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/list.h>
  30 #include <linux/mm.h>
  31 #include <linux/module.h>
  32 #include <linux/page-flags.h>
  33 #include <linux/migrate.h>
  34 #include <linux/node.h>
  35 #include <linux/compaction.h>
  36 #include <linux/percpu.h>
  37 #include <linux/preempt.h>
  38 #include <linux/workqueue.h>
  39 #include <linux/slab.h>
  40 #include <linux/spinlock.h>
  41 #include <linux/zpool.h>
  42 #include <linux/kmemleak.h>
  43
  44 /*
  45  * NCHUNKS_ORDER determines the internal allocation granularity, effectively
  46  * adjusting internal fragmentation.  It also determines the number of
  47  * freelists maintained in each pool. NCHUNKS_ORDER of 6 means that the
  48  * allocation granularity will be in chunks of size PAGE_SIZE/64. Some chunks
  49  * in the beginning of an allocated page are occupied by z3fold header, so
  50  * NCHUNKS will be calculated to 63 (or 62 in case CONFIG_DEBUG_SPINLOCK=y),
  51  * which shows the max number of free chunks in z3fold page, also there will
  52  * be 63, or 62, respectively, freelists per pool.
  53  */
  54 #define NCHUNKS_ORDER   6
  55
  56 #define CHUNK_SHIFT     (PAGE_SHIFT - NCHUNKS_ORDER)
  57 #define CHUNK_SIZE      (1 << CHUNK_SHIFT)
  58 #define ZHDR_SIZE_ALIGNED round_up(sizeof(struct z3fold_header), CHUNK_SIZE)
  59 #define ZHDR_CHUNKS     (ZHDR_SIZE_ALIGNED >> CHUNK_SHIFT)
  60 #define TOTAL_CHUNKS    (PAGE_SIZE >> CHUNK_SHIFT)
  61 #define NCHUNKS         (TOTAL_CHUNKS - ZHDR_CHUNKS)
  62
  63 #define BUDDY_MASK      (0x3)
  64 #define BUDDY_SHIFT     2
  65 #define SLOTS_ALIGN     (0x40)
  66
  67 /*****************
  68  * Structures
  69 *****************/
  70 struct z3fold_pool;
  71
  72 enum buddy {
  73         HEADLESS = 0,
  74         FIRST,
  75         MIDDLE,
  76         LAST,
  77         BUDDIES_MAX = LAST
  78 };
  79
  80 struct z3fold_buddy_slots {
  81         /*
  82          * we are using BUDDY_MASK in handle_to_buddy etc. so there should
  83          * be enough slots to hold all possible variants
  84          */
  85         unsigned long slot[BUDDY_MASK + 1];
  86         unsigned long pool; /* back link */
  87         rwlock_t lock;
  88 };
  89 #define HANDLE_FLAG_MASK        (0x03)
  90
  91 /*
  92  * struct z3fold_header - z3fold page metadata occupying first chunks of each
  93  *                      z3fold page, except for HEADLESS pages
  94  * @buddy:              links the z3fold page into the relevant list in the
  95  *                      pool
  96  * @page_lock:          per-page lock
  97  * @refcount:           reference count for the z3fold page
  98  * @work:               work_struct for page layout optimization
  99  * @slots:              pointer to the structure holding buddy slots
 100  * @pool:               pointer to the containing pool
 101  * @cpu:                CPU which this page "belongs" to
 102  * @first_chunks:       the size of the first buddy in chunks, 0 if free
 103  * @middle_chunks:      the size of the middle buddy in chunks, 0 if free
 104  * @last_chunks:        the size of the last buddy in chunks, 0 if free
 105  * @first_num:          the starting number (for the first handle)
 106  * @mapped_count:       the number of objects currently mapped
 107  */
 108 struct z3fold_header {
 109         struct list_head buddy;
 110         spinlock_t page_lock;
 111         struct kref refcount;
 112         struct work_struct work;
 113         struct z3fold_buddy_slots *slots;
 114         struct z3fold_pool *pool;
 115         short cpu;
 116         unsigned short first_chunks;
 117         unsigned short middle_chunks;
 118         unsigned short last_chunks;
 119         unsigned short start_middle;
 120         unsigned short first_num:2;
 121         unsigned short mapped_count:2;
 122         unsigned short foreign_handles:2;
 123 };
 124
 125 /**
 126  * struct z3fold_pool - stores metadata for each z3fold pool
 127  * @name:       pool name
 128  * @lock:       protects pool unbuddied lists
 129  * @stale_lock: protects pool stale page list
 130  * @unbuddied:  per-cpu array of lists tracking z3fold pages that contain 2-
 131  *              buddies; the list each z3fold page is added to depends on
 132  *              the size of its free region.
 133  * @stale:      list of pages marked for freeing
 134  * @pages_nr:   number of z3fold pages in the pool.
 135  * @c_handle:   cache for z3fold_buddy_slots allocation
 136  * @compact_wq: workqueue for page layout background optimization
 137  * @release_wq: workqueue for safe page release
 138  * @work:       work_struct for safe page release
 139  *
 140  * This structure is allocated at pool creation time and maintains metadata
 141  * pertaining to a particular z3fold pool.
 142  */
 143 struct z3fold_pool {
 144         const char *name;
 145         spinlock_t lock;
 146         spinlock_t stale_lock;
 147         struct list_head *unbuddied;
 148         struct list_head stale;
 149         atomic64_t pages_nr;
 150         struct kmem_cache *c_handle;
 151         struct workqueue_struct *compact_wq;
 152         struct workqueue_struct *release_wq;
 153         struct work_struct work;
 154 };
 155
 156 /*
 157  * Internal z3fold page flags
 158  */
 159 enum z3fold_page_flags {
 160         PAGE_HEADLESS = 0,
 161         MIDDLE_CHUNK_MAPPED,
 162         NEEDS_COMPACTING,
 163         PAGE_STALE,
 164         PAGE_CLAIMED, /* by either reclaim or free */
 165         PAGE_MIGRATED, /* page is migrated and soon to be released */
 166 };
 167
 168 /*
 169  * handle flags, go under HANDLE_FLAG_MASK
 170  */
 171 enum z3fold_handle_flags {
 172         HANDLES_NOFREE = 0,
 173 };
 174
 175 /*
 176  * Forward declarations
 177  */
 178 static struct z3fold_header *__z3fold_alloc(struct z3fold_pool *, size_t, bool);
 179 static void compact_page_work(struct work_struct *w);
 180
 181 /*****************
 182  * Helpers
 183 *****************/
 184
 185 /* Converts an allocation size in bytes to size in z3fold chunks */
 186 static int size_to_chunks(size_t size)
 187 {
 188         return (size + CHUNK_SIZE - 1) >> CHUNK_SHIFT;
 189 }
 190
 191 #define for_each_unbuddied_list(_iter, _begin) \
 192         for ((_iter) = (_begin); (_iter) < NCHUNKS; (_iter)++)
 193
 194 static inline struct z3fold_buddy_slots *alloc_slots(struct z3fold_pool *pool,
 195                                                         gfp_t gfp)
 196 {
 197         struct z3fold_buddy_slots *slots = kmem_cache_zalloc(pool->c_handle,
 198                                                              gfp);
 199
 200         if (slots) {
 201                 /* It will be freed separately in free_handle(). */
 202                 kmemleak_not_leak(slots);
 203                 slots->pool = (unsigned long)pool;
 204                 rwlock_init(&slots->lock);
 205         }
 206
 207         return slots;
 208 }
 209
 210 static inline struct z3fold_pool *slots_to_pool(struct z3fold_buddy_slots *s)
 211 {
 212         return (struct z3fold_pool *)(s->pool & ~HANDLE_FLAG_MASK);
 213 }
 214
 215 static inline struct z3fold_buddy_slots *handle_to_slots(unsigned long handle)
 216 {
 217         return (struct z3fold_buddy_slots *)(handle & ~(SLOTS_ALIGN - 1));
 218 }
 219
 220 /* Lock a z3fold page */
 221 static inline void z3fold_page_lock(struct z3fold_header *zhdr)
 222 {
 223         spin_lock(&zhdr->page_lock);
 224 }
 225
 226 /* Try to lock a z3fold page */
 227 static inline int z3fold_page_trylock(struct z3fold_header *zhdr)
 228 {
 229         return spin_trylock(&zhdr->page_lock);
 230 }
 231
 232 /* Unlock a z3fold page */
 233 static inline void z3fold_page_unlock(struct z3fold_header *zhdr)
 234 {
 235         spin_unlock(&zhdr->page_lock);
 236 }
 237
 238 /* return locked z3fold page if it's not headless */
 239 static inline struct z3fold_header *get_z3fold_header(unsigned long handle)
 240 {
 241         struct z3fold_buddy_slots *slots;
 242         struct z3fold_header *zhdr;
 243         int locked = 0;
 244
 245         if (!(handle & (1 << PAGE_HEADLESS))) {
 246                 slots = handle_to_slots(handle);
 247                 do {
 248                         unsigned long addr;
 249
 250                         read_lock(&slots->lock);
 251                         addr = *(unsigned long *)handle;
 252                         zhdr = (struct z3fold_header *)(addr & PAGE_MASK);
 253                         locked = z3fold_page_trylock(zhdr);
 254                         read_unlock(&slots->lock);
 255                         if (locked) {
 256                                 struct page *page = virt_to_page(zhdr);
 257
 258                                 if (!test_bit(PAGE_MIGRATED, &page->private))
 259                                         break;
 260                                 z3fold_page_unlock(zhdr);
 261                         }
 262                         cpu_relax();
 263                 } while (true);
 264         } else {
 265                 zhdr = (struct z3fold_header *)(handle & PAGE_MASK);
 266         }
 267
 268         return zhdr;
 269 }
 270
 271 static inline void put_z3fold_header(struct z3fold_header *zhdr)
 272 {
 273         struct page *page = virt_to_page(zhdr);
 274
 275         if (!test_bit(PAGE_HEADLESS, &page->private))
 276                 z3fold_page_unlock(zhdr);
 277 }
 278
 279 static inline void free_handle(unsigned long handle, struct z3fold_header *zhdr)
 280 {
 281         struct z3fold_buddy_slots *slots;
 282         int i;
 283         bool is_free;
 284
 285         if (WARN_ON(*(unsigned long *)handle == 0))
 286                 return;
 287
 288         slots = handle_to_slots(handle);
 289         write_lock(&slots->lock);
 290         *(unsigned long *)handle = 0;
 291
 292         if (test_bit(HANDLES_NOFREE, &slots->pool)) {
 293                 write_unlock(&slots->lock);
 294                 return; /* simple case, nothing else to do */
 295         }
 296
 297         if (zhdr->slots != slots)
 298                 zhdr->foreign_handles--;
 299
 300         is_free = true;
 301         for (i = 0; i <= BUDDY_MASK; i++) {
 302                 if (slots->slot[i]) {
 303                         is_free = false;
 304                         break;
 305                 }
 306         }
 307         write_unlock(&slots->lock);
 308
 309         if (is_free) {
 310                 struct z3fold_pool *pool = slots_to_pool(slots);
 311
 312                 if (zhdr->slots == slots)
 313                         zhdr->slots = NULL;
 314                 kmem_cache_free(pool->c_handle, slots);
 315         }
 316 }
 317
 318 /* Initializes the z3fold header of a newly allocated z3fold page */
 319 static struct z3fold_header *init_z3fold_page(struct page *page, bool headless,
 320                                         struct z3fold_pool *pool, gfp_t gfp)
 321 {
 322         struct z3fold_header *zhdr = page_address(page);
 323         struct z3fold_buddy_slots *slots;
 324
 325         clear_bit(PAGE_HEADLESS, &page->private);
 326         clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
 327         clear_bit(NEEDS_COMPACTING, &page->private);
 328         clear_bit(PAGE_STALE, &page->private);
 329         clear_bit(PAGE_CLAIMED, &page->private);
 330         clear_bit(PAGE_MIGRATED, &page->private);
 331         if (headless)
 332                 return zhdr;
 333
 334         slots = alloc_slots(pool, gfp);
 335         if (!slots)
 336                 return NULL;
 337
 338         memset(zhdr, 0, sizeof(*zhdr));
 339         spin_lock_init(&zhdr->page_lock);
 340         kref_init(&zhdr->refcount);
 341         zhdr->cpu = -1;
 342         zhdr->slots = slots;
 343         zhdr->pool = pool;
 344         INIT_LIST_HEAD(&zhdr->buddy);
 345         INIT_WORK(&zhdr->work, compact_page_work);
 346         return zhdr;
 347 }
 348
 349 /* Resets the struct page fields and frees the page */
 350 static void free_z3fold_page(struct page *page, bool headless)
 351 {
 352         if (!headless) {
 353                 lock_page(page);
 354                 __ClearPageMovable(page);
 355                 unlock_page(page);
 356         }
 357         __free_page(page);
 358 }
 359
 360 /* Helper function to build the index */
 361 static inline int __idx(struct z3fold_header *zhdr, enum buddy bud)
 362 {
 363         return (bud + zhdr->first_num) & BUDDY_MASK;
 364 }
 365
 366 /*
 367  * Encodes the handle of a particular buddy within a z3fold page
 368  * Pool lock should be held as this function accesses first_num
 369  */
 370 static unsigned long __encode_handle(struct z3fold_header *zhdr,
 371                                 struct z3fold_buddy_slots *slots,
 372                                 enum buddy bud)
 373 {
 374         unsigned long h = (unsigned long)zhdr;
 375         int idx = 0;
 376
 377         /*
 378          * For a headless page, its handle is its pointer with the extra
 379          * PAGE_HEADLESS bit set
 380          */
 381         if (bud == HEADLESS)
 382                 return h | (1 << PAGE_HEADLESS);
 383
 384         /* otherwise, return pointer to encoded handle */
 385         idx = __idx(zhdr, bud);
 386         h += idx;
 387         if (bud == LAST)
 388                 h |= (zhdr->last_chunks << BUDDY_SHIFT);
 389
 390         write_lock(&slots->lock);
 391         slots->slot[idx] = h;
 392         write_unlock(&slots->lock);
 393         return (unsigned long)&slots->slot[idx];
 394 }
 395
 396 static unsigned long encode_handle(struct z3fold_header *zhdr, enum buddy bud)
 397 {
 398         return __encode_handle(zhdr, zhdr->slots, bud);
 399 }
 400
 401 /* only for LAST bud, returns zero otherwise */
 402 static unsigned short handle_to_chunks(unsigned long handle)
 403 {
 404         struct z3fold_buddy_slots *slots = handle_to_slots(handle);
 405         unsigned long addr;
 406
 407         read_lock(&slots->lock);
 408         addr = *(unsigned long *)handle;
 409         read_unlock(&slots->lock);
 410         return (addr & ~PAGE_MASK) >> BUDDY_SHIFT;
 411 }
 412
 413 /*
 414  * (handle & BUDDY_MASK) < zhdr->first_num is possible in encode_handle
 415  *  but that doesn't matter. because the masking will result in the
 416  *  correct buddy number.
 417  */
 418 static enum buddy handle_to_buddy(unsigned long handle)
 419 {
 420         struct z3fold_header *zhdr;
 421         struct z3fold_buddy_slots *slots = handle_to_slots(handle);
 422         unsigned long addr;
 423
 424         read_lock(&slots->lock);
 425         WARN_ON(handle & (1 << PAGE_HEADLESS));
 426         addr = *(unsigned long *)handle;
 427         read_unlock(&slots->lock);
 428         zhdr = (struct z3fold_header *)(addr & PAGE_MASK);
 429         return (addr - zhdr->first_num) & BUDDY_MASK;
 430 }
 431
 432 static inline struct z3fold_pool *zhdr_to_pool(struct z3fold_header *zhdr)
 433 {
 434         return zhdr->pool;
 435 }
 436
 437 static void __release_z3fold_page(struct z3fold_header *zhdr, bool locked)
 438 {
 439         struct page *page = virt_to_page(zhdr);
 440         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 441
 442         WARN_ON(!list_empty(&zhdr->buddy));
 443         set_bit(PAGE_STALE, &page->private);
 444         clear_bit(NEEDS_COMPACTING, &page->private);
 445         spin_lock(&pool->lock);
 446         spin_unlock(&pool->lock);
 447
 448         if (locked)
 449                 z3fold_page_unlock(zhdr);
 450
 451         spin_lock(&pool->stale_lock);
 452         list_add(&zhdr->buddy, &pool->stale);
 453         queue_work(pool->release_wq, &pool->work);
 454         spin_unlock(&pool->stale_lock);
 455
 456         atomic64_dec(&pool->pages_nr);
 457 }
 458
 459 static void release_z3fold_page_locked(struct kref *ref)
 460 {
 461         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 462                                                 refcount);
 463         WARN_ON(z3fold_page_trylock(zhdr));
 464         __release_z3fold_page(zhdr, true);
 465 }
 466
 467 static void release_z3fold_page_locked_list(struct kref *ref)
 468 {
 469         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 470                                                refcount);
 471         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 472
 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 inline int put_z3fold_locked(struct z3fold_header *zhdr)
 482 {
 483         return kref_put(&zhdr->refcount, release_z3fold_page_locked);
 484 }
 485
 486 static inline int put_z3fold_locked_list(struct z3fold_header *zhdr)
 487 {
 488         return kref_put(&zhdr->refcount, release_z3fold_page_locked_list);
 489 }
 490
 491 static void free_pages_work(struct work_struct *w)
 492 {
 493         struct z3fold_pool *pool = container_of(w, struct z3fold_pool, work);
 494
 495         spin_lock(&pool->stale_lock);
 496         while (!list_empty(&pool->stale)) {
 497                 struct z3fold_header *zhdr = list_first_entry(&pool->stale,
 498                                                 struct z3fold_header, buddy);
 499                 struct page *page = virt_to_page(zhdr);
 500
 501                 list_del(&zhdr->buddy);
 502                 if (WARN_ON(!test_bit(PAGE_STALE, &page->private)))
 503                         continue;
 504                 spin_unlock(&pool->stale_lock);
 505                 cancel_work_sync(&zhdr->work);
 506                 free_z3fold_page(page, false);
 507                 cond_resched();
 508                 spin_lock(&pool->stale_lock);
 509         }
 510         spin_unlock(&pool->stale_lock);
 511 }
 512
 513 /*
 514  * Returns the number of free chunks in a z3fold page.
 515  * NB: can't be used with HEADLESS pages.
 516  */
 517 static int num_free_chunks(struct z3fold_header *zhdr)
 518 {
 519         int nfree;
 520         /*
 521          * If there is a middle object, pick up the bigger free space
 522          * either before or after it. Otherwise just subtract the number
 523          * of chunks occupied by the first and the last objects.
 524          */
 525         if (zhdr->middle_chunks != 0) {
 526                 int nfree_before = zhdr->first_chunks ?
 527                         0 : zhdr->start_middle - ZHDR_CHUNKS;
 528                 int nfree_after = zhdr->last_chunks ?
 529                         0 : TOTAL_CHUNKS -
 530                                 (zhdr->start_middle + zhdr->middle_chunks);
 531                 nfree = max(nfree_before, nfree_after);
 532         } else
 533                 nfree = NCHUNKS - zhdr->first_chunks - zhdr->last_chunks;
 534         return nfree;
 535 }
 536
 537 /* Add to the appropriate unbuddied list */
 538 static inline void add_to_unbuddied(struct z3fold_pool *pool,
 539                                 struct z3fold_header *zhdr)
 540 {
 541         if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0 ||
 542                         zhdr->middle_chunks == 0) {
 543                 struct list_head *unbuddied;
 544                 int freechunks = num_free_chunks(zhdr);
 545
 546                 migrate_disable();
 547                 unbuddied = this_cpu_ptr(pool->unbuddied);
 548                 spin_lock(&pool->lock);
 549                 list_add(&zhdr->buddy, &unbuddied[freechunks]);
 550                 spin_unlock(&pool->lock);
 551                 zhdr->cpu = smp_processor_id();
 552                 migrate_enable();
 553         }
 554 }
 555
 556 static inline enum buddy get_free_buddy(struct z3fold_header *zhdr, int chunks)
 557 {
 558         enum buddy bud = HEADLESS;
 559
 560         if (zhdr->middle_chunks) {
 561                 if (!zhdr->first_chunks &&
 562                     chunks <= zhdr->start_middle - ZHDR_CHUNKS)
 563                         bud = FIRST;
 564                 else if (!zhdr->last_chunks)
 565                         bud = LAST;
 566         } else {
 567                 if (!zhdr->first_chunks)
 568                         bud = FIRST;
 569                 else if (!zhdr->last_chunks)
 570                         bud = LAST;
 571                 else
 572                         bud = MIDDLE;
 573         }
 574
 575         return bud;
 576 }
 577
 578 static inline void *mchunk_memmove(struct z3fold_header *zhdr,
 579                                 unsigned short dst_chunk)
 580 {
 581         void *beg = zhdr;
 582         return memmove(beg + (dst_chunk << CHUNK_SHIFT),
 583                        beg + (zhdr->start_middle << CHUNK_SHIFT),
 584                        zhdr->middle_chunks << CHUNK_SHIFT);
 585 }
 586
 587 static inline bool buddy_single(struct z3fold_header *zhdr)
 588 {
 589         return !((zhdr->first_chunks && zhdr->middle_chunks) ||
 590                         (zhdr->first_chunks && zhdr->last_chunks) ||
 591                         (zhdr->middle_chunks && zhdr->last_chunks));
 592 }
 593
 594 static struct z3fold_header *compact_single_buddy(struct z3fold_header *zhdr)
 595 {
 596         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 597         void *p = zhdr;
 598         unsigned long old_handle = 0;
 599         size_t sz = 0;
 600         struct z3fold_header *new_zhdr = NULL;
 601         int first_idx = __idx(zhdr, FIRST);
 602         int middle_idx = __idx(zhdr, MIDDLE);
 603         int last_idx = __idx(zhdr, LAST);
 604         unsigned short *moved_chunks = NULL;
 605
 606         /*
 607          * No need to protect slots here -- all the slots are "local" and
 608          * the page lock is already taken
 609          */
 610         if (zhdr->first_chunks && zhdr->slots->slot[first_idx]) {
 611                 p += ZHDR_SIZE_ALIGNED;
 612                 sz = zhdr->first_chunks << CHUNK_SHIFT;
 613                 old_handle = (unsigned long)&zhdr->slots->slot[first_idx];
 614                 moved_chunks = &zhdr->first_chunks;
 615         } else if (zhdr->middle_chunks && zhdr->slots->slot[middle_idx]) {
 616                 p += zhdr->start_middle << CHUNK_SHIFT;
 617                 sz = zhdr->middle_chunks << CHUNK_SHIFT;
 618                 old_handle = (unsigned long)&zhdr->slots->slot[middle_idx];
 619                 moved_chunks = &zhdr->middle_chunks;
 620         } else if (zhdr->last_chunks && zhdr->slots->slot[last_idx]) {
 621                 p += PAGE_SIZE - (zhdr->last_chunks << CHUNK_SHIFT);
 622                 sz = zhdr->last_chunks << CHUNK_SHIFT;
 623                 old_handle = (unsigned long)&zhdr->slots->slot[last_idx];
 624                 moved_chunks = &zhdr->last_chunks;
 625         }
 626
 627         if (sz > 0) {
 628                 enum buddy new_bud = HEADLESS;
 629                 short chunks = size_to_chunks(sz);
 630                 void *q;
 631
 632                 new_zhdr = __z3fold_alloc(pool, sz, false);
 633                 if (!new_zhdr)
 634                         return NULL;
 635
 636                 if (WARN_ON(new_zhdr == zhdr))
 637                         goto out_fail;
 638
 639                 new_bud = get_free_buddy(new_zhdr, chunks);
 640                 q = new_zhdr;
 641                 switch (new_bud) {
 642                 case FIRST:
 643                         new_zhdr->first_chunks = chunks;
 644                         q += ZHDR_SIZE_ALIGNED;
 645                         break;
 646                 case MIDDLE:
 647                         new_zhdr->middle_chunks = chunks;
 648                         new_zhdr->start_middle =
 649                                 new_zhdr->first_chunks + ZHDR_CHUNKS;
 650                         q += new_zhdr->start_middle << CHUNK_SHIFT;
 651                         break;
 652                 case LAST:
 653                         new_zhdr->last_chunks = chunks;
 654                         q += PAGE_SIZE - (new_zhdr->last_chunks << CHUNK_SHIFT);
 655                         break;
 656                 default:
 657                         goto out_fail;
 658                 }
 659                 new_zhdr->foreign_handles++;
 660                 memcpy(q, p, sz);
 661                 write_lock(&zhdr->slots->lock);
 662                 *(unsigned long *)old_handle = (unsigned long)new_zhdr +
 663                         __idx(new_zhdr, new_bud);
 664                 if (new_bud == LAST)
 665                         *(unsigned long *)old_handle |=
 666                                         (new_zhdr->last_chunks << BUDDY_SHIFT);
 667                 write_unlock(&zhdr->slots->lock);
 668                 add_to_unbuddied(pool, new_zhdr);
 669                 z3fold_page_unlock(new_zhdr);
 670
 671                 *moved_chunks = 0;
 672         }
 673
 674         return new_zhdr;
 675
 676 out_fail:
 677         if (new_zhdr && !put_z3fold_locked(new_zhdr)) {
 678                 add_to_unbuddied(pool, new_zhdr);
 679                 z3fold_page_unlock(new_zhdr);
 680         }
 681         return NULL;
 682
 683 }
 684
 685 #define BIG_CHUNK_GAP   3
 686 /* Has to be called with lock held */
 687 static int z3fold_compact_page(struct z3fold_header *zhdr)
 688 {
 689         struct page *page = virt_to_page(zhdr);
 690
 691         if (test_bit(MIDDLE_CHUNK_MAPPED, &page->private))
 692                 return 0; /* can't move middle chunk, it's used */
 693
 694         if (unlikely(PageIsolated(page)))
 695                 return 0;
 696
 697         if (zhdr->middle_chunks == 0)
 698                 return 0; /* nothing to compact */
 699
 700         if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
 701                 /* move to the beginning */
 702                 mchunk_memmove(zhdr, ZHDR_CHUNKS);
 703                 zhdr->first_chunks = zhdr->middle_chunks;
 704                 zhdr->middle_chunks = 0;
 705                 zhdr->start_middle = 0;
 706                 zhdr->first_num++;
 707                 return 1;
 708         }
 709
 710         /*
 711          * moving data is expensive, so let's only do that if
 712          * there's substantial gain (at least BIG_CHUNK_GAP chunks)
 713          */
 714         if (zhdr->first_chunks != 0 && zhdr->last_chunks == 0 &&
 715             zhdr->start_middle - (zhdr->first_chunks + ZHDR_CHUNKS) >=
 716                         BIG_CHUNK_GAP) {
 717                 mchunk_memmove(zhdr, zhdr->first_chunks + ZHDR_CHUNKS);
 718                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
 719                 return 1;
 720         } else if (zhdr->last_chunks != 0 && zhdr->first_chunks == 0 &&
 721                    TOTAL_CHUNKS - (zhdr->last_chunks + zhdr->start_middle
 722                                         + zhdr->middle_chunks) >=
 723                         BIG_CHUNK_GAP) {
 724                 unsigned short new_start = TOTAL_CHUNKS - zhdr->last_chunks -
 725                         zhdr->middle_chunks;
 726                 mchunk_memmove(zhdr, new_start);
 727                 zhdr->start_middle = new_start;
 728                 return 1;
 729         }
 730
 731         return 0;
 732 }
 733
 734 static void do_compact_page(struct z3fold_header *zhdr, bool locked)
 735 {
 736         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 737         struct page *page;
 738
 739         page = virt_to_page(zhdr);
 740         if (locked)
 741                 WARN_ON(z3fold_page_trylock(zhdr));
 742         else
 743                 z3fold_page_lock(zhdr);
 744         if (WARN_ON(!test_and_clear_bit(NEEDS_COMPACTING, &page->private))) {
 745                 z3fold_page_unlock(zhdr);
 746                 return;
 747         }
 748         spin_lock(&pool->lock);
 749         list_del_init(&zhdr->buddy);
 750         spin_unlock(&pool->lock);
 751
 752         if (put_z3fold_locked(zhdr))
 753                 return;
 754
 755         if (test_bit(PAGE_STALE, &page->private) ||
 756             test_and_set_bit(PAGE_CLAIMED, &page->private)) {
 757                 z3fold_page_unlock(zhdr);
 758                 return;
 759         }
 760
 761         if (!zhdr->foreign_handles && buddy_single(zhdr) &&
 762             zhdr->mapped_count == 0 && compact_single_buddy(zhdr)) {
 763                 if (!put_z3fold_locked(zhdr)) {
 764                         clear_bit(PAGE_CLAIMED, &page->private);
 765                         z3fold_page_unlock(zhdr);
 766                 }
 767                 return;
 768         }
 769
 770         z3fold_compact_page(zhdr);
 771         add_to_unbuddied(pool, zhdr);
 772         clear_bit(PAGE_CLAIMED, &page->private);
 773         z3fold_page_unlock(zhdr);
 774 }
 775
 776 static void compact_page_work(struct work_struct *w)
 777 {
 778         struct z3fold_header *zhdr = container_of(w, struct z3fold_header,
 779                                                 work);
 780
 781         do_compact_page(zhdr, false);
 782 }
 783
 784 /* returns _locked_ z3fold page header or NULL */
 785 static inline struct z3fold_header *__z3fold_alloc(struct z3fold_pool *pool,
 786                                                 size_t size, bool can_sleep)
 787 {
 788         struct z3fold_header *zhdr = NULL;
 789         struct page *page;
 790         struct list_head *unbuddied;
 791         int chunks = size_to_chunks(size), i;
 792
 793 lookup:
 794         migrate_disable();
 795         /* First, try to find an unbuddied z3fold page. */
 796         unbuddied = this_cpu_ptr(pool->unbuddied);
 797         for_each_unbuddied_list(i, chunks) {
 798                 struct list_head *l = &unbuddied[i];
 799
 800                 zhdr = list_first_entry_or_null(READ_ONCE(l),
 801                                         struct z3fold_header, buddy);
 802
 803                 if (!zhdr)
 804                         continue;
 805
 806                 /* Re-check under lock. */
 807                 spin_lock(&pool->lock);
 808                 if (unlikely(zhdr != list_first_entry(READ_ONCE(l),
 809                                                 struct z3fold_header, buddy)) ||
 810                     !z3fold_page_trylock(zhdr)) {
 811                         spin_unlock(&pool->lock);
 812                         zhdr = NULL;
 813                         migrate_enable();
 814                         if (can_sleep)
 815                                 cond_resched();
 816                         goto lookup;
 817                 }
 818                 list_del_init(&zhdr->buddy);
 819                 zhdr->cpu = -1;
 820                 spin_unlock(&pool->lock);
 821
 822                 page = virt_to_page(zhdr);
 823                 if (test_bit(NEEDS_COMPACTING, &page->private) ||
 824                     test_bit(PAGE_CLAIMED, &page->private)) {
 825                         z3fold_page_unlock(zhdr);
 826                         zhdr = NULL;
 827                         migrate_enable();
 828                         if (can_sleep)
 829                                 cond_resched();
 830                         goto lookup;
 831                 }
 832
 833                 /*
 834                  * this page could not be removed from its unbuddied
 835                  * list while pool lock was held, and then we've taken
 836                  * page lock so kref_put could not be called before
 837                  * we got here, so it's safe to just call kref_get()
 838                  */
 839                 kref_get(&zhdr->refcount);
 840                 break;
 841         }
 842         migrate_enable();
 843
 844         if (!zhdr) {
 845                 int cpu;
 846
 847                 /* look for _exact_ match on other cpus' lists */
 848                 for_each_online_cpu(cpu) {
 849                         struct list_head *l;
 850
 851                         unbuddied = per_cpu_ptr(pool->unbuddied, cpu);
 852                         spin_lock(&pool->lock);
 853                         l = &unbuddied[chunks];
 854
 855                         zhdr = list_first_entry_or_null(READ_ONCE(l),
 856                                                 struct z3fold_header, buddy);
 857
 858                         if (!zhdr || !z3fold_page_trylock(zhdr)) {
 859                                 spin_unlock(&pool->lock);
 860                                 zhdr = NULL;
 861                                 continue;
 862                         }
 863                         list_del_init(&zhdr->buddy);
 864                         zhdr->cpu = -1;
 865                         spin_unlock(&pool->lock);
 866
 867                         page = virt_to_page(zhdr);
 868                         if (test_bit(NEEDS_COMPACTING, &page->private) ||
 869                             test_bit(PAGE_CLAIMED, &page->private)) {
 870                                 z3fold_page_unlock(zhdr);
 871                                 zhdr = NULL;
 872                                 if (can_sleep)
 873                                         cond_resched();
 874                                 continue;
 875                         }
 876                         kref_get(&zhdr->refcount);
 877                         break;
 878                 }
 879         }
 880
 881         if (zhdr && !zhdr->slots) {
 882                 zhdr->slots = alloc_slots(pool, GFP_ATOMIC);
 883                 if (!zhdr->slots)
 884                         goto out_fail;
 885         }
 886         return zhdr;
 887
 888 out_fail:
 889         if (!put_z3fold_locked(zhdr)) {
 890                 add_to_unbuddied(pool, zhdr);
 891                 z3fold_page_unlock(zhdr);
 892         }
 893         return NULL;
 894 }
 895
 896 /*
 897  * API Functions
 898  */
 899
 900 /**
 901  * z3fold_create_pool() - create a new z3fold pool
 902  * @name:       pool name
 903  * @gfp:        gfp flags when allocating the z3fold pool structure
 904  *
 905  * Return: pointer to the new z3fold pool or NULL if the metadata allocation
 906  * failed.
 907  */
 908 static struct z3fold_pool *z3fold_create_pool(const char *name, gfp_t gfp)
 909 {
 910         struct z3fold_pool *pool = NULL;
 911         int i, cpu;
 912
 913         pool = kzalloc(sizeof(struct z3fold_pool), gfp);
 914         if (!pool)
 915                 goto out;
 916         pool->c_handle = kmem_cache_create("z3fold_handle",
 917                                 sizeof(struct z3fold_buddy_slots),
 918                                 SLOTS_ALIGN, 0, NULL);
 919         if (!pool->c_handle)
 920                 goto out_c;
 921         spin_lock_init(&pool->lock);
 922         spin_lock_init(&pool->stale_lock);
 923         pool->unbuddied = __alloc_percpu(sizeof(struct list_head) * NCHUNKS,
 924                                          __alignof__(struct list_head));
 925         if (!pool->unbuddied)
 926                 goto out_pool;
 927         for_each_possible_cpu(cpu) {
 928                 struct list_head *unbuddied =
 929                                 per_cpu_ptr(pool->unbuddied, cpu);
 930                 for_each_unbuddied_list(i, 0)
 931                         INIT_LIST_HEAD(&unbuddied[i]);
 932         }
 933         INIT_LIST_HEAD(&pool->stale);
 934         atomic64_set(&pool->pages_nr, 0);
 935         pool->name = name;
 936         pool->compact_wq = create_singlethread_workqueue(pool->name);
 937         if (!pool->compact_wq)
 938                 goto out_unbuddied;
 939         pool->release_wq = create_singlethread_workqueue(pool->name);
 940         if (!pool->release_wq)
 941                 goto out_wq;
 942         INIT_WORK(&pool->work, free_pages_work);
 943         return pool;
 944
 945 out_wq:
 946         destroy_workqueue(pool->compact_wq);
 947 out_unbuddied:
 948         free_percpu(pool->unbuddied);
 949 out_pool:
 950         kmem_cache_destroy(pool->c_handle);
 951 out_c:
 952         kfree(pool);
 953 out:
 954         return NULL;
 955 }
 956
 957 /**
 958  * z3fold_destroy_pool() - destroys an existing z3fold pool
 959  * @pool:       the z3fold pool to be destroyed
 960  *
 961  * The pool should be emptied before this function is called.
 962  */
 963 static void z3fold_destroy_pool(struct z3fold_pool *pool)
 964 {
 965         kmem_cache_destroy(pool->c_handle);
 966
 967         /*
 968          * We need to destroy pool->compact_wq before pool->release_wq,
 969          * as any pending work on pool->compact_wq will call
 970          * queue_work(pool->release_wq, &pool->work).
 971          *
 972          * There are still outstanding pages until both workqueues are drained,
 973          * so we cannot unregister migration until then.
 974          */
 975
 976         destroy_workqueue(pool->compact_wq);
 977         destroy_workqueue(pool->release_wq);
 978         free_percpu(pool->unbuddied);
 979         kfree(pool);
 980 }
 981
 982 static const struct movable_operations z3fold_mops;
 983
 984 /**
 985  * z3fold_alloc() - allocates a region of a given size
 986  * @pool:       z3fold pool from which to allocate
 987  * @size:       size in bytes of the desired allocation
 988  * @gfp:        gfp flags used if the pool needs to grow
 989  * @handle:     handle of the new allocation
 990  *
 991  * This function will attempt to find a free region in the pool large enough to
 992  * satisfy the allocation request.  A search of the unbuddied lists is
 993  * performed first. If no suitable free region is found, then a new page is
 994  * allocated and added to the pool to satisfy the request.
 995  *
 996  * Return: 0 if success and handle is set, otherwise -EINVAL if the size or
 997  * gfp arguments are invalid or -ENOMEM if the pool was unable to allocate
 998  * a new page.
 999  */
1000 static int z3fold_alloc(struct z3fold_pool *pool, size_t size, gfp_t gfp,
1001                         unsigned long *handle)
1002 {
1003         int chunks = size_to_chunks(size);
1004         struct z3fold_header *zhdr = NULL;
1005         struct page *page = NULL;
1006         enum buddy bud;
1007         bool can_sleep = gfpflags_allow_blocking(gfp);
1008
1009         if (!size || (gfp & __GFP_HIGHMEM))
1010                 return -EINVAL;
1011
1012         if (size > PAGE_SIZE)
1013                 return -ENOSPC;
1014
1015         if (size > PAGE_SIZE - ZHDR_SIZE_ALIGNED - CHUNK_SIZE)
1016                 bud = HEADLESS;
1017         else {
1018 retry:
1019                 zhdr = __z3fold_alloc(pool, size, can_sleep);
1020                 if (zhdr) {
1021                         bud = get_free_buddy(zhdr, chunks);
1022                         if (bud == HEADLESS) {
1023                                 if (!put_z3fold_locked(zhdr))
1024                                         z3fold_page_unlock(zhdr);
1025                                 pr_err("No free chunks in unbuddied\n");
1026                                 WARN_ON(1);
1027                                 goto retry;
1028                         }
1029                         page = virt_to_page(zhdr);
1030                         goto found;
1031                 }
1032                 bud = FIRST;
1033         }
1034
1035         page = alloc_page(gfp);
1036         if (!page)
1037                 return -ENOMEM;
1038
1039         zhdr = init_z3fold_page(page, bud == HEADLESS, pool, gfp);
1040         if (!zhdr) {
1041                 __free_page(page);
1042                 return -ENOMEM;
1043         }
1044         atomic64_inc(&pool->pages_nr);
1045
1046         if (bud == HEADLESS) {
1047                 set_bit(PAGE_HEADLESS, &page->private);
1048                 goto headless;
1049         }
1050         if (can_sleep) {
1051                 lock_page(page);
1052                 __SetPageMovable(page, &z3fold_mops);
1053                 unlock_page(page);
1054         } else {
1055                 WARN_ON(!trylock_page(page));
1056                 __SetPageMovable(page, &z3fold_mops);
1057                 unlock_page(page);
1058         }
1059         z3fold_page_lock(zhdr);
1060
1061 found:
1062         if (bud == FIRST)
1063                 zhdr->first_chunks = chunks;
1064         else if (bud == LAST)
1065                 zhdr->last_chunks = chunks;
1066         else {
1067                 zhdr->middle_chunks = chunks;
1068                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
1069         }
1070         add_to_unbuddied(pool, zhdr);
1071
1072 headless:
1073         spin_lock(&pool->lock);
1074         *handle = encode_handle(zhdr, bud);
1075         spin_unlock(&pool->lock);
1076         if (bud != HEADLESS)
1077                 z3fold_page_unlock(zhdr);
1078
1079         return 0;
1080 }
1081
1082 /**
1083  * z3fold_free() - frees the allocation associated with the given handle
1084  * @pool:       pool in which the allocation resided
1085  * @handle:     handle associated with the allocation returned by z3fold_alloc()
1086  *
1087  * In the case that the z3fold page in which the allocation resides is under
1088  * reclaim, as indicated by the PAGE_CLAIMED flag being set, this function
1089  * only sets the first|middle|last_chunks to 0.  The page is actually freed
1090  * once all buddies are evicted (see z3fold_reclaim_page() below).
1091  */
1092 static void z3fold_free(struct z3fold_pool *pool, unsigned long handle)
1093 {
1094         struct z3fold_header *zhdr;
1095         struct page *page;
1096         enum buddy bud;
1097         bool page_claimed;
1098
1099         zhdr = get_z3fold_header(handle);
1100         page = virt_to_page(zhdr);
1101         page_claimed = test_and_set_bit(PAGE_CLAIMED, &page->private);
1102
1103         if (test_bit(PAGE_HEADLESS, &page->private)) {
1104                 /* if a headless page is under reclaim, just leave.
1105                  * NB: we use test_and_set_bit for a reason: if the bit
1106                  * has not been set before, we release this page
1107                  * immediately so we don't care about its value any more.
1108                  */
1109                 if (!page_claimed) {
1110                         put_z3fold_header(zhdr);
1111                         free_z3fold_page(page, true);
1112                         atomic64_dec(&pool->pages_nr);
1113                 }
1114                 return;
1115         }
1116
1117         /* Non-headless case */
1118         bud = handle_to_buddy(handle);
1119
1120         switch (bud) {
1121         case FIRST:
1122                 zhdr->first_chunks = 0;
1123                 break;
1124         case MIDDLE:
1125                 zhdr->middle_chunks = 0;
1126                 break;
1127         case LAST:
1128                 zhdr->last_chunks = 0;
1129                 break;
1130         default:
1131                 pr_err("%s: unknown bud %d\n", __func__, bud);
1132                 WARN_ON(1);
1133                 put_z3fold_header(zhdr);
1134                 return;
1135         }
1136
1137         if (!page_claimed)
1138                 free_handle(handle, zhdr);
1139         if (put_z3fold_locked_list(zhdr))
1140                 return;
1141         if (page_claimed) {
1142                 /* the page has not been claimed by us */
1143                 put_z3fold_header(zhdr);
1144                 return;
1145         }
1146         if (test_and_set_bit(NEEDS_COMPACTING, &page->private)) {
1147                 clear_bit(PAGE_CLAIMED, &page->private);
1148                 put_z3fold_header(zhdr);
1149                 return;
1150         }
1151         if (zhdr->cpu < 0 || !cpu_online(zhdr->cpu)) {
1152                 zhdr->cpu = -1;
1153                 kref_get(&zhdr->refcount);
1154                 clear_bit(PAGE_CLAIMED, &page->private);
1155                 do_compact_page(zhdr, true);
1156                 return;
1157         }
1158         kref_get(&zhdr->refcount);
1159         clear_bit(PAGE_CLAIMED, &page->private);
1160         queue_work_on(zhdr->cpu, pool->compact_wq, &zhdr->work);
1161         put_z3fold_header(zhdr);
1162 }
1163
1164 /**
1165  * z3fold_map() - maps the allocation associated with the given handle
1166  * @pool:       pool in which the allocation resides
1167  * @handle:     handle associated with the allocation to be mapped
1168  *
1169  * Extracts the buddy number from handle and constructs the pointer to the
1170  * correct starting chunk within the page.
1171  *
1172  * Returns: a pointer to the mapped allocation
1173  */
1174 static void *z3fold_map(struct z3fold_pool *pool, unsigned long handle)
1175 {
1176         struct z3fold_header *zhdr;
1177         struct page *page;
1178         void *addr;
1179         enum buddy buddy;
1180
1181         zhdr = get_z3fold_header(handle);
1182         addr = zhdr;
1183         page = virt_to_page(zhdr);
1184
1185         if (test_bit(PAGE_HEADLESS, &page->private))
1186                 goto out;
1187
1188         buddy = handle_to_buddy(handle);
1189         switch (buddy) {
1190         case FIRST:
1191                 addr += ZHDR_SIZE_ALIGNED;
1192                 break;
1193         case MIDDLE:
1194                 addr += zhdr->start_middle << CHUNK_SHIFT;
1195                 set_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1196                 break;
1197         case LAST:
1198                 addr += PAGE_SIZE - (handle_to_chunks(handle) << CHUNK_SHIFT);
1199                 break;
1200         default:
1201                 pr_err("unknown buddy id %d\n", buddy);
1202                 WARN_ON(1);
1203                 addr = NULL;
1204                 break;
1205         }
1206
1207         if (addr)
1208                 zhdr->mapped_count++;
1209 out:
1210         put_z3fold_header(zhdr);
1211         return addr;
1212 }
1213
1214 /**
1215  * z3fold_unmap() - unmaps the allocation associated with the given handle
1216  * @pool:       pool in which the allocation resides
1217  * @handle:     handle associated with the allocation to be unmapped
1218  */
1219 static void z3fold_unmap(struct z3fold_pool *pool, unsigned long handle)
1220 {
1221         struct z3fold_header *zhdr;
1222         struct page *page;
1223         enum buddy buddy;
1224
1225         zhdr = get_z3fold_header(handle);
1226         page = virt_to_page(zhdr);
1227
1228         if (test_bit(PAGE_HEADLESS, &page->private))
1229                 return;
1230
1231         buddy = handle_to_buddy(handle);
1232         if (buddy == MIDDLE)
1233                 clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1234         zhdr->mapped_count--;
1235         put_z3fold_header(zhdr);
1236 }
1237
1238 /**
1239  * z3fold_get_pool_size() - gets the z3fold pool size in pages
1240  * @pool:       pool whose size is being queried
1241  *
1242  * Returns: size in pages of the given pool.
1243  */
1244 static u64 z3fold_get_pool_size(struct z3fold_pool *pool)
1245 {
1246         return atomic64_read(&pool->pages_nr);
1247 }
1248
1249 static bool z3fold_page_isolate(struct page *page, isolate_mode_t mode)
1250 {
1251         struct z3fold_header *zhdr;
1252         struct z3fold_pool *pool;
1253
1254         VM_BUG_ON_PAGE(PageIsolated(page), page);
1255
1256         if (test_bit(PAGE_HEADLESS, &page->private))
1257                 return false;
1258
1259         zhdr = page_address(page);
1260         z3fold_page_lock(zhdr);
1261         if (test_bit(NEEDS_COMPACTING, &page->private) ||
1262             test_bit(PAGE_STALE, &page->private))
1263                 goto out;
1264
1265         if (zhdr->mapped_count != 0 || zhdr->foreign_handles != 0)
1266                 goto out;
1267
1268         if (test_and_set_bit(PAGE_CLAIMED, &page->private))
1269                 goto out;
1270         pool = zhdr_to_pool(zhdr);
1271         spin_lock(&pool->lock);
1272         if (!list_empty(&zhdr->buddy))
1273                 list_del_init(&zhdr->buddy);
1274         spin_unlock(&pool->lock);
1275
1276         kref_get(&zhdr->refcount);
1277         z3fold_page_unlock(zhdr);
1278         return true;
1279
1280 out:
1281         z3fold_page_unlock(zhdr);
1282         return false;
1283 }
1284
1285 static int z3fold_page_migrate(struct page *newpage, struct page *page,
1286                 enum migrate_mode mode)
1287 {
1288         struct z3fold_header *zhdr, *new_zhdr;
1289         struct z3fold_pool *pool;
1290
1291         VM_BUG_ON_PAGE(!PageIsolated(page), page);
1292         VM_BUG_ON_PAGE(!test_bit(PAGE_CLAIMED, &page->private), page);
1293         VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
1294
1295         zhdr = page_address(page);
1296         pool = zhdr_to_pool(zhdr);
1297
1298         if (!z3fold_page_trylock(zhdr))
1299                 return -EAGAIN;
1300         if (zhdr->mapped_count != 0 || zhdr->foreign_handles != 0) {
1301                 clear_bit(PAGE_CLAIMED, &page->private);
1302                 z3fold_page_unlock(zhdr);
1303                 return -EBUSY;
1304         }
1305         if (work_pending(&zhdr->work)) {
1306                 z3fold_page_unlock(zhdr);
1307                 return -EAGAIN;
1308         }
1309         new_zhdr = page_address(newpage);
1310         memcpy(new_zhdr, zhdr, PAGE_SIZE);
1311         newpage->private = page->private;
1312         set_bit(PAGE_MIGRATED, &page->private);
1313         z3fold_page_unlock(zhdr);
1314         spin_lock_init(&new_zhdr->page_lock);
1315         INIT_WORK(&new_zhdr->work, compact_page_work);
1316         /*
1317          * z3fold_page_isolate() ensures that new_zhdr->buddy is empty,
1318          * so we only have to reinitialize it.
1319          */
1320         INIT_LIST_HEAD(&new_zhdr->buddy);
1321         __ClearPageMovable(page);
1322
1323         get_page(newpage);
1324         z3fold_page_lock(new_zhdr);
1325         if (new_zhdr->first_chunks)
1326                 encode_handle(new_zhdr, FIRST);
1327         if (new_zhdr->last_chunks)
1328                 encode_handle(new_zhdr, LAST);
1329         if (new_zhdr->middle_chunks)
1330                 encode_handle(new_zhdr, MIDDLE);
1331         set_bit(NEEDS_COMPACTING, &newpage->private);
1332         new_zhdr->cpu = smp_processor_id();
1333         __SetPageMovable(newpage, &z3fold_mops);
1334         z3fold_page_unlock(new_zhdr);
1335
1336         queue_work_on(new_zhdr->cpu, pool->compact_wq, &new_zhdr->work);
1337
1338         /* PAGE_CLAIMED and PAGE_MIGRATED are cleared now. */
1339         page->private = 0;
1340         put_page(page);
1341         return 0;
1342 }
1343
1344 static void z3fold_page_putback(struct page *page)
1345 {
1346         struct z3fold_header *zhdr;
1347         struct z3fold_pool *pool;
1348
1349         zhdr = page_address(page);
1350         pool = zhdr_to_pool(zhdr);
1351
1352         z3fold_page_lock(zhdr);
1353         if (!list_empty(&zhdr->buddy))
1354                 list_del_init(&zhdr->buddy);
1355         INIT_LIST_HEAD(&page->lru);
1356         if (put_z3fold_locked(zhdr))
1357                 return;
1358         if (list_empty(&zhdr->buddy))
1359                 add_to_unbuddied(pool, zhdr);
1360         clear_bit(PAGE_CLAIMED, &page->private);
1361         z3fold_page_unlock(zhdr);
1362 }
1363
1364 static const struct movable_operations z3fold_mops = {
1365         .isolate_page = z3fold_page_isolate,
1366         .migrate_page = z3fold_page_migrate,
1367         .putback_page = z3fold_page_putback,
1368 };
1369
1370 /*****************
1371  * zpool
1372  ****************/
1373
1374 static void *z3fold_zpool_create(const char *name, gfp_t gfp)
1375 {
1376         return z3fold_create_pool(name, gfp);
1377 }
1378
1379 static void z3fold_zpool_destroy(void *pool)
1380 {
1381         z3fold_destroy_pool(pool);
1382 }
1383
1384 static int z3fold_zpool_malloc(void *pool, size_t size, gfp_t gfp,
1385                         unsigned long *handle)
1386 {
1387         return z3fold_alloc(pool, size, gfp, handle);
1388 }
1389 static void z3fold_zpool_free(void *pool, unsigned long handle)
1390 {
1391         z3fold_free(pool, handle);
1392 }
1393
1394 static void *z3fold_zpool_map(void *pool, unsigned long handle,
1395                         enum zpool_mapmode mm)
1396 {
1397         return z3fold_map(pool, handle);
1398 }
1399 static void z3fold_zpool_unmap(void *pool, unsigned long handle)
1400 {
1401         z3fold_unmap(pool, handle);
1402 }
1403
1404 static u64 z3fold_zpool_total_size(void *pool)
1405 {
1406         return z3fold_get_pool_size(pool) * PAGE_SIZE;
1407 }
1408
1409 static struct zpool_driver z3fold_zpool_driver = {
1410         .type =         "z3fold",
1411         .sleep_mapped = true,
1412         .owner =        THIS_MODULE,
1413         .create =       z3fold_zpool_create,
1414         .destroy =      z3fold_zpool_destroy,
1415         .malloc =       z3fold_zpool_malloc,
1416         .free =         z3fold_zpool_free,
1417         .map =          z3fold_zpool_map,
1418         .unmap =        z3fold_zpool_unmap,
1419         .total_size =   z3fold_zpool_total_size,
1420 };
1421
1422 MODULE_ALIAS("zpool-z3fold");
1423
1424 static int __init init_z3fold(void)
1425 {
1426         /*
1427          * Make sure the z3fold header is not larger than the page size and
1428          * there has remaining spaces for its buddy.
1429          */
1430         BUILD_BUG_ON(ZHDR_SIZE_ALIGNED > PAGE_SIZE - CHUNK_SIZE);
1431         zpool_register_driver(&z3fold_zpool_driver);
1432
1433         return 0;
1434 }
1435
1436 static void __exit exit_z3fold(void)
1437 {
1438         zpool_unregister_driver(&z3fold_zpool_driver);
1439 }
1440
1441 module_init(init_z3fold);
1442 module_exit(exit_z3fold);
1443
1444 MODULE_LICENSE("GPL");
1445 MODULE_AUTHOR("Vitaly Wool <vitalywool@gmail.com>");
1446 MODULE_DESCRIPTION("3-Fold Allocator for Compressed Pages");