 0;
 321         zhdr->cpu = -1;
 322         zhdr->slots = slots;
 323         INIT_LIST_HEAD(&zhdr->buddy);
 324         INIT_WORK(&zhdr->work, compact_page_work);
 325         return zhdr;
 326 }
 327
 328 /* Resets the struct page fields and frees the page */
 329 static void free_z3fold_page(struct page *page, bool headless)
 330 {
 331         if (!headless) {
 332                 lock_page(page);
 333                 __ClearPageMovable(page);
 334                 unlock_page(page);
 335         }
 336         ClearPagePrivate(page);
 337         __free_page(page);
 338 }
 339
 340 /* Lock a z3fold page */
 341 static inline void z3fold_page_lock(struct z3fold_header *zhdr)
 342 {
 343         spin_lock(&zhdr->page_lock);
 344 }
 345
 346 /* Try to lock a z3fold page */
 347 static inline int z3fold_page_trylock(struct z3fold_header *zhdr)
 348 {
 349         return spin_trylock(&zhdr->page_lock);
 350 }
 351
 352 /* Unlock a z3fold page */
 353 static inline void z3fold_page_unlock(struct z3fold_header *zhdr)
 354 {
 355         spin_unlock(&zhdr->page_lock);
 356 }
 357
 358 /* Helper function to build the index */
 359 static inline int __idx(struct z3fold_header *zhdr, enum buddy bud)
 360 {
 361         return (bud + zhdr->first_num) & BUDDY_MASK;
 362 }
 363
 364 /*
 365  * Encodes the handle of a particular buddy within a z3fold page
 366  * Pool lock should be held as this function accesses first_num
 367  */
 368 static unsigned long encode_handle(struct z3fold_header *zhdr, enum buddy bud)
 369 {
 370         struct z3fold_buddy_slots *slots;
 371         unsigned long h = (unsigned long)zhdr;
 372         int idx = 0;
 373
 374         /*
 375          * For a headless page, its handle is its pointer with the extra
 376          * PAGE_HEADLESS bit set
 377          */
 378         if (bud == HEADLESS)
 379                 return h | (1 << PAGE_HEADLESS);
 380
 381         /* otherwise, return pointer to encoded handle */
 382         idx = __idx(zhdr, bud);
 383         h += idx;
 384         if (bud == LAST)
 385                 h |= (zhdr->last_chunks << BUDDY_SHIFT);
 386
 387         slots = zhdr->slots;
 388         slots->slot[idx] = h;
 389         return (unsigned long)&slots->slot[idx];
 390 }
 391
 392 /* Returns the z3fold page where a given handle is stored */
 393 static inline struct z3fold_header *handle_to_z3fold_header(unsigned long h)
 394 {
 395         unsigned long addr = h;
 396
 397         if (!(addr & (1 << PAGE_HEADLESS)))
 398                 addr = *(unsigned long *)h;
 399
 400         return (struct z3fold_header *)(addr & PAGE_MASK);
 401 }
 402
 403 /* only for LAST bud, returns zero otherwise */
 404 static unsigned short handle_to_chunks(unsigned long handle)
 405 {
 406         unsigned long addr = *(unsigned long *)handle;
 407
 408         return (addr & ~PAGE_MASK) >> BUDDY_SHIFT;
 409 }
 410
 411 /*
 412  * (handle & BUDDY_MASK) < zhdr->first_num is possible in encode_handle
 413  *  but that doesn't matter. because the masking will result in the
 414  *  correct buddy number.
 415  */
 416 static enum buddy handle_to_buddy(unsigned long handle)
 417 {
 418         struct z3fold_header *zhdr;
 419         unsigned long addr;
 420
 421         WARN_ON(handle & (1 << PAGE_HEADLESS));
 422         addr = *(unsigned long *)handle;
 423         zhdr = (struct z3fold_header *)(addr & PAGE_MASK);
 424         return (addr - zhdr->first_num) & BUDDY_MASK;
 425 }
 426
 427 static inline struct z3fold_pool *zhdr_to_pool(struct z3fold_header *zhdr)
 428 {
 429         return slots_to_pool(zhdr->slots);
 430 }
 431
 432 static void __release_z3fold_page(struct z3fold_header *zhdr, bool locked)
 433 {
 434         struct page *page = virt_to_page(zhdr);
 435         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 436
 437         WARN_ON(!list_empty(&zhdr->buddy));
 438         set_bit(PAGE_STALE, &page->private);
 439         clear_bit(NEEDS_COMPACTING, &page->private);
 440         spin_lock(&pool->lock);
 441         if (!list_empty(&page->lru))
 442                 list_del_init(&page->lru);
 443         spin_unlock(&pool->lock);
 444         if (locked)
 445                 z3fold_page_unlock(zhdr);
 446         spin_lock(&pool->stale_lock);
 447         list_add(&zhdr->buddy, &pool->stale);
 448         queue_work(pool->release_wq, &pool->work);
 449         spin_unlock(&pool->stale_lock);
 450 }
 451
 452 static void __attribute__((__unused__))
 453                         release_z3fold_page(struct kref *ref)
 454 {
 455         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 456                                                 refcount);
 457         __release_z3fold_page(zhdr, false);
 458 }
 459
 460 static void release_z3fold_page_locked(struct kref *ref)
 461 {
 462         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 463                                                 refcount);
 464         WARN_ON(z3fold_page_trylock(zhdr));
 465         __release_z3fold_page(zhdr, true);
 466 }
 467
 468 static void release_z3fold_page_locked_list(struct kref *ref)
 469 {
 470         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 471                                                refcount);
 472         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 473         spin_lock(&pool->lock);
 474         list_del_init(&zhdr->buddy);
 475         spin_unlock(&pool->lock);
 476
 477         WARN_ON(z3fold_page_trylock(zhdr));
 478         __release_z3fold_page(zhdr, true);
 479 }
 480
 481 static void free_pages_work(struct work_struct *w)
 482 {
 483         struct z3fold_pool *pool = container_of(w, struct z3fold_pool, work);
 484
 485         spin_lock(&pool->stale_lock);
 486         while (!list_empty(&pool->stale)) {
 487                 struct z3fold_header *zhdr = list_first_entry(&pool->stale,
 488                                                 struct z3fold_header, buddy);
 489                 struct page *page = virt_to_page(zhdr);
 490
 491                 list_del(&zhdr->buddy);
 492                 if (WARN_ON(!test_bit(PAGE_STALE, &page->private)))
 493                         continue;
 494                 spin_unlock(&pool->stale_lock);
 495                 cancel_work_sync(&zhdr->work);
 496                 free_z3fold_page(page, false);
 497                 cond_resched();
 498                 spin_lock(&pool->stale_lock);
 499         }
 500         spin_unlock(&pool->stale_lock);
 501 }
 502
 503 /*
 504  * Returns the number of free chunks in a z3fold page.
 505  * NB: can't be used with HEADLESS pages.
 506  */
 507 static int num_free_chunks(struct z3fold_header *zhdr)
 508 {
 509         int nfree;
 510         /*
 511          * If there is a middle object, pick up the bigger free space
 512          * either before or after it. Otherwise just subtract the number
 513          * of chunks occupied by the first and the last objects.
 514          */
 515         if (zhdr->middle_chunks != 0) {
 516                 int nfree_before = zhdr->first_chunks ?
 517                         0 : zhdr->start_middle - ZHDR_CHUNKS;
 518                 int nfree_after = zhdr->last_chunks ?
 519                         0 : TOTAL_CHUNKS -
 520                                 (zhdr->start_middle + zhdr->middle_chunks);
 521                 nfree = max(nfree_before, nfree_after);
 522         } else
 523                 nfree = NCHUNKS - zhdr->first_chunks - zhdr->last_chunks;
 524         return nfree;
 525 }
 526
 527 /* Add to the appropriate unbuddied list */
 528 static inline void add_to_unbuddied(struct z3fold_pool *pool,
 529                                 struct z3fold_header *zhdr)
 530 {
 531         if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0 ||
 532                         zhdr->middle_chunks == 0) {
 533                 struct list_head *unbuddied = get_cpu_ptr(pool->unbuddied);
 534
 535                 int freechunks = num_free_chunks(zhdr);
 536                 spin_lock(&pool->lock);
 537                 list_add(&zhdr->buddy, &unbuddied[freechunks]);
 538                 spin_unlock(&pool->lock);
 539                 zhdr->cpu = smp_processor_id();
 540                 put_cpu_ptr(pool->unbuddied);
 541         }
 542 }
 543
 544 static inline void *mchunk_memmove(struct z3fold_header *zhdr,
 545                                 unsigned short dst_chunk)
 546 {
 547         void *beg = zhdr;
 548         return memmove(beg + (dst_chunk << CHUNK_SHIFT),
 549                        beg + (zhdr->start_middle << CHUNK_SHIFT),
 550                        zhdr->middle_chunks << CHUNK_SHIFT);
 551 }
 552
 553 #define BIG_CHUNK_GAP   3
 554 /* Has to be called with lock held */
 555 static int z3fold_compact_page(struct z3fold_header *zhdr)
 556 {
 557         struct page *page = virt_to_page(zhdr);
 558
 559         if (test_bit(MIDDLE_CHUNK_MAPPED, &page->private))
 560                 return 0; /* can't move middle chunk, it's used */
 561
 562         if (unlikely(PageIsolated(page)))
 563                 return 0;
 564
 565         if (zhdr->middle_chunks == 0)
 566                 return 0; /* nothing to compact */
 567
 568         if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
 569                 /* move to the beginning */
 570                 mchunk_memmove(zhdr, ZHDR_CHUNKS);
 571                 zhdr->first_chunks = zhdr->middle_chunks;
 572                 zhdr->middle_chunks = 0;
 573                 zhdr->start_middle = 0;
 574                 zhdr->first_num++;
 575                 return 1;
 576         }
 577
 578         /*
 579          * moving data is expensive, so let's only do that if
 580          * there's substantial gain (at least BIG_CHUNK_GAP chunks)
 581          */
 582         if (zhdr->first_chunks != 0 && zhdr->last_chunks == 0 &&
 583             zhdr->start_middle - (zhdr->first_chunks + ZHDR_CHUNKS) >=
 584                         BIG_CHUNK_GAP) {
 585                 mchunk_memmove(zhdr, zhdr->first_chunks + ZHDR_CHUNKS);
 586                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
 587                 return 1;
 588         } else if (zhdr->last_chunks != 0 && zhdr->first_chunks == 0 &&
 589                    TOTAL_CHUNKS - (zhdr->last_chunks + zhdr->start_middle
 590                                         + zhdr->middle_chunks) >=
 591                         BIG_CHUNK_GAP) {
 592                 unsigned short new_start = TOTAL_CHUNKS - zhdr->last_chunks -
 593                         zhdr->middle_chunks;
 594                 mchunk_memmove(zhdr, new_start);
 595                 zhdr->start_middle = new_start;
 596                 return 1;
 597         }
 598
 599         return 0;
 600 }
 601
 602 static void do_compact_page(struct z3fold_header *zhdr, bool locked)
 603 {
 604         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 605         struct page *page;
 606
 607         page = virt_to_page(zhdr);
 608         if (locked)
 609                 WARN_ON(z3fold_page_trylock(zhdr));
 610         else
 611                 z3fold_page_lock(zhdr);
 612         if (WARN_ON(!test_and_clear_bit(NEEDS_COMPACTING, &page->private))) {
 613                 z3fold_page_unlock(zhdr);
 614                 return;
 615         }
 616         spin_lock(&pool->lock);
 617         list_del_init(&zhdr->buddy);
 618         spin_unlock(&pool->lock);
 619
 620         if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
 621                 atomic64_dec(&pool->pages_nr);
 622                 return;
 623         }
 624
 625         if (unlikely(PageIsolated(page) ||
 626                      test_bit(PAGE_STALE, &page->private))) {
 627                 z3fold_page_unlock(zhdr);
 628                 return;
 629         }
 630
 631         z3fold_compact_page(zhdr);
 632         add_to_unbuddied(pool, zhdr);
 633         z3fold_page_unlock(zhdr);
 634 }
 635
 636 static void compact_page_work(struct work_struct *w)
 637 {
 638         struct z3fold_header *zhdr = container_of(w, struct z3fold_header,
 639                                                 work);
 640
 641         do_compact_page(zhdr, false);
 642 }
 643
 644 /* returns _locked_ z3fold page header or NULL */
 645 static inline struct z3fold_header *__z3fold_alloc(struct z3fold_pool *pool,
 646                                                 size_t size, bool can_sleep)
 647 {
 648         struct z3fold_header *zhdr = NULL;
 649         struct page *page;
 650         struct list_head *unbuddied;
 651         int chunks = size_to_chunks(size), i;
 652
 653 lookup:
 654         /* First, try to find an unbuddied z3fold page. */
 655         unbuddied = get_cpu_ptr(pool->unbuddied);
 656         for_each_unbuddied_list(i, chunks) {
 657                 struct list_head *l = &unbuddied[i];
 658
 659                 zhdr = list_first_entry_or_null(READ_ONCE(l),
 660                                         struct z3fold_header, buddy);
 661
 662                 if (!zhdr)
 663                         continue;
 664
 665                 /* Re-check under lock. */
 666                 spin_lock(&pool->lock);
 667                 l = &unbuddied[i];
 668                 if (unlikely(zhdr != list_first_entry(READ_ONCE(l),
 669                                                 struct z3fold_header, buddy)) ||
 670                     !z3fold_page_trylock(zhdr)) {
 671                         spin_unlock(&pool->lock);
 672                         zhdr = NULL;
 673                         put_cpu_ptr(pool->unbuddied);
 674                         if (can_sleep)
 675                                 cond_resched();
 676                         goto lookup;
 677                 }
 678                 list_del_init(&zhdr->buddy);
 679                 zhdr->cpu = -1;
 680                 spin_unlock(&pool->lock);
 681
 682                 page = virt_to_page(zhdr);
 683                 if (test_bit(NEEDS_COMPACTING, &page->private)) {
 684                         z3fold_page_unlock(zhdr);
 685                         zhdr = NULL;
 686                         put_cpu_ptr(pool->unbuddied);
 687                         if (can_sleep)
 688                                 cond_resched();
 689                         goto lookup;
 690                 }
 691
 692                 /*
 693                  * this page could not be removed from its unbuddied
 694                  * list while pool lock was held, and then we've taken
 695                  * page lock so kref_put could not be called before
 696                  * we got here, so it's safe to just call kref_get()
 697                  */
 698                 kref_get(&zhdr->refcount);
 699                 break;
 700         }
 701         put_cpu_ptr(pool->unbuddied);
 702
 703         if (!zhdr) {
 704                 int cpu;
 705
 706                 /* look for _exact_ match on other cpus' lists */
 707                 for_each_online_cpu(cpu) {
 708                         struct list_head *l;
 709
 710                         unbuddied = per_cpu_ptr(pool->unbuddied, cpu);
 711                         spin_lock(&pool->lock);
 712                         l = &unbuddied[chunks];
 713
 714                         zhdr = list_first_entry_or_null(READ_ONCE(l),
 715                                                 struct z3fold_header, buddy);
 716
 717                         if (!zhdr || !z3fold_page_trylock(zhdr)) {
 718                                 spin_unlock(&pool->lock);
 719                                 zhdr = NULL;
 720                                 continue;
 721                         }
 722                         list_del_init(&zhdr->buddy);
 723                         zhdr->cpu = -1;
 724                         spin_unlock(&pool->lock);
 725
 726                         page = virt_to_page(zhdr);
 727                         if (test_bit(NEEDS_COMPACTING, &page->private)) {
 728                                 z3fold_page_unlock(zhdr);
 729                                 zhdr = NULL;
 730                                 if (can_sleep)
 731                                         cond_resched();
 732                                 continue;
 733                         }
 734                         kref_get(&zhdr->refcount);
 735                         break;
 736                 }
 737         }
 738
 739         return zhdr;
 740 }
 741
 742 /*
 743  * API Functions
 744  */
 745
 746 /**
 747  * z3fold_create_pool() - create a new z3fold pool
 748  * @name:       pool name
 749  * @gfp:        gfp flags when allocating the z3fold pool structure
 750  * @ops:        user-defined operations for the z3fold pool
 751  *
 752  * Return: pointer to the new z3fold pool or NULL if the metadata allocation
 753  * failed.
 754  */
 755 static struct z3fold_pool *z3fold_create_pool(const char *name, gfp_t gfp,
 756                 const struct z3fold_ops *ops)
 757 {
 758         struct z3fold_pool *pool = NULL;
 759         int i, cpu;
 760
 761         pool = kzalloc(sizeof(struct z3fold_pool), gfp);
 762         if (!pool)
 763                 goto out;
 764         pool->c_handle = kmem_cache_create("z3fold_handle",
 765                                 sizeof(struct z3fold_buddy_slots),
 766                                 SLOTS_ALIGN, 0, NULL);
 767         if (!pool->c_handle)
 768                 goto out_c;
 769         spin_lock_init(&pool->lock);
 770         spin_lock_init(&pool->stale_lock);
 771         pool->unbuddied = __alloc_percpu(sizeof(struct list_head)*NCHUNKS, 2);
 772         if (!pool->unbuddied)
 773                 goto out_pool;
 774         for_each_possible_cpu(cpu) {
 775                 struct list_head *unbuddied =
 776                                 per_cpu_ptr(pool->unbuddied, cpu);
 777                 for_each_unbuddied_list(i, 0)
 778                         INIT_LIST_HEAD(&unbuddied[i]);
 779         }
 780         INIT_LIST_HEAD(&pool->lru);
 781         INIT_LIST_HEAD(&pool->stale);
 782         atomic64_set(&pool->pages_nr, 0);
 783         pool->name = name;
 784         pool->compact_wq = create_singlethread_workqueue(pool->name);
 785         if (!pool->compact_wq)
 786                 goto out_unbuddied;
 787         pool->release_wq = create_singlethread_workqueue(pool->name);
 788         if (!pool->release_wq)
 789                 goto out_wq;
 790         if (z3fold_register_migration(pool))
 791                 goto out_rwq;
 792         INIT_WORK(&pool->work, free_pages_work);
 793         pool->ops = ops;
 794         return pool;
 795
 796 out_rwq:
 797         destroy_workqueue(pool->release_wq);
 798 out_wq:
 799         destroy_workqueue(pool->compact_wq);
 800 out_unbuddied:
 801         free_percpu(pool->unbuddied);
 802 out_pool:
 803         kmem_cache_destroy(pool->c_handle);
 804 out_c:
 805         kfree(pool);
 806 out:
 807         return NULL;
 808 }
 809
 810 /**
 811  * z3fold_destroy_pool() - destroys an existing z3fold pool
 812  * @pool:       the z3fold pool to be destroyed
 813  *
 814  * The pool should be emptied before this function is called.
 815  */
 816 static void z3fold_destroy_pool(struct z3fold_pool *pool)
 817 {
 818         kmem_cache_destroy(pool->c_handle);
 819         z3fold_unregister_migration(pool);
 820         destroy_workqueue(pool->release_wq);
 821         destroy_workqueue(pool->compact_wq);
 822         kfree(pool);
 823 }
 824
 825 /**
 826  * z3fold_alloc() - allocates a region of a given size
 827  * @pool:       z3fold pool from which to allocate
 828  * @size:       size in bytes of the desired allocation
 829  * @gfp:        gfp flags used if the pool needs to grow
 830  * @handle:     handle of the new allocation
 831  *
 832  * This function will attempt to find a free region in the pool large enough to
 833  * satisfy the allocation request.  A search of the unbuddied lists is
 834  * performed first. If no suitable free region is found, then a new page is
 835  * allocated and added to the pool to satisfy the request.
 836  *
 837  * gfp should not set __GFP_HIGHMEM as highmem pages cannot be used
 838  * as z3fold pool pages.
 839  *
 840  * Return: 0 if success and handle is set, otherwise -EINVAL if the size or
 841  * gfp arguments are invalid or -ENOMEM if the pool was unable to allocate
 842  * a new page.
 843  */
 844 static int z3fold_alloc(struct z3fold_pool *pool, size_t size, gfp_t gfp,
 845                         unsigned long *handle)
 846 {
 847         int chunks = size_to_chunks(size);
 848         struct z3fold_header *zhdr = NULL;
 849         struct page *page = NULL;
 850         enum buddy bud;
 851         bool can_sleep = gfpflags_allow_blocking(gfp);
 852
 853         if (!size || (gfp & __GFP_HIGHMEM))
 854                 return -EINVAL;
 855
 856         if (size > PAGE_SIZE)
 857                 return -ENOSPC;
 858
 859         if (size > PAGE_SIZE - ZHDR_SIZE_ALIGNED - CHUNK_SIZE)
 860                 bud = HEADLESS;
 861         else {
 862 retry:
 863                 zhdr = __z3fold_alloc(pool, size, can_sleep);
 864                 if (zhdr) {
 865                         if (zhdr->first_chunks == 0) {
 866                                 if (zhdr->middle_chunks != 0 &&
 867                                     chunks >= zhdr->start_middle)
 868                                         bud = LAST;
 869                                 else
 870                                         bud = FIRST;
 871                         } else if (zhdr->last_chunks == 0)
 872                                 bud = LAST;
 873                         else if (zhdr->middle_chunks == 0)
 874                                 bud = MIDDLE;
 875                         else {
 876                                 if (kref_put(&zhdr->refcount,
 877                                              release_z3fold_page_locked))
 878                                         atomic64_dec(&pool->pages_nr);
 879                                 else
 880                                         z3fold_page_unlock(zhdr);
 881                                 pr_err("No free chunks in unbuddied\n");
 882                                 WARN_ON(1);
 883                                 goto retry;
 884                         }
 885                         page = virt_to_page(zhdr);
 886                         goto found;
 887                 }
 888                 bud = FIRST;
 889         }
 890
 891         page = NULL;
 892         if (can_sleep) {
 893                 spin_lock(&pool->stale_lock);
 894                 zhdr = list_first_entry_or_null(&pool->stale,
 895                                                 struct z3fold_header, buddy);
 896                 /*
 897                  * Before allocating a page, let's see if we can take one from
 898                  * the stale pages list. cancel_work_sync() can sleep so we
 899                  * limit this case to the contexts where we can sleep
 900                  */
 901                 if (zhdr) {
 902                         list_del(&zhdr->buddy);
 903                         spin_unlock(&pool->stale_lock);
 904                         cancel_work_sync(&zhdr->work);
 905                         page = virt_to_page(zhdr);
 906                 } else {
 907                         spin_unlock(&pool->stale_lock);
 908                 }
 909         }
 910         if (!page)
 911                 page = alloc_page(gfp);
 912
 913         if (!page)
 914                 return -ENOMEM;
 915
 916         zhdr = init_z3fold_page(page, pool, gfp);
 917         if (!zhdr) {
 918                 __free_page(page);
 919                 return -ENOMEM;
 920         }
 921         atomic64_inc(&pool->pages_nr);
 922
 923         if (bud == HEADLESS) {
 924                 set_bit(PAGE_HEADLESS, &page->private);
 925                 goto headless;
 926         }
 927         __SetPageMovable(page, pool->inode->i_mapping);
 928         z3fold_page_lock(zhdr);
 929
 930 found:
 931         if (bud == FIRST)
 932                 zhdr->first_chunks = chunks;
 933         else if (bud == LAST)
 934                 zhdr->last_chunks = chunks;
 935         else {
 936                 zhdr->middle_chunks = chunks;
 937                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
 938         }
 939         add_to_unbuddied(pool, zhdr);
 940
 941 headless:
 942         spin_lock(&pool->lock);
 943         /* Add/move z3fold page to beginning of LRU */
 944         if (!list_empty(&page->lru))
 945                 list_del(&page->lru);
 946
 947         list_add(&page->lru, &pool->lru);
 948
 949         *handle = encode_handle(zhdr, bud);
 950         spin_unlock(&pool->lock);
 951         if (bud != HEADLESS)
 952                 z3fold_page_unlock(zhdr);
 953
 954         return 0;
 955 }
 956
 957 /**
 958  * z3fold_free() - frees the allocation associated with the given handle
 959  * @pool:       pool in which the allocation resided
 960  * @handle:     handle associated with the allocation returned by z3fold_alloc()
 961  *
 962  * In the case that the z3fold page in which the allocation resides is under
 963  * reclaim, as indicated by the PG_reclaim flag being set, this function
 964  * only sets the first|last_chunks to 0.  The page is actually freed
 965  * once both buddies are evicted (see z3fold_reclaim_page() below).
 966  */
 967 static void z3fold_free(struct z3fold_pool *pool, unsigned long handle)
 968 {
 969         struct z3fold_header *zhdr;
 970         struct page *page;
 971         enum buddy bud;
 972
 973         zhdr = handle_to_z3fold_header(handle);
 974         page = virt_to_page(zhdr);
 975
 976         if (test_bit(PAGE_HEADLESS, &page->private)) {
 977                 /* if a headless page is under reclaim, just leave.
 978                  * NB: we use test_and_set_bit for a reason: if the bit
 979                  * has not been set before, we release this page
 980                  * immediately so we don't care about its value any more.
 981                  */
 982                 if (!test_and_set_bit(PAGE_CLAIMED, &page->private)) {
 983                         spin_lock(&pool->lock);
 984                         list_del(&page->lru);
 985                         spin_unlock(&pool->lock);
 986                         free_z3fold_page(page, true);
 987                         atomic64_dec(&pool->pages_nr);
 988                 }
 989                 return;
 990         }
 991
 992         /* Non-headless case */
 993         z3fold_page_lock(zhdr);
 994         bud = handle_to_buddy(handle);
 995
 996         switch (bud) {
 997         case FIRST:
 998                 zhdr->first_chunks = 0;
 999                 break;
1000         case MIDDLE:
1001                 zhdr->middle_chunks = 0;
1002                 break;
1003         case LAST:
1004                 zhdr->last_chunks = 0;
1005                 break;
1006         default:
1007                 pr_err("%s: unknown bud %d\n", __func__, bud);
1008                 WARN_ON(1);
1009                 z3fold_page_unlock(zhdr);
1010                 return;
1011         }
1012
1013         free_handle(handle);
1014         if (kref_put(&zhdr->refcount, release_z3fold_page_locked_list)) {
1015                 atomic64_dec(&pool->pages_nr);
1016                 return;
1017         }
1018         if (test_bit(PAGE_CLAIMED, &page->private)) {
1019                 z3fold_page_unlock(zhdr);
1020                 return;
1021         }
1022         if (unlikely(PageIsolated(page)) ||
1023             test_and_set_bit(NEEDS_COMPACTING, &page->private)) {
1024                 z3fold_page_unlock(zhdr);
1025                 return;
1026         }
1027         if (zhdr->cpu < 0 || !cpu_online(zhdr->cpu)) {
1028                 spin_lock(&pool->lock);
1029                 list_del_init(&zhdr->buddy);
1030                 spin_unlock(&pool->lock);
1031                 zhdr->cpu = -1;
1032                 kref_get(&zhdr->refcount);
1033                 do_compact_page(zhdr, true);
1034                 return;
1035         }
1036         kref_get(&zhdr->refcount);
1037         queue_work_on(zhdr->cpu, pool->compact_wq, &zhdr->work);
1038         z3fold_page_unlock(zhdr);
1039 }
1040
1041 /**
1042  * z3fold_reclaim_page() - evicts allocations from a pool page and frees it
1043  * @pool:       pool from which a page will attempt to be evicted
1044  * @retries:    number of pages on the LRU list for which eviction will
1045  *              be attempted before failing
1046  *
1047  * z3fold reclaim is different from normal system reclaim in that it is done
1048  * from the bottom, up. This is because only the bottom layer, z3fold, has
1049  * information on how the allocations are organized within each z3fold page.
1050  * This has the potential to create interesting locking situations between
1051  * z3fold and the user, however.
1052  *
1053  * To avoid these, this is how z3fold_reclaim_page() should be called:
1054  *
1055  * The user detects a page should be reclaimed and calls z3fold_reclaim_page().
1056  * z3fold_reclaim_page() will remove a z3fold page from the pool LRU list and
1057  * call the user-defined eviction handler with the pool and handle as
1058  * arguments.
1059  *
1060  * If the handle can not be evicted, the eviction handler should return
1061  * non-zero. z3fold_reclaim_page() will add the z3fold page back to the
1062  * appropriate list and try the next z3fold page on the LRU up to
1063  * a user defined number of retries.
1064  *
1065  * If the handle is successfully evicted, the eviction handler should
1066  * return 0 _and_ should have called z3fold_free() on the handle. z3fold_free()
1067  * contains logic to delay freeing the page if the page is under reclaim,
1068  * as indicated by the setting of the PG_reclaim flag on the underlying page.
1069  *
1070  * If all buddies in the z3fold page are successfully evicted, then the
1071  * z3fold page can be freed.
1072  *
1073  * Returns: 0 if page is successfully freed, otherwise -EINVAL if there are
1074  * no pages to evict or an eviction handler is not registered, -EAGAIN if
1075  * the retry limit was hit.
1076  */
1077 static int z3fold_reclaim_page(struct z3fold_pool *pool, unsigned int retries)
1078 {
1079         int i, ret = 0;
1080         struct z3fold_header *zhdr = NULL;
1081         struct page *page = NULL;
1082         struct list_head *pos;
1083         unsigned long first_handle = 0, middle_handle = 0, last_handle = 0;
1084
1085         spin_lock(&pool->lock);
1086         if (!pool->ops || !pool->ops->evict || retries == 0) {
1087                 spin_unlock(&pool->lock);
1088                 return -EINVAL;
1089         }
1090         for (i = 0; i < retries; i++) {
1091                 if (list_empty(&pool->lru)) {
1092                         spin_unlock(&pool->lock);
1093                         return -EINVAL;
1094                 }
1095                 list_for_each_prev(pos, &pool->lru) {
1096                         page = list_entry(pos, struct page, lru);
1097
1098                         /* this bit could have been set by free, in which case
1099                          * we pass over to the next page in the pool.
1100                          */
1101                         if (test_and_set_bit(PAGE_CLAIMED, &page->private))
1102                                 continue;
1103
1104                         if (unlikely(PageIsolated(page)))
1105                                 continue;
1106                         if (test_bit(PAGE_HEADLESS, &page->private))
1107                                 break;
1108
1109                         zhdr = page_address(page);
1110                         if (!z3fold_page_trylock(zhdr)) {
1111                                 zhdr = NULL;
1112                                 continue; /* can't evict at this point */
1113                         }
1114                         kref_get(&zhdr->refcount);
1115                         list_del_init(&zhdr->buddy);
1116                         zhdr->cpu = -1;
1117                         break;
1118                 }
1119
1120                 if (!zhdr)
1121                         break;
1122
1123                 list_del_init(&page->lru);
1124                 spin_unlock(&pool->lock);
1125
1126                 if (!test_bit(PAGE_HEADLESS, &page->private)) {
1127                         /*
1128                          * We need encode the handles before unlocking, since
1129                          * we can race with free that will set
1130                          * (first|last)_chunks to 0
1131                          */
1132                         first_handle = 0;
1133                         last_handle = 0;
1134                         middle_handle = 0;
1135                         if (zhdr->first_chunks)
1136                                 first_handle = encode_handle(zhdr, FIRST);
1137                         if (zhdr->middle_chunks)
1138                                 middle_handle = encode_handle(zhdr, MIDDLE);
1139                         if (zhdr->last_chunks)
1140                                 last_handle = encode_handle(zhdr, LAST);
1141                         /*
1142                          * it's safe to unlock here because we hold a
1143                          * reference to this page
1144                          */
1145                         z3fold_page_unlock(zhdr);
1146                 } else {
1147                         first_handle = encode_handle(zhdr, HEADLESS);
1148                         last_handle = middle_handle = 0;
1149                 }
1150
1151                 /* Issue the eviction callback(s) */
1152                 if (middle_handle) {
1153                         ret = pool->ops->evict(pool, middle_handle);
1154                         if (ret)
1155                                 goto next;
1156                 }
1157                 if (first_handle) {
1158                         ret = pool->ops->evict(pool, first_handle);
1159                         if (ret)
1160                                 goto next;
1161                 }
1162                 if (last_handle) {
1163                         ret = pool->ops->evict(pool, last_handle);
1164                         if (ret)
1165                                 goto next;
1166                 }
1167 next:
1168                 if (test_bit(PAGE_HEADLESS, &page->private)) {
1169                         if (ret == 0) {
1170                                 free_z3fold_page(page, true);
1171                                 atomic64_dec(&pool->pages_nr);
1172                                 return 0;
1173                         }
1174                         spin_lock(&pool->lock);
1175                         list_add(&page->lru, &pool->lru);
1176                         spin_unlock(&pool->lock);
1177                 } else {
1178                         z3fold_page_lock(zhdr);
1179                         clear_bit(PAGE_CLAIMED, &page->private);
1180                         if (kref_put(&zhdr->refcount,
1181                                         release_z3fold_page_locked)) {
1182                                 atomic64_dec(&pool->pages_nr);
1183                                 return 0;
1184                         }
1185                         /*
1186                          * if we are here, the page is still not completely
1187                          * free. Take the global pool lock then to be able
1188                          * to add it back to the lru list
1189                          */
1190                         spin_lock(&pool->lock);
1191                         list_add(&page->lru, &pool->lru);
1192                         spin_unlock(&pool->lock);
1193                         z3fold_page_unlock(zhdr);
1194                 }
1195
1196                 /* We started off locked to we need to lock the pool back */
1197                 spin_lock(&pool->lock);
1198         }
1199         spin_unlock(&pool->lock);
1200         return -EAGAIN;
1201 }
1202
1203 /**
1204  * z3fold_map() - maps the allocation associated with the given handle
1205  * @pool:       pool in which the allocation resides
1206  * @handle:     handle associated with the allocation to be mapped
1207  *
1208  * Extracts the buddy number from handle and constructs the pointer to the
1209  * correct starting chunk within the page.
1210  *
1211  * Returns: a pointer to the mapped allocation
1212  */
1213 static void *z3fold_map(struct z3fold_pool *pool, unsigned long handle)
1214 {
1215         struct z3fold_header *zhdr;
1216         struct page *page;
1217         void *addr;
1218         enum buddy buddy;
1219
1220         zhdr = handle_to_z3fold_header(handle);
1221         addr = zhdr;
1222         page = virt_to_page(zhdr);
1223
1224         if (test_bit(PAGE_HEADLESS, &page->private))
1225                 goto out;
1226
1227         z3fold_page_lock(zhdr);
1228         buddy = handle_to_buddy(handle);
1229         switch (buddy) {
1230         case FIRST:
1231                 addr += ZHDR_SIZE_ALIGNED;
1232                 break;
1233         case MIDDLE:
1234                 addr += zhdr->start_middle << CHUNK_SHIFT;
1235                 set_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1236                 break;
1237         case LAST:
1238                 addr += PAGE_SIZE - (handle_to_chunks(handle) << CHUNK_SHIFT);
1239                 break;
1240         default:
1241                 pr_err("unknown buddy id %d\n", buddy);
1242                 WARN_ON(1);
1243                 addr = NULL;
1244                 break;
1245         }
1246
1247         if (addr)
1248                 zhdr->mapped_count++;
1249         z3fold_page_unlock(zhdr);
1250 out:
1251         return addr;
1252 }
1253
1254 /**
1255  * z3fold_unmap() - unmaps the allocation associated with the given handle
1256  * @pool:       pool in which the allocation resides
1257  * @handle:     handle associated with the allocation to be unmapped
1258  */
1259 static void z3fold_unmap(struct z3fold_pool *pool, unsigned long handle)
1260 {
1261         struct z3fold_header *zhdr;
1262         struct page *page;
1263         enum buddy buddy;
1264
1265         zhdr = handle_to_z3fold_header(handle);
1266         page = virt_to_page(zhdr);
1267
1268         if (test_bit(PAGE_HEADLESS, &page->private))
1269                 return;
1270
1271         z3fold_page_lock(zhdr);
1272         buddy = handle_to_buddy(handle);
1273         if (buddy == MIDDLE)
1274                 clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1275         zhdr->mapped_count--;
1276         z3fold_page_unlock(zhdr);
1277 }
1278
1279 /**
1280  * z3fold_get_pool_size() - gets the z3fold pool size in pages
1281  * @pool:       pool whose size is being queried
1282  *
1283  * Returns: size in pages of the given pool.
1284  */
1285 static u64 z3fold_get_pool_size(struct z3fold_pool *pool)
1286 {
1287         return atomic64_read(&pool->pages_nr);
1288 }
1289
1290 static bool z3fold_page_isolate(struct page *page, isolate_mode_t mode)
1291 {
1292         struct z3fold_header *zhdr;
1293         struct z3fold_pool *pool;
1294
1295         VM_BUG_ON_PAGE(!PageMovable(page), page);
1296         VM_BUG_ON_PAGE(PageIsolated(page), page);
1297
1298         if (test_bit(PAGE_HEADLESS, &page->private))
1299                 return false;
1300
1301         zhdr = page_address(page);
1302         z3fold_page_lock(zhdr);
1303         if (test_bit(NEEDS_COMPACTING, &page->private) ||
1304             test_bit(PAGE_STALE, &page->private))
1305                 goto out;
1306
1307         pool = zhdr_to_pool(zhdr);
1308
1309         if (zhdr->mapped_count == 0) {
1310                 kref_get(&zhdr->refcount);
1311                 if (!list_empty(&zhdr->buddy))
1312                         list_del_init(&zhdr->buddy);
1313                 spin_lock(&pool->lock);
1314                 if (!list_empty(&page->lru))
1315                         list_del(&page->lru);
1316                 spin_unlock(&pool->lock);
1317                 z3fold_page_unlock(zhdr);
1318                 return true;
1319         }
1320 out:
1321         z3fold_page_unlock(zhdr);
1322         return false;
1323 }
1324
1325 static int z3fold_page_migrate(struct address_space *mapping, struct page *newpage,
1326                                struct page *page, enum migrate_mode mode)
1327 {
1328         struct z3fold_header *zhdr, *new_zhdr;
1329         struct z3fold_pool *pool;
1330         struct address_space *new_mapping;
1331
1332         VM_BUG_ON_PAGE(!PageMovable(page), page);
1333         VM_BUG_ON_PAGE(!PageIsolated(page), page);
1334
1335         zhdr = page_address(page);
1336         pool = zhdr_to_pool(zhdr);
1337
1338         if (!trylock_page(page))
1339                 return -EAGAIN;
1340
1341         if (!z3fold_page_trylock(zhdr)) {
1342                 unlock_page(page);
1343                 return -EAGAIN;
1344         }
1345         if (zhdr->mapped_count != 0) {
1346                 z3fold_page_unlock(zhdr);
1347                 unlock_page(page);
1348                 return -EBUSY;
1349         }
1350         new_zhdr = page_address(newpage);
1351         memcpy(new_zhdr, zhdr, PAGE_SIZE);
1352         newpage->private = page->private;
1353         page->private = 0;
1354         z3fold_page_unlock(zhdr);
1355         spin_lock_init(&new_zhdr->page_lock);
1356         new_mapping = page_mapping(page);
1357         __ClearPageMovable(page);
1358         ClearPagePrivate(page);
1359
1360         get_page(newpage);
1361         z3fold_page_lock(new_zhdr);
1362         if (new_zhdr->first_chunks)
1363                 encode_handle(new_zhdr, FIRST);
1364         if (new_zhdr->last_chunks)
1365                 encode_handle(new_zhdr, LAST);
1366         if (new_zhdr->middle_chunks)
1367                 encode_handle(new_zhdr, MIDDLE);
1368         set_bit(NEEDS_COMPACTING, &newpage->private);
1369         new_zhdr->cpu = smp_processor_id();
1370         spin_lock(&pool->lock);
1371         list_add(&newpage->lru, &pool->lru);
1372         spin_unlock(&pool->lock);
1373         __SetPageMovable(newpage, new_mapping);
1374         z3fold_page_unlock(new_zhdr);
1375
1376         queue_work_on(new_zhdr->cpu, pool->compact_wq, &new_zhdr->work);
1377
1378         page_mapcount_reset(page);
1379         unlock_page(page);
1380         put_page(page);
1381         return 0;
1382 }
1383
1384 static void z3fold_page_putback(struct page *page)
1385 {
1386         struct z3fold_header *zhdr;
1387         struct z3fold_pool *pool;
1388
1389         zhdr = page_address(page);
1390         pool = zhdr_to_pool(zhdr);
1391
1392         z3fold_page_lock(zhdr);
1393         if (!list_empty(&zhdr->buddy))
1394                 list_del_init(&zhdr->buddy);
1395         INIT_LIST_HEAD(&page->lru);
1396         if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
1397                 atomic64_dec(&pool->pages_nr);
1398                 return;
1399         }
1400         spin_lock(&pool->lock);
1401         list_add(&page->lru, &pool->lru);
1402         spin_unlock(&pool->lock);
1403         z3fold_page_unlock(zhdr);
1404 }
1405
1406 static const struct address_space_operations z3fold_aops = {
1407         .isolate_page = z3fold_page_isolate,
1408         .migratepage = z3fold_page_migrate,
1409         .putback_page = z3fold_page_putback,
1410 };
1411
1412 /*****************
1413  * zpool
1414  ****************/
1415
1416 static int z3fold_zpool_evict(struct z3fold_pool *pool, unsigned long handle)
1417 {
1418         if (pool->zpool && pool->zpool_ops && pool->zpool_ops->evict)
1419                 return pool->zpool_ops->evict(pool->zpool, handle);
1420         else
1421                 return -ENOENT;
1422 }
1423
1424 static const struct z3fold_ops z3fold_zpool_ops = {
1425         .evict =        z3fold_zpool_evict
1426 };
1427
1428 static void *z3fold_zpool_create(const char *name, gfp_t gfp,
1429                                const struct zpool_ops *zpool_ops,
1430                                struct zpool *zpool)
1431 {
1432         struct z3fold_pool *pool;
1433
1434         pool = z3fold_create_pool(name, gfp,
1435                                 zpool_ops ? &z3fold_zpool_ops : NULL);
1436         if (pool) {
1437                 pool->zpool = zpool;
1438                 pool->zpool_ops = zpool_ops;
1439         }
1440         return pool;
1441 }
1442
1443 static void z3fold_zpool_destroy(void *pool)
1444 {
1445         z3fold_destroy_pool(pool);
1446 }
1447
1448 static int z3fold_zpool_malloc(void *pool, size_t size, gfp_t gfp,
1449                         unsigned long *handle)
1450 {
1451         return z3fold_alloc(pool, size, gfp, handle);
1452 }
1453 static void z3fold_zpool_free(void *pool, unsigned long handle)
1454 {
1455         z3fold_free(pool, handle);
1456 }
1457
1458 static int z3fold_zpool_shrink(void *pool, unsigned int pages,
1459                         unsigned int *reclaimed)
1460 {
1461         unsigned int total = 0;
1462         int ret = -EINVAL;
1463
1464         while (total < pages) {
1465                 ret = z3fold_reclaim_page(pool, 8);
1466                 if (ret < 0)
1467                         break;
1468                 total++;
1469         }
1470
1471         if (reclaimed)
1472                 *reclaimed = total;
1473
1474         return ret;
1475 }
1476
1477 static void *z3fold_zpool_map(void *pool, unsigned long handle,
1478                         enum zpool_mapmode mm)
1479 {
1480         return z3fold_map(pool, handle);
1481 }
1482 static void z3fold_zpool_unmap(void *pool, unsigned long handle)
1483 {
1484         z3fold_unmap(pool, handle);
1485 }
1486
1487 static u64 z3fold_zpool_total_size(void *pool)
1488 {
1489         return z3fold_get_pool_size(pool) * PAGE_SIZE;
1490 }
1491
1492 static struct zpool_driver z3fold_zpool_driver = {
1493         .type =         "z3fold",
1494         .owner =        THIS_MODULE,
1495         .create =       z3fold_zpool_create,
1496         .destroy =      z3fold_zpool_destroy,
1497         .malloc =       z3fold_zpool_malloc,
1498         .free =         z3fold_zpool_free,
1499         .shrink =       z3fold_zpool_shrink,
1500         .map =          z3fold_zpool_map,
1501         .unmap =        z3fold_zpool_unmap,
1502         .total_size =   z3fold_zpool_total_size,
1503 };
1504
1505 MODULE_ALIAS("zpool-z3fold");
1506
1507 static int __init init_z3fold(void)
1508 {
1509         int ret;
1510
1511         /* Make sure the z3fold header is not larger than the page size */
1512         BUILD_BUG_ON(ZHDR_SIZE_ALIGNED > PAGE_SIZE);
1513         ret = z3fold_mount();
1514         if (ret)
1515                 return ret;
1516
1517         zpool_register_driver(&z3fold_zpool_driver);
1518
1519         return 0;
1520 }
1521
1522 static void __exit exit_z3fold(void)
1523 {
1524         z3fold_unmount();
1525         zpool_unregister_driver(&z3fold_zpool_driver);
1526 }
1527
1528 module_init(init_z3fold);
1529 module_exit(exit_z3fold);
1530
1531 MODULE_LICENSE("GPL");
1532 MODULE_AUTHOR("Vitaly Wool <vitalywool@gmail.com>");
1533 MODULE_DESCRIPTION("3-Fold Allocator for Compressed Pages");